diff --git a/diffusers/pipelines/deprecated/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py b/diffusers/pipelines/deprecated/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..bb3009d238a44f007ccfa79e7793a2deccb3a1f3 --- /dev/null +++ b/diffusers/pipelines/deprecated/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py @@ -0,0 +1,732 @@ +import inspect +from itertools import repeat +from typing import Callable + +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....image_processor import VaeImageProcessor +from ....models import AutoencoderKL, UNet2DConditionModel +from ....pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import deprecate, is_torch_xla_available, logging +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import SemanticStableDiffusionPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class SemanticStableDiffusionPipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin): + _last_supported_version = "0.33.1" + r""" + Pipeline for text-to-image generation using Stable Diffusion with latent editing. + + This model inherits from [`DiffusionPipeline`] and builds on the [`StableDiffusionPipeline`]. Check the superclass + documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular + device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`Q16SafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + editing_prompt: str | list[str] | None = None, + editing_prompt_embeddings: torch.Tensor | None = None, + reverse_editing_direction: bool | list[bool] | None = False, + edit_guidance_scale: float | list[float] | None = 5, + edit_warmup_steps: int | list[int] | None = 10, + edit_cooldown_steps: int | list[int] | None = None, + edit_threshold: float | list[float] | None = 0.9, + edit_momentum_scale: float | None = 0.1, + edit_mom_beta: float | None = 0.4, + edit_weights: list[float] | None = None, + sem_guidance: list[torch.Tensor] | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + editing_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to use for semantic guidance. Semantic guidance is disabled by setting + `editing_prompt = None`. Guidance direction of prompt should be specified via + `reverse_editing_direction`. + editing_prompt_embeddings (`torch.Tensor`, *optional*): + Pre-computed embeddings to use for semantic guidance. Guidance direction of embedding should be + specified via `reverse_editing_direction`. + reverse_editing_direction (`bool` or `list[bool]`, *optional*, defaults to `False`): + Whether the corresponding prompt in `editing_prompt` should be increased or decreased. + edit_guidance_scale (`float` or `list[float]`, *optional*, defaults to 5): + Guidance scale for semantic guidance. If provided as a list, values should correspond to + `editing_prompt`. + edit_warmup_steps (`float` or `list[float]`, *optional*, defaults to 10): + Number of diffusion steps (for each prompt) for which semantic guidance is not applied. Momentum is + calculated for those steps and applied once all warmup periods are over. + edit_cooldown_steps (`float` or `list[float]`, *optional*, defaults to `None`): + Number of diffusion steps (for each prompt) after which semantic guidance is longer applied. + edit_threshold (`float` or `list[float]`, *optional*, defaults to 0.9): + Threshold of semantic guidance. + edit_momentum_scale (`float`, *optional*, defaults to 0.1): + Scale of the momentum to be added to the semantic guidance at each diffusion step. If set to 0.0, + momentum is disabled. Momentum is already built up during warmup (for diffusion steps smaller than + `sld_warmup_steps`). Momentum is only added to latent guidance once all warmup periods are finished. + edit_mom_beta (`float`, *optional*, defaults to 0.4): + Defines how semantic guidance momentum builds up. `edit_mom_beta` indicates how much of the previous + momentum is kept. Momentum is already built up during warmup (for diffusion steps smaller than + `edit_warmup_steps`). + edit_weights (`list[float]`, *optional*, defaults to `None`): + Indicates how much each individual concept should influence the overall guidance. If no weights are + provided all concepts are applied equally. + sem_guidance (`list[torch.Tensor]`, *optional*): + list of pre-generated guidance vectors to be applied at generation. Length of the list has to + correspond to `num_inference_steps`. + + Examples: + + ```py + >>> import torch + >>> from diffusers import SemanticStableDiffusionPipeline + + >>> pipe = SemanticStableDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> out = pipe( + ... prompt="a photo of the face of a woman", + ... num_images_per_prompt=1, + ... guidance_scale=7, + ... editing_prompt=[ + ... "smiling, smile", # Concepts to apply + ... "glasses, wearing glasses", + ... "curls, wavy hair, curly hair", + ... "beard, full beard, mustache", + ... ], + ... reverse_editing_direction=[ + ... False, + ... False, + ... False, + ... False, + ... ], # Direction of guidance i.e. increase all concepts + ... edit_warmup_steps=[10, 10, 10, 10], # Warmup period for each concept + ... edit_guidance_scale=[4, 5, 5, 5.4], # Guidance scale for each concept + ... edit_threshold=[ + ... 0.99, + ... 0.975, + ... 0.925, + ... 0.96, + ... ], # Threshold for each concept. Threshold equals the percentile of the latent space that will be discarded. I.e. threshold=0.99 uses 1% of the latent dimensions + ... edit_momentum_scale=0.3, # Momentum scale that will be added to the latent guidance + ... edit_mom_beta=0.6, # Momentum beta + ... edit_weights=[1, 1, 1, 1, 1], # Weights of the individual concepts against each other + ... ) + >>> image = out.images[0] + ``` + + Returns: + [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, + [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] is returned, otherwise a + `tuple` is returned where the first element is a list with the generated images and the second element + is a list of `bool`s indicating whether the corresponding generated image contains "not-safe-for-work" + (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + + if editing_prompt: + enable_edit_guidance = True + if isinstance(editing_prompt, str): + editing_prompt = [editing_prompt] + enabled_editing_prompts = len(editing_prompt) + elif editing_prompt_embeddings is not None: + enable_edit_guidance = True + enabled_editing_prompts = editing_prompt_embeddings.shape[0] + else: + enabled_editing_prompts = 0 + enable_edit_guidance = False + + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + if text_input_ids.shape[-1] > self.tokenizer.model_max_length: + removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + text_embeddings = self.text_encoder(text_input_ids.to(device))[0] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = text_embeddings.shape + text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) + text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if enable_edit_guidance: + # get safety text embeddings + if editing_prompt_embeddings is None: + edit_concepts_input = self.tokenizer( + [x for item in editing_prompt for x in repeat(item, batch_size)], + padding="max_length", + max_length=self.tokenizer.model_max_length, + return_tensors="pt", + ) + + edit_concepts_input_ids = edit_concepts_input.input_ids + + if edit_concepts_input_ids.shape[-1] > self.tokenizer.model_max_length: + removed_text = self.tokenizer.batch_decode( + edit_concepts_input_ids[:, self.tokenizer.model_max_length :] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + edit_concepts_input_ids = edit_concepts_input_ids[:, : self.tokenizer.model_max_length] + edit_concepts = self.text_encoder(edit_concepts_input_ids.to(device))[0] + else: + edit_concepts = editing_prompt_embeddings.to(device).repeat(batch_size, 1, 1) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed_edit, seq_len_edit, _ = edit_concepts.shape + edit_concepts = edit_concepts.repeat(1, num_images_per_prompt, 1) + edit_concepts = edit_concepts.view(bs_embed_edit * num_images_per_prompt, seq_len_edit, -1) + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + # get unconditional embeddings for classifier free guidance + + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(device))[0] + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if enable_edit_guidance: + text_embeddings = torch.cat([uncond_embeddings, text_embeddings, edit_concepts]) + else: + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + # get the initial random noise unless the user supplied it + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + text_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # Initialize edit_momentum to None + edit_momentum = None + + self.uncond_estimates = None + self.text_estimates = None + self.edit_estimates = None + self.sem_guidance = None + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = ( + torch.cat([latents] * (2 + enabled_editing_prompts)) if do_classifier_free_guidance else latents + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_out = noise_pred.chunk(2 + enabled_editing_prompts) # [b,4, 64, 64] + noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1] + noise_pred_edit_concepts = noise_pred_out[2:] + + # default text guidance + noise_guidance = guidance_scale * (noise_pred_text - noise_pred_uncond) + # noise_guidance = (noise_pred_text - noise_pred_edit_concepts[0]) + + if self.uncond_estimates is None: + self.uncond_estimates = torch.zeros((num_inference_steps + 1, *noise_pred_uncond.shape)) + self.uncond_estimates[i] = noise_pred_uncond.detach().cpu() + + if self.text_estimates is None: + self.text_estimates = torch.zeros((num_inference_steps + 1, *noise_pred_text.shape)) + self.text_estimates[i] = noise_pred_text.detach().cpu() + + if self.edit_estimates is None and enable_edit_guidance: + self.edit_estimates = torch.zeros( + (num_inference_steps + 1, len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape) + ) + + if self.sem_guidance is None: + self.sem_guidance = torch.zeros((num_inference_steps + 1, *noise_pred_text.shape)) + + if edit_momentum is None: + edit_momentum = torch.zeros_like(noise_guidance) + + if enable_edit_guidance: + concept_weights = torch.zeros( + (len(noise_pred_edit_concepts), noise_guidance.shape[0]), + device=device, + dtype=noise_guidance.dtype, + ) + noise_guidance_edit = torch.zeros( + (len(noise_pred_edit_concepts), *noise_guidance.shape), + device=device, + dtype=noise_guidance.dtype, + ) + # noise_guidance_edit = torch.zeros_like(noise_guidance) + warmup_inds = [] + for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts): + self.edit_estimates[i, c] = noise_pred_edit_concept + if isinstance(edit_guidance_scale, list): + edit_guidance_scale_c = edit_guidance_scale[c] + else: + edit_guidance_scale_c = edit_guidance_scale + + if isinstance(edit_threshold, list): + edit_threshold_c = edit_threshold[c] + else: + edit_threshold_c = edit_threshold + if isinstance(reverse_editing_direction, list): + reverse_editing_direction_c = reverse_editing_direction[c] + else: + reverse_editing_direction_c = reverse_editing_direction + if edit_weights: + edit_weight_c = edit_weights[c] + else: + edit_weight_c = 1.0 + if isinstance(edit_warmup_steps, list): + edit_warmup_steps_c = edit_warmup_steps[c] + else: + edit_warmup_steps_c = edit_warmup_steps + + if isinstance(edit_cooldown_steps, list): + edit_cooldown_steps_c = edit_cooldown_steps[c] + elif edit_cooldown_steps is None: + edit_cooldown_steps_c = i + 1 + else: + edit_cooldown_steps_c = edit_cooldown_steps + if i >= edit_warmup_steps_c: + warmup_inds.append(c) + if i >= edit_cooldown_steps_c: + noise_guidance_edit[c, :, :, :, :] = torch.zeros_like(noise_pred_edit_concept) + continue + + noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond + # tmp_weights = (noise_pred_text - noise_pred_edit_concept).sum(dim=(1, 2, 3)) + tmp_weights = (noise_guidance - noise_pred_edit_concept).sum(dim=(1, 2, 3)) + + tmp_weights = torch.full_like(tmp_weights, edit_weight_c) # * (1 / enabled_editing_prompts) + if reverse_editing_direction_c: + noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1 + concept_weights[c, :] = tmp_weights + + noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c + + # torch.quantile function expects float32 + if noise_guidance_edit_tmp.dtype == torch.float32: + tmp = torch.quantile( + torch.abs(noise_guidance_edit_tmp).flatten(start_dim=2), + edit_threshold_c, + dim=2, + keepdim=False, + ) + else: + tmp = torch.quantile( + torch.abs(noise_guidance_edit_tmp).flatten(start_dim=2).to(torch.float32), + edit_threshold_c, + dim=2, + keepdim=False, + ).to(noise_guidance_edit_tmp.dtype) + + noise_guidance_edit_tmp = torch.where( + torch.abs(noise_guidance_edit_tmp) >= tmp[:, :, None, None], + noise_guidance_edit_tmp, + torch.zeros_like(noise_guidance_edit_tmp), + ) + noise_guidance_edit[c, :, :, :, :] = noise_guidance_edit_tmp + + # noise_guidance_edit = noise_guidance_edit + noise_guidance_edit_tmp + + warmup_inds = torch.tensor(warmup_inds).to(device) + if len(noise_pred_edit_concepts) > warmup_inds.shape[0] > 0: + concept_weights = concept_weights.to("cpu") # Offload to cpu + noise_guidance_edit = noise_guidance_edit.to("cpu") + + concept_weights_tmp = torch.index_select(concept_weights.to(device), 0, warmup_inds) + concept_weights_tmp = torch.where( + concept_weights_tmp < 0, torch.zeros_like(concept_weights_tmp), concept_weights_tmp + ) + concept_weights_tmp = concept_weights_tmp / concept_weights_tmp.sum(dim=0) + # concept_weights_tmp = torch.nan_to_num(concept_weights_tmp) + + noise_guidance_edit_tmp = torch.index_select(noise_guidance_edit.to(device), 0, warmup_inds) + noise_guidance_edit_tmp = torch.einsum( + "cb,cbijk->bijk", concept_weights_tmp, noise_guidance_edit_tmp + ) + noise_guidance = noise_guidance + noise_guidance_edit_tmp + + self.sem_guidance[i] = noise_guidance_edit_tmp.detach().cpu() + + del noise_guidance_edit_tmp + del concept_weights_tmp + concept_weights = concept_weights.to(device) + noise_guidance_edit = noise_guidance_edit.to(device) + + concept_weights = torch.where( + concept_weights < 0, torch.zeros_like(concept_weights), concept_weights + ) + + concept_weights = torch.nan_to_num(concept_weights) + + noise_guidance_edit = torch.einsum("cb,cbijk->bijk", concept_weights, noise_guidance_edit) + noise_guidance_edit = noise_guidance_edit.to(edit_momentum.device) + + noise_guidance_edit = noise_guidance_edit + edit_momentum_scale * edit_momentum + + edit_momentum = edit_mom_beta * edit_momentum + (1 - edit_mom_beta) * noise_guidance_edit + + if warmup_inds.shape[0] == len(noise_pred_edit_concepts): + noise_guidance = noise_guidance + noise_guidance_edit + self.sem_guidance[i] = noise_guidance_edit.detach().cpu() + + if sem_guidance is not None: + edit_guidance = sem_guidance[i].to(device) + noise_guidance = noise_guidance + edit_guidance + + noise_pred = noise_pred_uncond + noise_guidance + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post-processing + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, text_embeddings.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + if not return_dict: + return (image, has_nsfw_concept) + + return SemanticStableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/spectrogram_diffusion/__init__.py b/diffusers/pipelines/deprecated/spectrogram_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..150954baa0eb8f8a7216b4891effc14a71e21b1b --- /dev/null +++ b/diffusers/pipelines/deprecated/spectrogram_diffusion/__init__.py @@ -0,0 +1,75 @@ +# flake8: noqa +from typing import TYPE_CHECKING +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + _LazyModule, + is_note_seq_available, + OptionalDependencyNotAvailable, + is_torch_available, + is_transformers_available, + get_objects_from_module, +) + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["continous_encoder"] = ["SpectrogramContEncoder"] + _import_structure["notes_encoder"] = ["SpectrogramNotesEncoder"] + _import_structure["pipeline_spectrogram_diffusion"] = [ + "SpectrogramContEncoder", + "SpectrogramDiffusionPipeline", + "T5FilmDecoder", + ] +try: + if not (is_transformers_available() and is_torch_available() and is_note_seq_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_transformers_and_torch_and_note_seq_objects + + _dummy_objects.update(get_objects_from_module(dummy_transformers_and_torch_and_note_seq_objects)) +else: + _import_structure["midi_utils"] = ["MidiProcessor"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_spectrogram_diffusion import SpectrogramDiffusionPipeline + from .pipeline_spectrogram_diffusion import SpectrogramContEncoder + from .pipeline_spectrogram_diffusion import SpectrogramNotesEncoder + from .pipeline_spectrogram_diffusion import T5FilmDecoder + + try: + if not (is_transformers_available() and is_torch_available() and is_note_seq_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_transformers_and_torch_and_note_seq_objects import * + + else: + from .midi_utils import MidiProcessor + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/spectrogram_diffusion/continuous_encoder.py b/diffusers/pipelines/deprecated/spectrogram_diffusion/continuous_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..51716e6acc6578f1a3bee200836054ee5ca10f51 --- /dev/null +++ b/diffusers/pipelines/deprecated/spectrogram_diffusion/continuous_encoder.py @@ -0,0 +1,92 @@ +# Copyright 2022 The Music Spectrogram Diffusion Authors. +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +from transformers.modeling_utils import ModuleUtilsMixin +from transformers.models.t5.modeling_t5 import ( + T5Block, + T5Config, + T5LayerNorm, +) + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models import ModelMixin + + +class SpectrogramContEncoder(ModelMixin, ConfigMixin, ModuleUtilsMixin): + @register_to_config + def __init__( + self, + input_dims: int, + targets_context_length: int, + d_model: int, + dropout_rate: float, + num_layers: int, + num_heads: int, + d_kv: int, + d_ff: int, + feed_forward_proj: str, + is_decoder: bool = False, + ): + super().__init__() + + self.input_proj = nn.Linear(input_dims, d_model, bias=False) + + self.position_encoding = nn.Embedding(targets_context_length, d_model) + self.position_encoding.weight.requires_grad = False + + self.dropout_pre = nn.Dropout(p=dropout_rate) + + t5config = T5Config( + d_model=d_model, + num_heads=num_heads, + d_kv=d_kv, + d_ff=d_ff, + feed_forward_proj=feed_forward_proj, + dropout_rate=dropout_rate, + is_decoder=is_decoder, + is_encoder_decoder=False, + ) + self.encoders = nn.ModuleList() + for lyr_num in range(num_layers): + lyr = T5Block(t5config) + self.encoders.append(lyr) + + self.layer_norm = T5LayerNorm(d_model) + self.dropout_post = nn.Dropout(p=dropout_rate) + + def forward(self, encoder_inputs, encoder_inputs_mask): + x = self.input_proj(encoder_inputs) + + # terminal relative positional encodings + max_positions = encoder_inputs.shape[1] + input_positions = torch.arange(max_positions, device=encoder_inputs.device) + + seq_lens = encoder_inputs_mask.sum(-1) + input_positions = torch.roll(input_positions.unsqueeze(0), tuple(seq_lens.tolist()), dims=0) + x += self.position_encoding(input_positions) + + x = self.dropout_pre(x) + + # inverted the attention mask + input_shape = encoder_inputs.size() + extended_attention_mask = self.get_extended_attention_mask(encoder_inputs_mask, input_shape) + + for lyr in self.encoders: + x = lyr(x, extended_attention_mask)[0] + x = self.layer_norm(x) + + return self.dropout_post(x), encoder_inputs_mask diff --git a/diffusers/pipelines/deprecated/spectrogram_diffusion/midi_utils.py b/diffusers/pipelines/deprecated/spectrogram_diffusion/midi_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..3c7a6c151583d4b9bf1cfb8a4bb25aacd9e999cc --- /dev/null +++ b/diffusers/pipelines/deprecated/spectrogram_diffusion/midi_utils.py @@ -0,0 +1,667 @@ +# Copyright 2022 The Music Spectrogram Diffusion Authors. +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import dataclasses +import math +import os +from typing import Any, Callable, Mapping, MutableMapping, Sequence + +import numpy as np +import torch +import torch.nn.functional as F + +from ....utils import is_note_seq_available +from .pipeline_spectrogram_diffusion import TARGET_FEATURE_LENGTH + + +if is_note_seq_available(): + import note_seq +else: + raise ImportError("Please install note-seq via `pip install note-seq`") + + +INPUT_FEATURE_LENGTH = 2048 + +SAMPLE_RATE = 16000 +HOP_SIZE = 320 +FRAME_RATE = int(SAMPLE_RATE // HOP_SIZE) + +DEFAULT_STEPS_PER_SECOND = 100 +DEFAULT_MAX_SHIFT_SECONDS = 10 +DEFAULT_NUM_VELOCITY_BINS = 1 + +SLAKH_CLASS_PROGRAMS = { + "Acoustic Piano": 0, + "Electric Piano": 4, + "Chromatic Percussion": 8, + "Organ": 16, + "Acoustic Guitar": 24, + "Clean Electric Guitar": 26, + "Distorted Electric Guitar": 29, + "Acoustic Bass": 32, + "Electric Bass": 33, + "Violin": 40, + "Viola": 41, + "Cello": 42, + "Contrabass": 43, + "Orchestral Harp": 46, + "Timpani": 47, + "String Ensemble": 48, + "Synth Strings": 50, + "Choir and Voice": 52, + "Orchestral Hit": 55, + "Trumpet": 56, + "Trombone": 57, + "Tuba": 58, + "French Horn": 60, + "Brass Section": 61, + "Soprano/Alto Sax": 64, + "Tenor Sax": 66, + "Baritone Sax": 67, + "Oboe": 68, + "English Horn": 69, + "Bassoon": 70, + "Clarinet": 71, + "Pipe": 73, + "Synth Lead": 80, + "Synth Pad": 88, +} + + +@dataclasses.dataclass +class NoteRepresentationConfig: + """Configuration note representations.""" + + onsets_only: bool + include_ties: bool + + +@dataclasses.dataclass +class NoteEventData: + pitch: int + velocity: int | None = None + program: int | None = None + is_drum: bool | None = None + instrument: int | None = None + + +@dataclasses.dataclass +class NoteEncodingState: + """Encoding state for note transcription, keeping track of active pitches.""" + + # velocity bin for active pitches and programs + active_pitches: MutableMapping[tuple[int, int], int] = dataclasses.field(default_factory=dict) + + +@dataclasses.dataclass +class EventRange: + type: str + min_value: int + max_value: int + + +@dataclasses.dataclass +class Event: + type: str + value: int + + +class Tokenizer: + def __init__(self, regular_ids: int): + # The special tokens: 0=PAD, 1=EOS, and 2=UNK + self._num_special_tokens = 3 + self._num_regular_tokens = regular_ids + + def encode(self, token_ids): + encoded = [] + for token_id in token_ids: + if not 0 <= token_id < self._num_regular_tokens: + raise ValueError( + f"token_id {token_id} does not fall within valid range of [0, {self._num_regular_tokens})" + ) + encoded.append(token_id + self._num_special_tokens) + + # Add EOS token + encoded.append(1) + + # Pad to till INPUT_FEATURE_LENGTH + encoded = encoded + [0] * (INPUT_FEATURE_LENGTH - len(encoded)) + + return encoded + + +class Codec: + """Encode and decode events. + + Useful for declaring what certain ranges of a vocabulary should be used for. This is intended to be used from + Python before encoding or after decoding with GenericTokenVocabulary. This class is more lightweight and does not + include things like EOS or UNK token handling. + + To ensure that 'shift' events are always the first block of the vocab and start at 0, that event type is required + and specified separately. + """ + + def __init__(self, max_shift_steps: int, steps_per_second: float, event_ranges: list[EventRange]): + """Define Codec. + + Args: + max_shift_steps: Maximum number of shift steps that can be encoded. + steps_per_second: Shift steps will be interpreted as having a duration of + 1 / steps_per_second. + event_ranges: Other supported event types and their ranges. + """ + self.steps_per_second = steps_per_second + self._shift_range = EventRange(type="shift", min_value=0, max_value=max_shift_steps) + self._event_ranges = [self._shift_range] + event_ranges + # Ensure all event types have unique names. + assert len(self._event_ranges) == len({er.type for er in self._event_ranges}) + + @property + def num_classes(self) -> int: + return sum(er.max_value - er.min_value + 1 for er in self._event_ranges) + + # The next couple methods are simplified special case methods just for shift + # events that are intended to be used from within autograph functions. + + def is_shift_event_index(self, index: int) -> bool: + return (self._shift_range.min_value <= index) and (index <= self._shift_range.max_value) + + @property + def max_shift_steps(self) -> int: + return self._shift_range.max_value + + def encode_event(self, event: Event) -> int: + """Encode an event to an index.""" + offset = 0 + for er in self._event_ranges: + if event.type == er.type: + if not er.min_value <= event.value <= er.max_value: + raise ValueError( + f"Event value {event.value} is not within valid range " + f"[{er.min_value}, {er.max_value}] for type {event.type}" + ) + return offset + event.value - er.min_value + offset += er.max_value - er.min_value + 1 + + raise ValueError(f"Unknown event type: {event.type}") + + def event_type_range(self, event_type: str) -> tuple[int, int]: + """Return [min_id, max_id] for an event type.""" + offset = 0 + for er in self._event_ranges: + if event_type == er.type: + return offset, offset + (er.max_value - er.min_value) + offset += er.max_value - er.min_value + 1 + + raise ValueError(f"Unknown event type: {event_type}") + + def decode_event_index(self, index: int) -> Event: + """Decode an event index to an Event.""" + offset = 0 + for er in self._event_ranges: + if offset <= index <= offset + er.max_value - er.min_value: + return Event(type=er.type, value=er.min_value + index - offset) + offset += er.max_value - er.min_value + 1 + + raise ValueError(f"Unknown event index: {index}") + + +@dataclasses.dataclass +class ProgramGranularity: + # both tokens_map_fn and program_map_fn should be idempotent + tokens_map_fn: Callable[[Sequence[int], Codec], Sequence[int]] + program_map_fn: Callable[[int], int] + + +def drop_programs(tokens, codec: Codec): + """Drops program change events from a token sequence.""" + min_program_id, max_program_id = codec.event_type_range("program") + return tokens[(tokens < min_program_id) | (tokens > max_program_id)] + + +def programs_to_midi_classes(tokens, codec): + """Modifies program events to be the first program in the MIDI class.""" + min_program_id, max_program_id = codec.event_type_range("program") + is_program = (tokens >= min_program_id) & (tokens <= max_program_id) + return np.where(is_program, min_program_id + 8 * ((tokens - min_program_id) // 8), tokens) + + +PROGRAM_GRANULARITIES = { + # "flat" granularity; drop program change tokens and set NoteSequence + # programs to zero + "flat": ProgramGranularity(tokens_map_fn=drop_programs, program_map_fn=lambda program: 0), + # map each program to the first program in its MIDI class + "midi_class": ProgramGranularity( + tokens_map_fn=programs_to_midi_classes, program_map_fn=lambda program: 8 * (program // 8) + ), + # leave programs as is + "full": ProgramGranularity(tokens_map_fn=lambda tokens, codec: tokens, program_map_fn=lambda program: program), +} + + +def frame(signal, frame_length, frame_step, pad_end=False, pad_value=0, axis=-1): + """ + equivalent of tf.signal.frame + """ + signal_length = signal.shape[axis] + if pad_end: + frames_overlap = frame_length - frame_step + rest_samples = np.abs(signal_length - frames_overlap) % np.abs(frame_length - frames_overlap) + pad_size = int(frame_length - rest_samples) + + if pad_size != 0: + pad_axis = [0] * signal.ndim + pad_axis[axis] = pad_size + signal = F.pad(signal, pad_axis, "constant", pad_value) + frames = signal.unfold(axis, frame_length, frame_step) + return frames + + +def program_to_slakh_program(program): + # this is done very hackily, probably should use a custom mapping + for slakh_program in sorted(SLAKH_CLASS_PROGRAMS.values(), reverse=True): + if program >= slakh_program: + return slakh_program + + +def audio_to_frames( + samples, + hop_size: int, + frame_rate: int, +) -> tuple[Sequence[Sequence[int]], torch.Tensor]: + """Convert audio samples to non-overlapping frames and frame times.""" + frame_size = hop_size + samples = np.pad(samples, [0, frame_size - len(samples) % frame_size], mode="constant") + + # Split audio into frames. + frames = frame( + torch.Tensor(samples).unsqueeze(0), + frame_length=frame_size, + frame_step=frame_size, + pad_end=False, # TODO check why its off by 1 here when True + ) + + num_frames = len(samples) // frame_size + + times = np.arange(num_frames) / frame_rate + return frames, times + + +def note_sequence_to_onsets_and_offsets_and_programs( + ns: note_seq.NoteSequence, +) -> tuple[Sequence[float], Sequence[NoteEventData]]: + """Extract onset & offset times and pitches & programs from a NoteSequence. + + The onset & offset times will not necessarily be in sorted order. + + Args: + ns: NoteSequence from which to extract onsets and offsets. + + Returns: + times: A list of note onset and offset times. values: A list of NoteEventData objects where velocity is zero for + note + offsets. + """ + # Sort by program and pitch and put offsets before onsets as a tiebreaker for + # subsequent stable sort. + notes = sorted(ns.notes, key=lambda note: (note.is_drum, note.program, note.pitch)) + times = [note.end_time for note in notes if not note.is_drum] + [note.start_time for note in notes] + values = [ + NoteEventData(pitch=note.pitch, velocity=0, program=note.program, is_drum=False) + for note in notes + if not note.is_drum + ] + [ + NoteEventData(pitch=note.pitch, velocity=note.velocity, program=note.program, is_drum=note.is_drum) + for note in notes + ] + return times, values + + +def num_velocity_bins_from_codec(codec: Codec): + """Get number of velocity bins from event codec.""" + lo, hi = codec.event_type_range("velocity") + return hi - lo + + +# segment an array into segments of length n +def segment(a, n): + return [a[i : i + n] for i in range(0, len(a), n)] + + +def velocity_to_bin(velocity, num_velocity_bins): + if velocity == 0: + return 0 + else: + return math.ceil(num_velocity_bins * velocity / note_seq.MAX_MIDI_VELOCITY) + + +def note_event_data_to_events( + state: NoteEncodingState | None, + value: NoteEventData, + codec: Codec, +) -> Sequence[Event]: + """Convert note event data to a sequence of events.""" + if value.velocity is None: + # onsets only, no program or velocity + return [Event("pitch", value.pitch)] + else: + num_velocity_bins = num_velocity_bins_from_codec(codec) + velocity_bin = velocity_to_bin(value.velocity, num_velocity_bins) + if value.program is None: + # onsets + offsets + velocities only, no programs + if state is not None: + state.active_pitches[(value.pitch, 0)] = velocity_bin + return [Event("velocity", velocity_bin), Event("pitch", value.pitch)] + else: + if value.is_drum: + # drum events use a separate vocabulary + return [Event("velocity", velocity_bin), Event("drum", value.pitch)] + else: + # program + velocity + pitch + if state is not None: + state.active_pitches[(value.pitch, value.program)] = velocity_bin + return [ + Event("program", value.program), + Event("velocity", velocity_bin), + Event("pitch", value.pitch), + ] + + +def note_encoding_state_to_events(state: NoteEncodingState) -> Sequence[Event]: + """Output program and pitch events for active notes plus a final tie event.""" + events = [] + for pitch, program in sorted(state.active_pitches.keys(), key=lambda k: k[::-1]): + if state.active_pitches[(pitch, program)]: + events += [Event("program", program), Event("pitch", pitch)] + events.append(Event("tie", 0)) + return events + + +def encode_and_index_events( + state, event_times, event_values, codec, frame_times, encode_event_fn, encoding_state_to_events_fn=None +): + """Encode a sequence of timed events and index to audio frame times. + + Encodes time shifts as repeated single step shifts for later run length encoding. + + Optionally, also encodes a sequence of "state events", keeping track of the current encoding state at each audio + frame. This can be used e.g. to prepend events representing the current state to a targets segment. + + Args: + state: Initial event encoding state. + event_times: Sequence of event times. + event_values: Sequence of event values. + encode_event_fn: Function that transforms event value into a sequence of one + or more Event objects. + codec: An Codec object that maps Event objects to indices. + frame_times: Time for every audio frame. + encoding_state_to_events_fn: Function that transforms encoding state into a + sequence of one or more Event objects. + + Returns: + events: Encoded events and shifts. event_start_indices: Corresponding start event index for every audio frame. + Note: one event can correspond to multiple audio indices due to sampling rate differences. This makes + splitting sequences tricky because the same event can appear at the end of one sequence and the beginning of + another. + event_end_indices: Corresponding end event index for every audio frame. Used + to ensure when slicing that one chunk ends where the next begins. Should always be true that + event_end_indices[i] = event_start_indices[i + 1]. + state_events: Encoded "state" events representing the encoding state before + each event. + state_event_indices: Corresponding state event index for every audio frame. + """ + indices = np.argsort(event_times, kind="stable") + event_steps = [round(event_times[i] * codec.steps_per_second) for i in indices] + event_values = [event_values[i] for i in indices] + + events = [] + state_events = [] + event_start_indices = [] + state_event_indices = [] + + cur_step = 0 + cur_event_idx = 0 + cur_state_event_idx = 0 + + def fill_event_start_indices_to_cur_step(): + while ( + len(event_start_indices) < len(frame_times) + and frame_times[len(event_start_indices)] < cur_step / codec.steps_per_second + ): + event_start_indices.append(cur_event_idx) + state_event_indices.append(cur_state_event_idx) + + for event_step, event_value in zip(event_steps, event_values): + while event_step > cur_step: + events.append(codec.encode_event(Event(type="shift", value=1))) + cur_step += 1 + fill_event_start_indices_to_cur_step() + cur_event_idx = len(events) + cur_state_event_idx = len(state_events) + if encoding_state_to_events_fn: + # Dump state to state events *before* processing the next event, because + # we want to capture the state prior to the occurrence of the event. + for e in encoding_state_to_events_fn(state): + state_events.append(codec.encode_event(e)) + + for e in encode_event_fn(state, event_value, codec): + events.append(codec.encode_event(e)) + + # After the last event, continue filling out the event_start_indices array. + # The inequality is not strict because if our current step lines up exactly + # with (the start of) an audio frame, we need to add an additional shift event + # to "cover" that frame. + while cur_step / codec.steps_per_second <= frame_times[-1]: + events.append(codec.encode_event(Event(type="shift", value=1))) + cur_step += 1 + fill_event_start_indices_to_cur_step() + cur_event_idx = len(events) + + # Now fill in event_end_indices. We need this extra array to make sure that + # when we slice events, each slice ends exactly where the subsequent slice + # begins. + event_end_indices = event_start_indices[1:] + [len(events)] + + events = np.array(events).astype(np.int32) + state_events = np.array(state_events).astype(np.int32) + event_start_indices = segment(np.array(event_start_indices).astype(np.int32), TARGET_FEATURE_LENGTH) + event_end_indices = segment(np.array(event_end_indices).astype(np.int32), TARGET_FEATURE_LENGTH) + state_event_indices = segment(np.array(state_event_indices).astype(np.int32), TARGET_FEATURE_LENGTH) + + outputs = [] + for start_indices, end_indices, event_indices in zip(event_start_indices, event_end_indices, state_event_indices): + outputs.append( + { + "inputs": events, + "event_start_indices": start_indices, + "event_end_indices": end_indices, + "state_events": state_events, + "state_event_indices": event_indices, + } + ) + + return outputs + + +def extract_sequence_with_indices(features, state_events_end_token=None, feature_key="inputs"): + """Extract target sequence corresponding to audio token segment.""" + features = features.copy() + start_idx = features["event_start_indices"][0] + end_idx = features["event_end_indices"][-1] + + features[feature_key] = features[feature_key][start_idx:end_idx] + + if state_events_end_token is not None: + # Extract the state events corresponding to the audio start token, and + # prepend them to the targets array. + state_event_start_idx = features["state_event_indices"][0] + state_event_end_idx = state_event_start_idx + 1 + while features["state_events"][state_event_end_idx - 1] != state_events_end_token: + state_event_end_idx += 1 + features[feature_key] = np.concatenate( + [ + features["state_events"][state_event_start_idx:state_event_end_idx], + features[feature_key], + ], + axis=0, + ) + + return features + + +def map_midi_programs( + feature, codec: Codec, granularity_type: str = "full", feature_key: str = "inputs" +) -> Mapping[str, Any]: + """Apply MIDI program map to token sequences.""" + granularity = PROGRAM_GRANULARITIES[granularity_type] + + feature[feature_key] = granularity.tokens_map_fn(feature[feature_key], codec) + return feature + + +def run_length_encode_shifts_fn( + features, + codec: Codec, + feature_key: str = "inputs", + state_change_event_types: Sequence[str] = (), +) -> Callable[[Mapping[str, Any]], Mapping[str, Any]]: + """Return a function that run-length encodes shifts for a given codec. + + Args: + codec: The Codec to use for shift events. + feature_key: The feature key for which to run-length encode shifts. + state_change_event_types: A list of event types that represent state + changes; tokens corresponding to these event types will be interpreted as state changes and redundant ones + will be removed. + + Returns: + A preprocessing function that run-length encodes single-step shifts. + """ + state_change_event_ranges = [codec.event_type_range(event_type) for event_type in state_change_event_types] + + def run_length_encode_shifts(features: MutableMapping[str, Any]) -> Mapping[str, Any]: + """Combine leading/interior shifts, trim trailing shifts. + + Args: + features: Dict of features to process. + + Returns: + A dict of features. + """ + events = features[feature_key] + + shift_steps = 0 + total_shift_steps = 0 + output = np.array([], dtype=np.int32) + + current_state = np.zeros(len(state_change_event_ranges), dtype=np.int32) + + for event in events: + if codec.is_shift_event_index(event): + shift_steps += 1 + total_shift_steps += 1 + + else: + # If this event is a state change and has the same value as the current + # state, we can skip it entirely. + is_redundant = False + for i, (min_index, max_index) in enumerate(state_change_event_ranges): + if (min_index <= event) and (event <= max_index): + if current_state[i] == event: + is_redundant = True + current_state[i] = event + if is_redundant: + continue + + # Once we've reached a non-shift event, RLE all previous shift events + # before outputting the non-shift event. + if shift_steps > 0: + shift_steps = total_shift_steps + while shift_steps > 0: + output_steps = np.minimum(codec.max_shift_steps, shift_steps) + output = np.concatenate([output, [output_steps]], axis=0) + shift_steps -= output_steps + output = np.concatenate([output, [event]], axis=0) + + features[feature_key] = output + return features + + return run_length_encode_shifts(features) + + +def note_representation_processor_chain(features, codec: Codec, note_representation_config: NoteRepresentationConfig): + tie_token = codec.encode_event(Event("tie", 0)) + state_events_end_token = tie_token if note_representation_config.include_ties else None + + features = extract_sequence_with_indices( + features, state_events_end_token=state_events_end_token, feature_key="inputs" + ) + + features = map_midi_programs(features, codec) + + features = run_length_encode_shifts_fn(features, codec, state_change_event_types=["velocity", "program"]) + + return features + + +class MidiProcessor: + def __init__(self): + self.codec = Codec( + max_shift_steps=DEFAULT_MAX_SHIFT_SECONDS * DEFAULT_STEPS_PER_SECOND, + steps_per_second=DEFAULT_STEPS_PER_SECOND, + event_ranges=[ + EventRange("pitch", note_seq.MIN_MIDI_PITCH, note_seq.MAX_MIDI_PITCH), + EventRange("velocity", 0, DEFAULT_NUM_VELOCITY_BINS), + EventRange("tie", 0, 0), + EventRange("program", note_seq.MIN_MIDI_PROGRAM, note_seq.MAX_MIDI_PROGRAM), + EventRange("drum", note_seq.MIN_MIDI_PITCH, note_seq.MAX_MIDI_PITCH), + ], + ) + self.tokenizer = Tokenizer(self.codec.num_classes) + self.note_representation_config = NoteRepresentationConfig(onsets_only=False, include_ties=True) + + def __call__(self, midi: bytes | os.PathLike | str): + if not isinstance(midi, bytes): + with open(midi, "rb") as f: + midi = f.read() + + ns = note_seq.midi_to_note_sequence(midi) + ns_sus = note_seq.apply_sustain_control_changes(ns) + + for note in ns_sus.notes: + if not note.is_drum: + note.program = program_to_slakh_program(note.program) + + samples = np.zeros(int(ns_sus.total_time * SAMPLE_RATE)) + + _, frame_times = audio_to_frames(samples, HOP_SIZE, FRAME_RATE) + times, values = note_sequence_to_onsets_and_offsets_and_programs(ns_sus) + + events = encode_and_index_events( + state=NoteEncodingState(), + event_times=times, + event_values=values, + frame_times=frame_times, + codec=self.codec, + encode_event_fn=note_event_data_to_events, + encoding_state_to_events_fn=note_encoding_state_to_events, + ) + + events = [ + note_representation_processor_chain(event, self.codec, self.note_representation_config) for event in events + ] + input_tokens = [self.tokenizer.encode(event["inputs"]) for event in events] + + return input_tokens diff --git a/diffusers/pipelines/deprecated/spectrogram_diffusion/notes_encoder.py b/diffusers/pipelines/deprecated/spectrogram_diffusion/notes_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..a0b0e37e86d49afe0c61fd06bbfbc3b421a8f481 --- /dev/null +++ b/diffusers/pipelines/deprecated/spectrogram_diffusion/notes_encoder.py @@ -0,0 +1,86 @@ +# Copyright 2022 The Music Spectrogram Diffusion Authors. +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +from transformers.modeling_utils import ModuleUtilsMixin +from transformers.models.t5.modeling_t5 import T5Block, T5Config, T5LayerNorm + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models import ModelMixin + + +class SpectrogramNotesEncoder(ModelMixin, ConfigMixin, ModuleUtilsMixin): + @register_to_config + def __init__( + self, + max_length: int, + vocab_size: int, + d_model: int, + dropout_rate: float, + num_layers: int, + num_heads: int, + d_kv: int, + d_ff: int, + feed_forward_proj: str, + is_decoder: bool = False, + ): + super().__init__() + + self.token_embedder = nn.Embedding(vocab_size, d_model) + + self.position_encoding = nn.Embedding(max_length, d_model) + self.position_encoding.weight.requires_grad = False + + self.dropout_pre = nn.Dropout(p=dropout_rate) + + t5config = T5Config( + vocab_size=vocab_size, + d_model=d_model, + num_heads=num_heads, + d_kv=d_kv, + d_ff=d_ff, + dropout_rate=dropout_rate, + feed_forward_proj=feed_forward_proj, + is_decoder=is_decoder, + is_encoder_decoder=False, + ) + + self.encoders = nn.ModuleList() + for lyr_num in range(num_layers): + lyr = T5Block(t5config) + self.encoders.append(lyr) + + self.layer_norm = T5LayerNorm(d_model) + self.dropout_post = nn.Dropout(p=dropout_rate) + + def forward(self, encoder_input_tokens, encoder_inputs_mask): + x = self.token_embedder(encoder_input_tokens) + + seq_length = encoder_input_tokens.shape[1] + inputs_positions = torch.arange(seq_length, device=encoder_input_tokens.device) + x += self.position_encoding(inputs_positions) + + x = self.dropout_pre(x) + + # inverted the attention mask + input_shape = encoder_input_tokens.size() + extended_attention_mask = self.get_extended_attention_mask(encoder_inputs_mask, input_shape) + + for lyr in self.encoders: + x = lyr(x, extended_attention_mask)[0] + x = self.layer_norm(x) + + return self.dropout_post(x), encoder_inputs_mask diff --git a/diffusers/pipelines/deprecated/spectrogram_diffusion/pipeline_spectrogram_diffusion.py b/diffusers/pipelines/deprecated/spectrogram_diffusion/pipeline_spectrogram_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..ead87da95c87539171f540e1c798cfb99c0e0dc1 --- /dev/null +++ b/diffusers/pipelines/deprecated/spectrogram_diffusion/pipeline_spectrogram_diffusion.py @@ -0,0 +1,271 @@ +# Copyright 2022 The Music Spectrogram Diffusion Authors. +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from typing import Any, Callable + +import numpy as np +import torch + +from ....models import T5FilmDecoder +from ....schedulers import DDPMScheduler +from ....utils import is_onnx_available, logging +from ....utils.torch_utils import randn_tensor + + +if is_onnx_available(): + from ...onnx_utils import OnnxRuntimeModel + +from ...pipeline_utils import AudioPipelineOutput, DiffusionPipeline +from .continuous_encoder import SpectrogramContEncoder +from .notes_encoder import SpectrogramNotesEncoder + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +TARGET_FEATURE_LENGTH = 256 + + +class SpectrogramDiffusionPipeline(DiffusionPipeline): + r""" + Pipeline for unconditional audio generation. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + notes_encoder ([`SpectrogramNotesEncoder`]): + continuous_encoder ([`SpectrogramContEncoder`]): + decoder ([`T5FilmDecoder`]): + A [`T5FilmDecoder`] to denoise the encoded audio latents. + scheduler ([`DDPMScheduler`]): + A scheduler to be used in combination with `decoder` to denoise the encoded audio latents. + melgan ([`OnnxRuntimeModel`]): + """ + + _optional_components = ["melgan"] + + def __init__( + self, + notes_encoder: SpectrogramNotesEncoder, + continuous_encoder: SpectrogramContEncoder, + decoder: T5FilmDecoder, + scheduler: DDPMScheduler, + melgan: OnnxRuntimeModel if is_onnx_available() else Any, + ) -> None: + super().__init__() + + # From MELGAN + self.min_value = math.log(1e-5) # Matches MelGAN training. + self.max_value = 4.0 # Largest value for most examples + self.n_dims = 128 + + self.register_modules( + notes_encoder=notes_encoder, + continuous_encoder=continuous_encoder, + decoder=decoder, + scheduler=scheduler, + melgan=melgan, + ) + + def scale_features(self, features, output_range=(-1.0, 1.0), clip=False): + """Linearly scale features to network outputs range.""" + min_out, max_out = output_range + if clip: + features = torch.clip(features, self.min_value, self.max_value) + # Scale to [0, 1]. + zero_one = (features - self.min_value) / (self.max_value - self.min_value) + # Scale to [min_out, max_out]. + return zero_one * (max_out - min_out) + min_out + + def scale_to_features(self, outputs, input_range=(-1.0, 1.0), clip=False): + """Invert by linearly scaling network outputs to features range.""" + min_out, max_out = input_range + outputs = torch.clip(outputs, min_out, max_out) if clip else outputs + # Scale to [0, 1]. + zero_one = (outputs - min_out) / (max_out - min_out) + # Scale to [self.min_value, self.max_value]. + return zero_one * (self.max_value - self.min_value) + self.min_value + + def encode(self, input_tokens, continuous_inputs, continuous_mask): + tokens_mask = input_tokens > 0 + tokens_encoded, tokens_mask = self.notes_encoder( + encoder_input_tokens=input_tokens, encoder_inputs_mask=tokens_mask + ) + + continuous_encoded, continuous_mask = self.continuous_encoder( + encoder_inputs=continuous_inputs, encoder_inputs_mask=continuous_mask + ) + + return [(tokens_encoded, tokens_mask), (continuous_encoded, continuous_mask)] + + def decode(self, encodings_and_masks, input_tokens, noise_time): + timesteps = noise_time + if not torch.is_tensor(timesteps): + timesteps = torch.tensor([timesteps], dtype=torch.long, device=input_tokens.device) + elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0: + timesteps = timesteps[None].to(input_tokens.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps * torch.ones(input_tokens.shape[0], dtype=timesteps.dtype, device=timesteps.device) + + logits = self.decoder( + encodings_and_masks=encodings_and_masks, decoder_input_tokens=input_tokens, decoder_noise_time=timesteps + ) + return logits + + @torch.no_grad() + def __call__( + self, + input_tokens: list[list[int]], + generator: torch.Generator | None = None, + num_inference_steps: int = 100, + return_dict: bool = True, + output_type: str = "np", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ) -> AudioPipelineOutput | tuple: + r""" + The call function to the pipeline for generation. + + Args: + input_tokens (`list[list[int]]`): + The tokenized MIDI inputs to generate audio from. Each element is a list of integer tokens produced by + the `MidiProcessor`. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality audio at the + expense of slower inference. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.AudioPipelineOutput`] instead of a plain tuple. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated audio. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Example: + + ```py + >>> from diffusers import SpectrogramDiffusionPipeline, MidiProcessor + + >>> pipe = SpectrogramDiffusionPipeline.from_pretrained("google/music-spectrogram-diffusion") + >>> pipe = pipe.to("cuda") + >>> processor = MidiProcessor() + + >>> # Download MIDI from: wget http://www.piano-midi.de/midis/beethoven/beethoven_hammerklavier_2.mid + >>> output = pipe(processor("beethoven_hammerklavier_2.mid")) + + >>> audio = output.audios[0] + ``` + + Returns: + [`pipelines.AudioPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`pipelines.AudioPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated audio. + """ + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + pred_mel = np.zeros([1, TARGET_FEATURE_LENGTH, self.n_dims], dtype=np.float32) + full_pred_mel = np.zeros([1, 0, self.n_dims], np.float32) + ones = torch.ones((1, TARGET_FEATURE_LENGTH), dtype=bool, device=self.device) + + for i, encoder_input_tokens in enumerate(input_tokens): + if i == 0: + encoder_continuous_inputs = torch.from_numpy(pred_mel[:1].copy()).to( + device=self.device, dtype=self.decoder.dtype + ) + # The first chunk has no previous context. + encoder_continuous_mask = torch.zeros((1, TARGET_FEATURE_LENGTH), dtype=bool, device=self.device) + else: + # The full song pipeline does not feed in a context feature, so the mask + # will be all 0s after the feature converter. Because we know we're + # feeding in a full context chunk from the previous prediction, set it + # to all 1s. + encoder_continuous_mask = ones + + encoder_continuous_inputs = self.scale_features( + encoder_continuous_inputs, output_range=[-1.0, 1.0], clip=True + ) + + encodings_and_masks = self.encode( + input_tokens=torch.IntTensor([encoder_input_tokens]).to(device=self.device), + continuous_inputs=encoder_continuous_inputs, + continuous_mask=encoder_continuous_mask, + ) + + # Sample encoder_continuous_inputs shaped gaussian noise to begin loop + x = randn_tensor( + shape=encoder_continuous_inputs.shape, + generator=generator, + device=self.device, + dtype=self.decoder.dtype, + ) + + # set step values + self.scheduler.set_timesteps(num_inference_steps) + + # Denoising diffusion loop + for j, t in enumerate(self.progress_bar(self.scheduler.timesteps)): + output = self.decode( + encodings_and_masks=encodings_and_masks, + input_tokens=x, + noise_time=t / self.scheduler.config.num_train_timesteps, # rescale to [0, 1) + ) + + # Compute previous output: x_t -> x_t-1 + x = self.scheduler.step(output, t, x, generator=generator).prev_sample + + mel = self.scale_to_features(x, input_range=[-1.0, 1.0]) + encoder_continuous_inputs = mel[:1] + pred_mel = mel.cpu().float().numpy() + + full_pred_mel = np.concatenate([full_pred_mel, pred_mel[:1]], axis=1) + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, full_pred_mel) + + logger.info("Generated segment", i) + + if output_type == "np" and not is_onnx_available(): + raise ValueError( + "Cannot return output in 'np' format if ONNX is not available. Make sure to have ONNX installed or set 'output_type' to 'mel'." + ) + elif output_type == "np" and self.melgan is None: + raise ValueError( + "Cannot return output in 'np' format if melgan component is not defined. Make sure to define `self.melgan` or set 'output_type' to 'mel'." + ) + + if output_type == "np": + output = self.melgan(input_features=full_pred_mel.astype(np.float32)) + else: + output = full_pred_mel + + if not return_dict: + return (output,) + + return AudioPipelineOutput(audios=output) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_attend_and_excite/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_attend_and_excite/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..2087f09ea580c3a3b47c70e1ae0f2dac7af7f022 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_attend_and_excite/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py b/diffusers/pipelines/deprecated/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py new file mode 100644 index 0000000000000000000000000000000000000000..6a9e031f8d1478444c8d9e0bbe606acda06f3b6b --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py @@ -0,0 +1,1113 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import numpy as np +import torch +from torch.nn import functional as F +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....image_processor import VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.attention_processor import Attention +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionAttendAndExcitePipeline + + >>> pipe = StableDiffusionAttendAndExcitePipeline.from_pretrained( + ... "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16 + ... ).to("cuda") + + + >>> prompt = "a cat and a frog" + + >>> # use get_indices function to find out indices of the tokens you want to alter + >>> pipe.get_indices(prompt) + {0: '<|startoftext|>', 1: 'a', 2: 'cat', 3: 'and', 4: 'a', 5: 'frog', 6: '<|endoftext|>'} + + >>> token_indices = [2, 5] + >>> seed = 6141 + >>> generator = torch.Generator("cuda").manual_seed(seed) + + >>> images = pipe( + ... prompt=prompt, + ... token_indices=token_indices, + ... guidance_scale=7.5, + ... generator=generator, + ... num_inference_steps=50, + ... max_iter_to_alter=25, + ... ).images + + >>> image = images[0] + >>> image.save(f"../images/{prompt}_{seed}.png") + ``` +""" + + +class AttentionStore: + @staticmethod + def get_empty_store(): + return {"down": [], "mid": [], "up": []} + + def __call__(self, attn, is_cross: bool, place_in_unet: str): + if self.cur_att_layer >= 0 and is_cross: + if attn.shape[1] == np.prod(self.attn_res): + self.step_store[place_in_unet].append(attn) + + self.cur_att_layer += 1 + if self.cur_att_layer == self.num_att_layers: + self.cur_att_layer = 0 + self.between_steps() + + def between_steps(self): + self.attention_store = self.step_store + self.step_store = self.get_empty_store() + + def get_average_attention(self): + average_attention = self.attention_store + return average_attention + + def aggregate_attention(self, from_where: list[str]) -> torch.Tensor: + """Aggregates the attention across the different layers and heads at the specified resolution.""" + out = [] + attention_maps = self.get_average_attention() + for location in from_where: + for item in attention_maps[location]: + cross_maps = item.reshape(-1, self.attn_res[0], self.attn_res[1], item.shape[-1]) + out.append(cross_maps) + out = torch.cat(out, dim=0) + out = out.sum(0) / out.shape[0] + return out + + def reset(self): + self.cur_att_layer = 0 + self.step_store = self.get_empty_store() + self.attention_store = {} + + def __init__(self, attn_res): + """ + Initialize an empty AttentionStore :param step_index: used to visualize only a specific step in the diffusion + process + """ + self.num_att_layers = -1 + self.cur_att_layer = 0 + self.step_store = self.get_empty_store() + self.attention_store = {} + self.curr_step_index = 0 + self.attn_res = attn_res + + +class AttendExciteAttnProcessor: + def __init__(self, attnstore, place_in_unet): + super().__init__() + self.attnstore = attnstore + self.place_in_unet = place_in_unet + + def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + query = attn.to_q(hidden_states) + + is_cross = encoder_hidden_states is not None + encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + + # only need to store attention maps during the Attend and Excite process + if attention_probs.requires_grad: + self.attnstore(attention_probs, is_cross, self.place_in_unet) + + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +class StableDiffusionAttendAndExcitePipeline( + DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin +): + r""" + Pipeline for text-to-image generation using Stable Diffusion and Attend-and-Excite. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + indices, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + indices_is_list_ints = isinstance(indices, list) and isinstance(indices[0], int) + indices_is_list_list_ints = ( + isinstance(indices, list) and isinstance(indices[0], list) and isinstance(indices[0][0], int) + ) + + if not indices_is_list_ints and not indices_is_list_list_ints: + raise TypeError("`indices` must be a list of ints or a list of a list of ints") + + if indices_is_list_ints: + indices_batch_size = 1 + elif indices_is_list_list_ints: + indices_batch_size = len(indices) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if indices_batch_size != prompt_batch_size: + raise ValueError( + f"indices batch size must be same as prompt batch size. indices batch size: {indices_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @staticmethod + def _compute_max_attention_per_index( + attention_maps: torch.Tensor, + indices: list[int], + ) -> list[torch.Tensor]: + """Computes the maximum attention value for each of the tokens we wish to alter.""" + attention_for_text = attention_maps[:, :, 1:-1] + attention_for_text *= 100 + attention_for_text = torch.nn.functional.softmax(attention_for_text, dim=-1) + + # Shift indices since we removed the first token + indices = [index - 1 for index in indices] + + # Extract the maximum values + max_indices_list = [] + for i in indices: + image = attention_for_text[:, :, i] + smoothing = GaussianSmoothing().to(attention_maps.device) + input = F.pad(image.unsqueeze(0).unsqueeze(0), (1, 1, 1, 1), mode="reflect") + image = smoothing(input).squeeze(0).squeeze(0) + max_indices_list.append(image.max()) + return max_indices_list + + def _aggregate_and_get_max_attention_per_token( + self, + indices: list[int], + ): + """Aggregates the attention for each token and computes the max activation value for each token to alter.""" + attention_maps = self.attention_store.aggregate_attention( + from_where=("up", "down", "mid"), + ) + max_attention_per_index = self._compute_max_attention_per_index( + attention_maps=attention_maps, + indices=indices, + ) + return max_attention_per_index + + @staticmethod + def _compute_loss(max_attention_per_index: list[torch.Tensor]) -> torch.Tensor: + """Computes the attend-and-excite loss using the maximum attention value for each token.""" + losses = [max(0, 1.0 - curr_max) for curr_max in max_attention_per_index] + loss = max(losses) + return loss + + @staticmethod + def _update_latent(latents: torch.Tensor, loss: torch.Tensor, step_size: float) -> torch.Tensor: + """Update the latent according to the computed loss.""" + grad_cond = torch.autograd.grad(loss.requires_grad_(True), [latents], retain_graph=True)[0] + latents = latents - step_size * grad_cond + return latents + + def _perform_iterative_refinement_step( + self, + latents: torch.Tensor, + indices: list[int], + loss: torch.Tensor, + threshold: float, + text_embeddings: torch.Tensor, + step_size: float, + t: int, + max_refinement_steps: int = 20, + ): + """ + Performs the iterative latent refinement introduced in the paper. Here, we continuously update the latent code + according to our loss objective until the given threshold is reached for all tokens. + """ + iteration = 0 + target_loss = max(0, 1.0 - threshold) + while loss > target_loss: + iteration += 1 + + latents = latents.clone().detach().requires_grad_(True) + self.unet(latents, t, encoder_hidden_states=text_embeddings).sample + self.unet.zero_grad() + + # Get max activation value for each subject token + max_attention_per_index = self._aggregate_and_get_max_attention_per_token( + indices=indices, + ) + + loss = self._compute_loss(max_attention_per_index) + + if loss != 0: + latents = self._update_latent(latents, loss, step_size) + + logger.info(f"\t Try {iteration}. loss: {loss}") + + if iteration >= max_refinement_steps: + logger.info(f"\t Exceeded max number of iterations ({max_refinement_steps})! ") + break + + # Run one more time but don't compute gradients and update the latents. + # We just need to compute the new loss - the grad update will occur below + latents = latents.clone().detach().requires_grad_(True) + _ = self.unet(latents, t, encoder_hidden_states=text_embeddings).sample + self.unet.zero_grad() + + # Get max activation value for each subject token + max_attention_per_index = self._aggregate_and_get_max_attention_per_token( + indices=indices, + ) + loss = self._compute_loss(max_attention_per_index) + logger.info(f"\t Finished with loss of: {loss}") + return loss, latents, max_attention_per_index + + def register_attention_control(self): + attn_procs = {} + cross_att_count = 0 + for name in self.unet.attn_processors.keys(): + if name.startswith("mid_block"): + place_in_unet = "mid" + elif name.startswith("up_blocks"): + place_in_unet = "up" + elif name.startswith("down_blocks"): + place_in_unet = "down" + else: + continue + + cross_att_count += 1 + attn_procs[name] = AttendExciteAttnProcessor(attnstore=self.attention_store, place_in_unet=place_in_unet) + + self.unet.set_attn_processor(attn_procs) + self.attention_store.num_att_layers = cross_att_count + + def get_indices(self, prompt: str) -> dict[str, int]: + """Utility function to list the indices of the tokens you wish to alte""" + ids = self.tokenizer(prompt).input_ids + indices = {i: tok for tok, i in zip(self.tokenizer.convert_ids_to_tokens(ids), range(len(ids)))} + return indices + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + token_indices: list[int] | list[list[int]], + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + max_iter_to_alter: int = 25, + thresholds: dict = {0: 0.05, 10: 0.5, 20: 0.8}, + scale_factor: int = 20, + attn_res: tuple[int] | None = (16, 16), + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + token_indices (`list[int]`): + The token indices to alter with attend-and-excite. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + max_iter_to_alter (`int`, *optional*, defaults to `25`): + Number of denoising steps to apply attend-and-excite. The `max_iter_to_alter` denoising steps are when + attend-and-excite is applied. For example, if `max_iter_to_alter` is `25` and there are a total of `30` + denoising steps, the first `25` denoising steps applies attend-and-excite and the last `5` will not. + thresholds (`dict`, *optional*, defaults to `{0: 0.05, 10: 0.5, 20: 0.8}`): + Dictionary defining the iterations and desired thresholds to apply iterative latent refinement in. + scale_factor (`int`, *optional*, default to 20): + Scale factor to control the step size of each attend-and-excite update. + attn_res (`tuple`, *optional*, default computed from width and height): + The 2D resolution of the semantic attention map. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + token_indices, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + if attn_res is None: + attn_res = int(np.ceil(width / 32)), int(np.ceil(height / 32)) + self.attention_store = AttentionStore(attn_res) + original_attn_proc = self.unet.attn_processors + self.register_attention_control() + + # default config for step size from original repo + scale_range = np.linspace(1.0, 0.5, len(self.scheduler.timesteps)) + step_size = scale_factor * np.sqrt(scale_range) + + text_embeddings = ( + prompt_embeds[batch_size * num_images_per_prompt :] if do_classifier_free_guidance else prompt_embeds + ) + + if isinstance(token_indices[0], int): + token_indices = [token_indices] + + indices = [] + + for ind in token_indices: + indices = indices + [ind] * num_images_per_prompt + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Attend and excite process + with torch.enable_grad(): + latents = latents.clone().detach().requires_grad_(True) + updated_latents = [] + for latent, index, text_embedding in zip(latents, indices, text_embeddings): + # Forward pass of denoising with text conditioning + latent = latent.unsqueeze(0) + text_embedding = text_embedding.unsqueeze(0) + + self.unet( + latent, + t, + encoder_hidden_states=text_embedding, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + self.unet.zero_grad() + + # Get max activation value for each subject token + max_attention_per_index = self._aggregate_and_get_max_attention_per_token( + indices=index, + ) + + loss = self._compute_loss(max_attention_per_index=max_attention_per_index) + + # If this is an iterative refinement step, verify we have reached the desired threshold for all + if i in thresholds.keys() and loss > 1.0 - thresholds[i]: + loss, latent, max_attention_per_index = self._perform_iterative_refinement_step( + latents=latent, + indices=index, + loss=loss, + threshold=thresholds[i], + text_embeddings=text_embedding, + step_size=step_size[i], + t=t, + ) + + # Perform gradient update + if i < max_iter_to_alter: + if loss != 0: + latent = self._update_latent( + latents=latent, + loss=loss, + step_size=step_size[i], + ) + logger.info(f"Iteration {i} | Loss: {loss:0.4f}") + + updated_latents.append(latent) + + latents = torch.cat(updated_latents, dim=0) + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post-processing + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + +class GaussianSmoothing(torch.nn.Module): + """ + Arguments: + Apply gaussian smoothing on a 1d, 2d or 3d tensor. Filtering is performed separately for each channel in the input + using a depthwise convolution. + channels (int, sequence): Number of channels of the input tensors. Output will + have this number of channels as well. + kernel_size (int, sequence): Size of the gaussian kernel. sigma (float, sequence): Standard deviation of the + gaussian kernel. dim (int, optional): The number of dimensions of the data. + Default value is 2 (spatial). + """ + + # channels=1, kernel_size=kernel_size, sigma=sigma, dim=2 + def __init__( + self, + channels: int = 1, + kernel_size: int = 3, + sigma: float = 0.5, + dim: int = 2, + ): + super().__init__() + + if isinstance(kernel_size, int): + kernel_size = [kernel_size] * dim + if isinstance(sigma, float): + sigma = [sigma] * dim + + # The gaussian kernel is the product of the + # gaussian function of each dimension. + kernel = 1 + meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size]) + for size, std, mgrid in zip(kernel_size, sigma, meshgrids): + mean = (size - 1) / 2 + kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2)) + + # Make sure sum of values in gaussian kernel equals 1. + kernel = kernel / torch.sum(kernel) + + # Reshape to depthwise convolutional weight + kernel = kernel.view(1, 1, *kernel.size()) + kernel = kernel.repeat(channels, *[1] * (kernel.dim() - 1)) + + self.register_buffer("weight", kernel) + self.groups = channels + + if dim == 1: + self.conv = F.conv1d + elif dim == 2: + self.conv = F.conv2d + elif dim == 3: + self.conv = F.conv3d + else: + raise RuntimeError("Only 1, 2 and 3 dimensions are supported. Received {}.".format(dim)) + + def forward(self, input): + """ + Arguments: + Apply gaussian filter to input. + input (torch.Tensor): Input to apply gaussian filter on. + Returns: + filtered (torch.Tensor): Filtered output. + """ + return self.conv(input, weight=self.weight.to(input.dtype), groups=self.groups) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_diffedit/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_diffedit/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3924c610274c96febf9fe43a488424b1b984506d --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_diffedit/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_diffedit import StableDiffusionDiffEditPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py b/diffusers/pipelines/deprecated/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py new file mode 100644 index 0000000000000000000000000000000000000000..ee8675678f2d2b5b5a2553b975835d21e578b743 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py @@ -0,0 +1,1548 @@ +# Copyright 2025 DiffEdit Authors and Pix2Pix Zero Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....configuration_utils import FrozenDict +from ....image_processor import VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import DDIMInverseScheduler, KarrasDiffusionSchedulers +from ....utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class DiffEditInversionPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + latents (`torch.Tensor`) + inverted latents tensor + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `num_timesteps * batch_size` or numpy array of shape `(num_timesteps, + batch_size, height, width, num_channels)`. PIL images or numpy array present the denoised images of the + diffusion pipeline. + """ + + latents: torch.Tensor + images: list[PIL.Image.Image] | np.ndarray + + +EXAMPLE_DOC_STRING = """ + + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionDiffEditPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png" + + >>> init_image = download_image(img_url).resize((768, 768)) + + >>> pipeline = StableDiffusionDiffEditPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16 + ... ) + + >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.enable_model_cpu_offload() + + >>> mask_prompt = "A bowl of fruits" + >>> prompt = "A bowl of pears" + + >>> mask_image = pipeline.generate_mask(image=init_image, source_prompt=prompt, target_prompt=mask_prompt) + >>> image_latents = pipeline.invert(image=init_image, prompt=mask_prompt).latents + >>> image = pipeline(prompt=prompt, mask_image=mask_image, image_latents=image_latents).images[0] + ``` +""" + +EXAMPLE_INVERT_DOC_STRING = """ + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionDiffEditPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png" + + >>> init_image = download_image(img_url).resize((768, 768)) + + >>> pipeline = StableDiffusionDiffEditPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16 + ... ) + + >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.enable_model_cpu_offload() + + >>> prompt = "A bowl of fruits" + + >>> inverted_latents = pipeline.invert(image=init_image, prompt=prompt).latents + ``` +""" + + +def auto_corr_loss(hidden_states, generator=None): + reg_loss = 0.0 + for i in range(hidden_states.shape[0]): + for j in range(hidden_states.shape[1]): + noise = hidden_states[i : i + 1, j : j + 1, :, :] + while True: + roll_amount = torch.randint(noise.shape[2] // 2, (1,), generator=generator).item() + reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=2)).mean() ** 2 + reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=3)).mean() ** 2 + + if noise.shape[2] <= 8: + break + noise = torch.nn.functional.avg_pool2d(noise, kernel_size=2) + return reg_loss + + +def kl_divergence(hidden_states): + return hidden_states.var() + hidden_states.mean() ** 2 - 1 - torch.log(hidden_states.var() + 1e-7) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +def preprocess_mask(mask, batch_size: int = 1): + if not isinstance(mask, torch.Tensor): + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list): + if isinstance(mask[0], PIL.Image.Image): + mask = [np.array(m.convert("L")).astype(np.float32) / 255.0 for m in mask] + if isinstance(mask[0], np.ndarray): + mask = np.stack(mask, axis=0) if mask[0].ndim < 3 else np.concatenate(mask, axis=0) + mask = torch.from_numpy(mask) + elif isinstance(mask[0], torch.Tensor): + mask = torch.stack(mask, dim=0) if mask[0].ndim < 3 else torch.cat(mask, dim=0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + # Check mask shape + if batch_size > 1: + if mask.shape[0] == 1: + mask = torch.cat([mask] * batch_size) + elif mask.shape[0] > 1 and mask.shape[0] != batch_size: + raise ValueError( + f"`mask_image` with batch size {mask.shape[0]} cannot be broadcasted to batch size {batch_size} " + f"inferred by prompt inputs" + ) + + if mask.shape[1] != 1: + raise ValueError(f"`mask_image` must have 1 channel, but has {mask.shape[1]} channels") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("`mask_image` should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + return mask + + +class StableDiffusionDiffEditPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, +): + r""" + > [!WARNING] > This is an experimental feature! + + Pipeline for text-guided image inpainting using Stable Diffusion and DiffEdit. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading and saving methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + inverse_scheduler ([`DDIMInverseScheduler`]): + A scheduler to be used in combination with `unet` to fill in the unmasked part of the input latents. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "inverse_scheduler"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + inverse_scheduler: DDIMInverseScheduler, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "skip_prk_steps", True) is False: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration" + " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make" + " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to" + " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face" + " Hub, it would be very nice if you could open a Pull request for the" + " `scheduler/scheduler_config.json` file" + ) + deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["skip_prk_steps"] = True + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + inverse_scheduler=inverse_scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if (strength is None) or (strength is not None and (strength < 0 or strength > 1)): + raise ValueError( + f"The value of `strength` should in [0.0, 1.0] but is, but is {strength} of type {type(strength)}." + ) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def check_source_inputs( + self, + source_prompt=None, + source_negative_prompt=None, + source_prompt_embeds=None, + source_negative_prompt_embeds=None, + ): + if source_prompt is not None and source_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `source_prompt`: {source_prompt} and `source_prompt_embeds`: {source_prompt_embeds}." + " Please make sure to only forward one of the two." + ) + elif source_prompt is None and source_prompt_embeds is None: + raise ValueError( + "Provide either `source_image` or `source_prompt_embeds`. Cannot leave all both of the arguments undefined." + ) + elif source_prompt is not None and ( + not isinstance(source_prompt, str) and not isinstance(source_prompt, list) + ): + raise ValueError(f"`source_prompt` has to be of type `str` or `list` but is {type(source_prompt)}") + + if source_negative_prompt is not None and source_negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `source_negative_prompt`: {source_negative_prompt} and `source_negative_prompt_embeds`:" + f" {source_negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if source_prompt_embeds is not None and source_negative_prompt_embeds is not None: + if source_prompt_embeds.shape != source_negative_prompt_embeds.shape: + raise ValueError( + "`source_prompt_embeds` and `source_negative_prompt_embeds` must have the same shape when passed" + f" directly, but got: `source_prompt_embeds` {source_prompt_embeds.shape} !=" + f" `source_negative_prompt_embeds` {source_negative_prompt_embeds.shape}." + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + + return timesteps, num_inference_steps - t_start + + def get_inverse_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + + # safety for t_start overflow to prevent empty timsteps slice + if t_start == 0: + return self.inverse_scheduler.timesteps, num_inference_steps + timesteps = self.inverse_scheduler.timesteps[:-t_start] + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_latents(self, image, batch_size, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if isinstance(generator, list): + latents = [ + self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + latents = torch.cat(latents, dim=0) + else: + latents = self.vae.encode(image).latent_dist.sample(generator) + + latents = self.vae.config.scaling_factor * latents + + if batch_size != latents.shape[0]: + if batch_size % latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_latents_per_image = batch_size // latents.shape[0] + latents = torch.cat([latents] * additional_latents_per_image, dim=0) + else: + raise ValueError( + f"Cannot duplicate `image` of batch size {latents.shape[0]} to {batch_size} text prompts." + ) + else: + latents = torch.cat([latents], dim=0) + + return latents + + def get_epsilon(self, model_output: torch.Tensor, sample: torch.Tensor, timestep: int): + pred_type = self.inverse_scheduler.config.prediction_type + alpha_prod_t = self.inverse_scheduler.alphas_cumprod[timestep] + + beta_prod_t = 1 - alpha_prod_t + + if pred_type == "epsilon": + return model_output + elif pred_type == "sample": + return (sample - alpha_prod_t ** (0.5) * model_output) / beta_prod_t ** (0.5) + elif pred_type == "v_prediction": + return (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {pred_type} must be one of `epsilon`, `sample`, or `v_prediction`" + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def generate_mask( + self, + image: torch.Tensor | PIL.Image.Image = None, + target_prompt: str | list[str] | None = None, + target_negative_prompt: str | list[str] | None = None, + target_prompt_embeds: torch.Tensor | None = None, + target_negative_prompt_embeds: torch.Tensor | None = None, + source_prompt: str | list[str] | None = None, + source_negative_prompt: str | list[str] | None = None, + source_prompt_embeds: torch.Tensor | None = None, + source_negative_prompt_embeds: torch.Tensor | None = None, + num_maps_per_mask: int | None = 10, + mask_encode_strength: float | None = 0.5, + mask_thresholding_ratio: float | None = 3.0, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "np", + cross_attention_kwargs: dict[str, Any] | None = None, + ): + r""" + Generate a latent mask given a mask prompt, a target prompt, and an image. + + Args: + image (`PIL.Image.Image`): + `Image` or tensor representing an image batch to be used for computing the mask. + target_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide semantic mask generation. If not defined, you need to pass + `prompt_embeds`. + target_negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + target_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + target_negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + source_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide semantic mask generation using DiffEdit. If not defined, you need to + pass `source_prompt_embeds` or `source_image` instead. + source_negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide semantic mask generation away from using DiffEdit. If not defined, you + need to pass `source_negative_prompt_embeds` or `source_image` instead. + source_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings to guide the semantic mask generation. Can be used to easily tweak text + inputs (prompt weighting). If not provided, text embeddings are generated from `source_prompt` input + argument. + source_negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings to negatively guide the semantic mask generation. Can be used to easily + tweak text inputs (prompt weighting). If not provided, text embeddings are generated from + `source_negative_prompt` input argument. + num_maps_per_mask (`int`, *optional*, defaults to 10): + The number of noise maps sampled to generate the semantic mask using DiffEdit. + mask_encode_strength (`float`, *optional*, defaults to 0.5): + The strength of the noise maps sampled to generate the semantic mask using DiffEdit. Must be between 0 + and 1. + mask_thresholding_ratio (`float`, *optional*, defaults to 3.0): + The maximum multiple of the mean absolute difference used to clamp the semantic guidance map before + mask binarization. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the + [`~models.attention_processor.AttnProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + + Examples: + + Returns: + `list[PIL.Image.Image]` or `np.array`: + When returning a `list[PIL.Image.Image]`, the list consists of a batch of single-channel binary images + with dimensions `(height // self.vae_scale_factor, width // self.vae_scale_factor)`. If it's + `np.array`, the shape is `(batch_size, height // self.vae_scale_factor, width // + self.vae_scale_factor)`. + """ + + # 1. Check inputs (Provide dummy argument for callback_steps) + self.check_inputs( + target_prompt, + mask_encode_strength, + 1, + target_negative_prompt, + target_prompt_embeds, + target_negative_prompt_embeds, + ) + + self.check_source_inputs( + source_prompt, + source_negative_prompt, + source_prompt_embeds, + source_negative_prompt_embeds, + ) + + if (num_maps_per_mask is None) or ( + num_maps_per_mask is not None and (not isinstance(num_maps_per_mask, int) or num_maps_per_mask <= 0) + ): + raise ValueError( + f"`num_maps_per_mask` has to be a positive integer but is {num_maps_per_mask} of type" + f" {type(num_maps_per_mask)}." + ) + + if mask_thresholding_ratio is None or mask_thresholding_ratio <= 0: + raise ValueError( + f"`mask_thresholding_ratio` has to be positive but is {mask_thresholding_ratio} of type" + f" {type(mask_thresholding_ratio)}." + ) + + # 2. Define call parameters + if target_prompt is not None and isinstance(target_prompt, str): + batch_size = 1 + elif target_prompt is not None and isinstance(target_prompt, list): + batch_size = len(target_prompt) + else: + batch_size = target_prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompts + (cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None) + target_negative_prompt_embeds, target_prompt_embeds = self.encode_prompt( + target_prompt, + device, + num_maps_per_mask, + do_classifier_free_guidance, + target_negative_prompt, + prompt_embeds=target_prompt_embeds, + negative_prompt_embeds=target_negative_prompt_embeds, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + target_prompt_embeds = torch.cat([target_negative_prompt_embeds, target_prompt_embeds]) + + source_negative_prompt_embeds, source_prompt_embeds = self.encode_prompt( + source_prompt, + device, + num_maps_per_mask, + do_classifier_free_guidance, + source_negative_prompt, + prompt_embeds=source_prompt_embeds, + negative_prompt_embeds=source_negative_prompt_embeds, + ) + if do_classifier_free_guidance: + source_prompt_embeds = torch.cat([source_negative_prompt_embeds, source_prompt_embeds]) + + # 4. Preprocess image + image = self.image_processor.preprocess(image).repeat_interleave(num_maps_per_mask, dim=0) + + # 5. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, _ = self.get_timesteps(num_inference_steps, mask_encode_strength, device) + encode_timestep = timesteps[0] + + # 6. Prepare image latents and add noise with specified strength + image_latents = self.prepare_image_latents( + image, batch_size * num_maps_per_mask, self.vae.dtype, device, generator + ) + noise = randn_tensor(image_latents.shape, generator=generator, device=device, dtype=self.vae.dtype) + image_latents = self.scheduler.add_noise(image_latents, noise, encode_timestep) + + latent_model_input = torch.cat([image_latents] * (4 if do_classifier_free_guidance else 2)) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, encode_timestep) + + # 7. Predict the noise residual + prompt_embeds = torch.cat([source_prompt_embeds, target_prompt_embeds]) + noise_pred = self.unet( + latent_model_input, + encode_timestep, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + if do_classifier_free_guidance: + noise_pred_neg_src, noise_pred_source, noise_pred_uncond, noise_pred_target = noise_pred.chunk(4) + noise_pred_source = noise_pred_neg_src + guidance_scale * (noise_pred_source - noise_pred_neg_src) + noise_pred_target = noise_pred_uncond + guidance_scale * (noise_pred_target - noise_pred_uncond) + else: + noise_pred_source, noise_pred_target = noise_pred.chunk(2) + + # 8. Compute the mask from the absolute difference of predicted noise residuals + # TODO: Consider smoothing mask guidance map + mask_guidance_map = ( + torch.abs(noise_pred_target - noise_pred_source) + .reshape(batch_size, num_maps_per_mask, *noise_pred_target.shape[-3:]) + .mean([1, 2]) + ) + clamp_magnitude = mask_guidance_map.mean() * mask_thresholding_ratio + semantic_mask_image = mask_guidance_map.clamp(0, clamp_magnitude) / clamp_magnitude + semantic_mask_image = torch.where(semantic_mask_image <= 0.5, 0, 1) + mask_image = semantic_mask_image.cpu().numpy() + + # 9. Convert to Numpy array or PIL. + if output_type == "pil": + mask_image = self.image_processor.numpy_to_pil(mask_image) + + # Offload all models + self.maybe_free_model_hooks() + + return mask_image + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_INVERT_DOC_STRING) + def invert( + self, + prompt: str | list[str] | None = None, + image: torch.Tensor | PIL.Image.Image = None, + num_inference_steps: int = 50, + inpaint_strength: float = 0.8, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + decode_latents: bool = False, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int | None = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + lambda_auto_corr: float = 20.0, + lambda_kl: float = 20.0, + num_reg_steps: int = 0, + num_auto_corr_rolls: int = 5, + ): + r""" + Generate inverted latents given a prompt and image. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`PIL.Image.Image`): + `Image` or tensor representing an image batch to produce the inverted latents guided by `prompt`. + inpaint_strength (`float`, *optional*, defaults to 0.8): + Indicates extent of the noising process to run latent inversion. Must be between 0 and 1. When + `inpaint_strength` is 1, the inversion process is run for the full number of iterations specified in + `num_inference_steps`. `image` is used as a reference for the inversion process, and adding more noise + increases `inpaint_strength`. If `inpaint_strength` is 0, no inpainting occurs. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + decode_latents (`bool`, *optional*, defaults to `False`): + Whether or not to decode the inverted latents into a generated image. Setting this argument to `True` + decodes all inverted latents for each timestep into a list of generated images. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.DiffEditInversionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the + [`~models.attention_processor.AttnProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + lambda_auto_corr (`float`, *optional*, defaults to 20.0): + Lambda parameter to control auto correction. + lambda_kl (`float`, *optional*, defaults to 20.0): + Lambda parameter to control Kullback-Leibler divergence output. + num_reg_steps (`int`, *optional*, defaults to 0): + Number of regularization loss steps. + num_auto_corr_rolls (`int`, *optional*, defaults to 5): + Number of auto correction roll steps. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.pipeline_stable_diffusion_diffedit.DiffEditInversionPipelineOutput`] or + `tuple`: + If `return_dict` is `True`, + [`~pipelines.stable_diffusion.pipeline_stable_diffusion_diffedit.DiffEditInversionPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is the inverted latents tensors + ordered by increasing noise, and the second is the corresponding decoded images if `decode_latents` is + `True`, otherwise `None`. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + inpaint_strength, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Preprocess image + image = self.image_processor.preprocess(image) + + # 4. Prepare latent variables + num_images_per_prompt = 1 + latents = self.prepare_image_latents( + image, batch_size * num_images_per_prompt, self.vae.dtype, device, generator + ) + + # 5. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 6. Prepare timesteps + self.inverse_scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_inverse_timesteps(num_inference_steps, inpaint_strength, device) + + # 7. Noising loop where we obtain the intermediate noised latent image for each timestep. + num_warmup_steps = len(timesteps) - num_inference_steps * self.inverse_scheduler.order + inverted_latents = [] + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.inverse_scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # regularization of the noise prediction (not in original code or paper but borrowed from Pix2PixZero) + if num_reg_steps > 0: + with torch.enable_grad(): + for _ in range(num_reg_steps): + if lambda_auto_corr > 0: + for _ in range(num_auto_corr_rolls): + var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True) + + # Derive epsilon from model output before regularizing to IID standard normal + var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t) + + l_ac = auto_corr_loss(var_epsilon, generator=generator) + l_ac.backward() + + grad = var.grad.detach() / num_auto_corr_rolls + noise_pred = noise_pred - lambda_auto_corr * grad + + if lambda_kl > 0: + var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True) + + # Derive epsilon from model output before regularizing to IID standard normal + var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t) + + l_kld = kl_divergence(var_epsilon) + l_kld.backward() + + grad = var.grad.detach() + noise_pred = noise_pred - lambda_kl * grad + + noise_pred = noise_pred.detach() + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.inverse_scheduler.step(noise_pred, t, latents).prev_sample + inverted_latents.append(latents.detach().clone()) + + # call the callback, if provided + if i == len(timesteps) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.inverse_scheduler.order == 0 + ): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + assert len(inverted_latents) == len(timesteps) + latents = torch.stack(list(reversed(inverted_latents)), 1) + + # 8. Post-processing + image = None + if decode_latents: + image = self.decode_latents(latents.flatten(0, 1)) + + # 9. Convert to PIL. + if decode_latents and output_type == "pil": + image = self.image_processor.numpy_to_pil(image) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (latents, image) + + return DiffEditInversionPipelineOutput(latents=latents, images=image) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + mask_image: torch.Tensor | PIL.Image.Image = None, + image_latents: torch.Tensor | PIL.Image.Image = None, + inpaint_strength: float | None = 0.8, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + mask_image (`PIL.Image.Image`): + `Image` or tensor representing an image batch to mask the generated image. White pixels in the mask are + repainted, while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, 1, H, W)`. + image_latents (`PIL.Image.Image` or `torch.Tensor`): + Partially noised image latents from the inversion process to be used as inputs for image generation. + inpaint_strength (`float`, *optional*, defaults to 0.8): + Indicates extent to inpaint the masked area. Must be between 0 and 1. When `inpaint_strength` is 1, the + denoising process is run on the masked area for the full number of iterations specified in + `num_inference_steps`. `image_latents` is used as a reference for the masked area, and adding more + noise to a region increases `inpaint_strength`. If `inpaint_strength` is 0, no inpainting occurs. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + inpaint_strength, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + if mask_image is None: + raise ValueError( + "`mask_image` input cannot be undefined. Use `generate_mask()` to compute `mask_image` from text prompts." + ) + if image_latents is None: + raise ValueError( + "`image_latents` input cannot be undefined. Use `invert()` to compute `image_latents` from input images." + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Preprocess mask + mask_image = preprocess_mask(mask_image, batch_size) + latent_height, latent_width = mask_image.shape[-2:] + mask_image = torch.cat([mask_image] * num_images_per_prompt) + mask_image = mask_image.to(device=device, dtype=prompt_embeds.dtype) + + # 5. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, inpaint_strength, device) + + # 6. Preprocess image latents + if isinstance(image_latents, list) and any(isinstance(l, torch.Tensor) and l.ndim == 5 for l in image_latents): + image_latents = torch.cat(image_latents).detach() + elif isinstance(image_latents, torch.Tensor) and image_latents.ndim == 5: + image_latents = image_latents.detach() + else: + image_latents = self.image_processor.preprocess(image_latents).detach() + + latent_shape = (self.vae.config.latent_channels, latent_height, latent_width) + if image_latents.shape[-3:] != latent_shape: + raise ValueError( + f"Each latent image in `image_latents` must have shape {latent_shape}, " + f"but has shape {image_latents.shape[-3:]}" + ) + if image_latents.ndim == 4: + image_latents = image_latents.reshape(batch_size, len(timesteps), *latent_shape) + if image_latents.shape[:2] != (batch_size, len(timesteps)): + raise ValueError( + f"`image_latents` must have batch size {batch_size} with latent images from {len(timesteps)}" + f" timesteps, but has batch size {image_latents.shape[0]} with latent images from" + f" {image_latents.shape[1]} timesteps." + ) + image_latents = image_latents.transpose(0, 1).repeat_interleave(num_images_per_prompt, dim=1) + image_latents = image_latents.to(device=device, dtype=prompt_embeds.dtype) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + latents = image_latents[0].clone() + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # mask with inverted latents from appropriate timestep - use original image latent for last step + latents = latents * mask_image + image_latents[i] * (1 - mask_image) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_gligen/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_gligen/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..81c8b8b99cd83de1f007e3a2c3a0108b7d5bbe9d --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_gligen/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_gligen"] = ["StableDiffusionGLIGENPipeline"] + _import_structure["pipeline_stable_diffusion_gligen_text_image"] = ["StableDiffusionGLIGENTextImagePipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_gligen import StableDiffusionGLIGENPipeline + from .pipeline_stable_diffusion_gligen_text_image import StableDiffusionGLIGENTextImagePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py b/diffusers/pipelines/deprecated/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py new file mode 100644 index 0000000000000000000000000000000000000000..38f5af842e1b07a1beaf6c241143e1c7300bcf45 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py @@ -0,0 +1,861 @@ +# Copyright 2025 The GLIGEN Authors and HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import warnings +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....image_processor import VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.attention import GatedSelfAttentionDense +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionGLIGENPipeline + >>> from diffusers.utils import load_image + + >>> # Insert objects described by text at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENPipeline.from_pretrained( + ... "masterful/gligen-1-4-inpainting-text-box", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> input_image = load_image( + ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/livingroom_modern.png" + ... ) + >>> prompt = "a birthday cake" + >>> boxes = [[0.2676, 0.6088, 0.4773, 0.7183]] + >>> phrases = ["a birthday cake"] + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_inpaint_image=input_image, + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-1-4-inpainting-text-box.jpg") + + >>> # Generate an image described by the prompt and + >>> # insert objects described by text at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENPipeline.from_pretrained( + ... "masterful/gligen-1-4-generation-text-box", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a waterfall and a modern high speed train running through the tunnel in a beautiful forest with fall foliage" + >>> boxes = [[0.1387, 0.2051, 0.4277, 0.7090], [0.4980, 0.4355, 0.8516, 0.7266]] + >>> phrases = ["a waterfall", "a modern high speed train running through the tunnel"] + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-1-4-generation-text-box.jpg") + ``` +""" + + +class StableDiffusionGLIGENPipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + _optional_components = ["safety_checker", "feature_extractor"] + model_cpu_offload_seq = "text_encoder->unet->vae" + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + gligen_phrases, + gligen_boxes, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if len(gligen_phrases) != len(gligen_boxes): + raise ValueError( + "length of `gligen_phrases` and `gligen_boxes` has to be same, but" + f" got: `gligen_phrases` {len(gligen_phrases)} != `gligen_boxes` {len(gligen_boxes)}" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def enable_fuser(self, enabled=True): + for module in self.unet.modules(): + if type(module) is GatedSelfAttentionDense: + module.enabled = enabled + + def draw_inpaint_mask_from_boxes(self, boxes, size): + inpaint_mask = torch.ones(size[0], size[1]) + for box in boxes: + x0, x1 = box[0] * size[0], box[2] * size[0] + y0, y1 = box[1] * size[1], box[3] * size[1] + inpaint_mask[int(y0) : int(y1), int(x0) : int(x1)] = 0 + return inpaint_mask + + def crop(self, im, new_width, new_height): + width, height = im.size + left = (width - new_width) / 2 + top = (height - new_height) / 2 + right = (width + new_width) / 2 + bottom = (height + new_height) / 2 + return im.crop((left, top, right, bottom)) + + def target_size_center_crop(self, im, new_hw): + width, height = im.size + if width != height: + im = self.crop(im, min(height, width), min(height, width)) + return im.resize((new_hw, new_hw), PIL.Image.LANCZOS) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + gligen_scheduled_sampling_beta: float = 0.3, + gligen_phrases: list[str] = None, + gligen_boxes: list[list[float]] = None, + gligen_inpaint_image: PIL.Image.Image | None = None, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + gligen_phrases (`list[str]`): + The phrases to guide what to include in each of the regions defined by the corresponding + `gligen_boxes`. There should only be one phrase per bounding box. + gligen_boxes (`list[list[float]]`): + The bounding boxes that identify rectangular regions of the image that are going to be filled with the + content described by the corresponding `gligen_phrases`. Each rectangular box is defined as a + `list[float]` of 4 elements `[xmin, ymin, xmax, ymax]` where each value is between [0,1]. + gligen_inpaint_image (`PIL.Image.Image`, *optional*): + The input image, if provided, is inpainted with objects described by the `gligen_boxes` and + `gligen_phrases`. Otherwise, it is treated as a generation task on a blank input image. + gligen_scheduled_sampling_beta (`float`, defaults to 0.3): + Scheduled Sampling factor from [GLIGEN: Open-Set Grounded Text-to-Image + Generation](https://huggingface.co/papers/2301.07093). Scheduled Sampling factor is only varied for + scheduled sampling during inference for improved quality and controllability. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + gligen_phrases, + gligen_boxes, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5.1 Prepare GLIGEN variables + max_objs = 30 + if len(gligen_boxes) > max_objs: + warnings.warn( + f"More that {max_objs} objects found. Only first {max_objs} objects will be processed.", + FutureWarning, + ) + gligen_phrases = gligen_phrases[:max_objs] + gligen_boxes = gligen_boxes[:max_objs] + # prepare batched input to the GLIGENTextBoundingboxProjection (boxes, phrases, mask) + # Get tokens for phrases from pre-trained CLIPTokenizer + tokenizer_inputs = self.tokenizer(gligen_phrases, padding=True, return_tensors="pt").to(device) + # For the token, we use the same pre-trained text encoder + # to obtain its text feature + _text_embeddings = self.text_encoder(**tokenizer_inputs).pooler_output + n_objs = len(gligen_boxes) + # For each entity, described in phrases, is denoted with a bounding box, + # we represent the location information as (xmin,ymin,xmax,ymax) + boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype) + boxes[:n_objs] = torch.tensor(gligen_boxes) + text_embeddings = torch.zeros( + max_objs, self.unet.config.cross_attention_dim, device=device, dtype=self.text_encoder.dtype + ) + text_embeddings[:n_objs] = _text_embeddings + # Generate a mask for each object that is entity described by phrases + masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + masks[:n_objs] = 1 + + repeat_batch = batch_size * num_images_per_prompt + boxes = boxes.unsqueeze(0).expand(repeat_batch, -1, -1).clone() + text_embeddings = text_embeddings.unsqueeze(0).expand(repeat_batch, -1, -1).clone() + masks = masks.unsqueeze(0).expand(repeat_batch, -1).clone() + if do_classifier_free_guidance: + repeat_batch = repeat_batch * 2 + boxes = torch.cat([boxes] * 2) + text_embeddings = torch.cat([text_embeddings] * 2) + masks = torch.cat([masks] * 2) + masks[: repeat_batch // 2] = 0 + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + cross_attention_kwargs["gligen"] = {"boxes": boxes, "positive_embeddings": text_embeddings, "masks": masks} + + # Prepare latent variables for GLIGEN inpainting + if gligen_inpaint_image is not None: + # if the given input image is not of the same size as expected by VAE + # center crop and resize the input image to expected shape + if gligen_inpaint_image.size != (self.vae.sample_size, self.vae.sample_size): + gligen_inpaint_image = self.target_size_center_crop(gligen_inpaint_image, self.vae.sample_size) + # Convert a single image into a batch of images with a batch size of 1 + # The resulting shape becomes (1, C, H, W), where C is the number of channels, + # and H and W are the height and width of the image. + # scales the pixel values to a range [-1, 1] + gligen_inpaint_image = self.image_processor.preprocess(gligen_inpaint_image) + gligen_inpaint_image = gligen_inpaint_image.to(dtype=self.vae.dtype, device=self.vae.device) + # Run AutoEncoder to get corresponding latents + gligen_inpaint_latent = self.vae.encode(gligen_inpaint_image).latent_dist.sample() + gligen_inpaint_latent = self.vae.config.scaling_factor * gligen_inpaint_latent + # Generate an inpainting mask + # pixel value = 0, where the object is present (defined by bounding boxes above) + # 1, everywhere else + gligen_inpaint_mask = self.draw_inpaint_mask_from_boxes(gligen_boxes, gligen_inpaint_latent.shape[2:]) + gligen_inpaint_mask = gligen_inpaint_mask.to( + dtype=gligen_inpaint_latent.dtype, device=gligen_inpaint_latent.device + ) + gligen_inpaint_mask = gligen_inpaint_mask[None, None] + gligen_inpaint_mask_addition = torch.cat( + (gligen_inpaint_latent * gligen_inpaint_mask, gligen_inpaint_mask), dim=1 + ) + # Convert a single mask into a batch of masks with a batch size of 1 + gligen_inpaint_mask_addition = gligen_inpaint_mask_addition.expand(repeat_batch, -1, -1, -1).clone() + + num_grounding_steps = int(gligen_scheduled_sampling_beta * len(timesteps)) + self.enable_fuser(True) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Scheduled sampling + if i == num_grounding_steps: + self.enable_fuser(False) + + if latents.shape[1] != 4: + latents = torch.randn_like(latents[:, :4]) + + if gligen_inpaint_image is not None: + gligen_inpaint_latent_with_noise = ( + self.scheduler.add_noise( + gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t]) + ) + .expand(latents.shape[0], -1, -1, -1) + .clone() + ) + latents = gligen_inpaint_latent_with_noise * gligen_inpaint_mask + latents * ( + 1 - gligen_inpaint_mask + ) + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if gligen_inpaint_image is not None: + latent_model_input = torch.cat((latent_model_input, gligen_inpaint_mask_addition), dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py b/diffusers/pipelines/deprecated/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py new file mode 100644 index 0000000000000000000000000000000000000000..d72d12a64945224bcb1de60cb27f627296293221 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py @@ -0,0 +1,1053 @@ +# Copyright 2025 The GLIGEN Authors and HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import warnings +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ....image_processor import VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.attention import GatedSelfAttentionDense +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion import StableDiffusionPipelineOutput +from ...stable_diffusion.clip_image_project_model import CLIPImageProjection +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionGLIGENTextImagePipeline + >>> from diffusers.utils import load_image + + >>> # Insert objects described by image at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained( + ... "anhnct/Gligen_Inpainting_Text_Image", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> input_image = load_image( + ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/livingroom_modern.png" + ... ) + >>> prompt = "a backpack" + >>> boxes = [[0.2676, 0.4088, 0.4773, 0.7183]] + >>> phrases = None + >>> gligen_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/backpack.jpeg" + ... ) + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_inpaint_image=input_image, + ... gligen_boxes=boxes, + ... gligen_images=[gligen_image], + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-inpainting-text-image-box.jpg") + + >>> # Generate an image described by the prompt and + >>> # insert objects described by text and image at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained( + ... "anhnct/Gligen_Text_Image", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a flower sitting on the beach" + >>> boxes = [[0.0, 0.09, 0.53, 0.76]] + >>> phrases = ["flower"] + >>> gligen_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/pexels-pixabay-60597.jpg" + ... ) + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_images=[gligen_image], + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-generation-text-image-box.jpg") + + >>> # Generate an image described by the prompt and + >>> # transfer style described by image at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained( + ... "anhnct/Gligen_Text_Image", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a dragon flying on the sky" + >>> boxes = [[0.4, 0.2, 1.0, 0.8], [0.0, 1.0, 0.0, 1.0]] # Set `[0.0, 1.0, 0.0, 1.0]` for the style + + >>> gligen_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/landscape.png" + ... ) + + >>> gligen_placeholder = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/landscape.png" + ... ) + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=[ + ... "dragon", + ... "placeholder", + ... ], # Can use any text instead of `placeholder` token, because we will use mask here + ... gligen_images=[ + ... gligen_placeholder, + ... gligen_image, + ... ], # Can use any image in gligen_placeholder, because we will use mask here + ... input_phrases_mask=[1, 0], # Set 0 for the placeholder token + ... input_images_mask=[0, 1], # Set 0 for the placeholder image + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-generation-text-image-box-style-transfer.jpg") + ``` +""" + + +class StableDiffusionGLIGENTextImagePipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + processor ([`~transformers.CLIPProcessor`]): + A `CLIPProcessor` to process reference image. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + image_project ([`CLIPImageProjection`]): + A `CLIPImageProjection` to project image embedding into phrases embedding space. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + processor: CLIPProcessor, + image_encoder: CLIPVisionModelWithProjection, + image_project: CLIPImageProjection, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + image_encoder=image_encoder, + processor=processor, + image_project=image_project, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + gligen_images, + gligen_phrases, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if gligen_images is not None and gligen_phrases is not None: + if len(gligen_images) != len(gligen_phrases): + raise ValueError( + "`gligen_images` and `gligen_phrases` must have the same length when both are provided, but" + f" got: `gligen_images` with length {len(gligen_images)} != `gligen_phrases` with length {len(gligen_phrases)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def enable_fuser(self, enabled=True): + for module in self.unet.modules(): + if type(module) is GatedSelfAttentionDense: + module.enabled = enabled + + def draw_inpaint_mask_from_boxes(self, boxes, size): + """ + Create an inpainting mask based on given boxes. This function generates an inpainting mask using the provided + boxes to mark regions that need to be inpainted. + """ + inpaint_mask = torch.ones(size[0], size[1]) + for box in boxes: + x0, x1 = box[0] * size[0], box[2] * size[0] + y0, y1 = box[1] * size[1], box[3] * size[1] + inpaint_mask[int(y0) : int(y1), int(x0) : int(x1)] = 0 + return inpaint_mask + + def crop(self, im, new_width, new_height): + """ + Crop the input image to the specified dimensions. + """ + width, height = im.size + left = (width - new_width) / 2 + top = (height - new_height) / 2 + right = (width + new_width) / 2 + bottom = (height + new_height) / 2 + return im.crop((left, top, right, bottom)) + + def target_size_center_crop(self, im, new_hw): + """ + Crop and resize the image to the target size while keeping the center. + """ + width, height = im.size + if width != height: + im = self.crop(im, min(height, width), min(height, width)) + return im.resize((new_hw, new_hw), PIL.Image.LANCZOS) + + def complete_mask(self, has_mask, max_objs, device): + """ + Based on the input mask corresponding value `0 or 1` for each phrases and image, mask the features + corresponding to phrases and images. + """ + mask = torch.ones(1, max_objs).type(self.text_encoder.dtype).to(device) + if has_mask is None: + return mask + + if isinstance(has_mask, int): + return mask * has_mask + else: + for idx, value in enumerate(has_mask): + mask[0, idx] = value + return mask + + def get_clip_feature(self, input, normalize_constant, device, is_image=False): + """ + Get image and phrases embedding by using CLIP pretrain model. The image embedding is transformed into the + phrases embedding space through a projection. + """ + if is_image: + if input is None: + return None + inputs = self.processor(images=[input], return_tensors="pt").to(device) + inputs["pixel_values"] = inputs["pixel_values"].to(self.image_encoder.dtype) + + outputs = self.image_encoder(**inputs) + feature = outputs.image_embeds + feature = self.image_project(feature).squeeze(0) + feature = (feature / feature.norm()) * normalize_constant + feature = feature.unsqueeze(0) + else: + if input is None: + return None + inputs = self.tokenizer(input, return_tensors="pt", padding=True).to(device) + outputs = self.text_encoder(**inputs) + feature = outputs.pooler_output + return feature + + def get_cross_attention_kwargs_with_grounded( + self, + hidden_size, + gligen_phrases, + gligen_images, + gligen_boxes, + input_phrases_mask, + input_images_mask, + repeat_batch, + normalize_constant, + max_objs, + device, + ): + """ + Prepare the cross-attention kwargs containing information about the grounded input (boxes, mask, image + embedding, phrases embedding). + """ + phrases, images = gligen_phrases, gligen_images + images = [None] * len(phrases) if images is None else images + phrases = [None] * len(images) if phrases is None else phrases + + boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype) + masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + image_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + image_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + + text_features = [] + image_features = [] + for phrase, image in zip(phrases, images): + text_features.append(self.get_clip_feature(phrase, normalize_constant, device, is_image=False)) + image_features.append(self.get_clip_feature(image, normalize_constant, device, is_image=True)) + + for idx, (box, text_feature, image_feature) in enumerate(zip(gligen_boxes, text_features, image_features)): + boxes[idx] = torch.tensor(box) + masks[idx] = 1 + if text_feature is not None: + phrases_embeddings[idx] = text_feature + phrases_masks[idx] = 1 + if image_feature is not None: + image_embeddings[idx] = image_feature + image_masks[idx] = 1 + + input_phrases_mask = self.complete_mask(input_phrases_mask, max_objs, device) + phrases_masks = phrases_masks.unsqueeze(0).repeat(repeat_batch, 1) * input_phrases_mask + input_images_mask = self.complete_mask(input_images_mask, max_objs, device) + image_masks = image_masks.unsqueeze(0).repeat(repeat_batch, 1) * input_images_mask + boxes = boxes.unsqueeze(0).repeat(repeat_batch, 1, 1) + masks = masks.unsqueeze(0).repeat(repeat_batch, 1) + phrases_embeddings = phrases_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1) + image_embeddings = image_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1) + + out = { + "boxes": boxes, + "masks": masks, + "phrases_masks": phrases_masks, + "image_masks": image_masks, + "phrases_embeddings": phrases_embeddings, + "image_embeddings": image_embeddings, + } + + return out + + def get_cross_attention_kwargs_without_grounded(self, hidden_size, repeat_batch, max_objs, device): + """ + Prepare the cross-attention kwargs without information about the grounded input (boxes, mask, image embedding, + phrases embedding) (All are zero tensor). + """ + boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype) + masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + image_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + image_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + + out = { + "boxes": boxes.unsqueeze(0).repeat(repeat_batch, 1, 1), + "masks": masks.unsqueeze(0).repeat(repeat_batch, 1), + "phrases_masks": phrases_masks.unsqueeze(0).repeat(repeat_batch, 1), + "image_masks": image_masks.unsqueeze(0).repeat(repeat_batch, 1), + "phrases_embeddings": phrases_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1), + "image_embeddings": image_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1), + } + + return out + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + gligen_scheduled_sampling_beta: float = 0.3, + gligen_phrases: list[str] = None, + gligen_images: list[PIL.Image.Image] = None, + input_phrases_mask: int | list[int] = None, + input_images_mask: int | list[int] = None, + gligen_boxes: list[list[float]] = None, + gligen_inpaint_image: PIL.Image.Image | None = None, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + gligen_normalize_constant: float = 28.7, + clip_skip: int = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + gligen_phrases (`list[str]`): + The phrases to guide what to include in each of the regions defined by the corresponding + `gligen_boxes`. There should only be one phrase per bounding box. + gligen_images (`list[PIL.Image.Image]`): + The images to guide what to include in each of the regions defined by the corresponding `gligen_boxes`. + There should only be one image per bounding box + input_phrases_mask (`int` or `list[int]`): + pre phrases mask input defined by the correspongding `input_phrases_mask` + input_images_mask (`int` or `list[int]`): + pre images mask input defined by the correspongding `input_images_mask` + gligen_boxes (`list[list[float]]`): + The bounding boxes that identify rectangular regions of the image that are going to be filled with the + content described by the corresponding `gligen_phrases`. Each rectangular box is defined as a + `list[float]` of 4 elements `[xmin, ymin, xmax, ymax]` where each value is between [0,1]. + gligen_inpaint_image (`PIL.Image.Image`, *optional*): + The input image, if provided, is inpainted with objects described by the `gligen_boxes` and + `gligen_phrases`. Otherwise, it is treated as a generation task on a blank input image. + gligen_scheduled_sampling_beta (`float`, defaults to 0.3): + Scheduled Sampling factor from [GLIGEN: Open-Set Grounded Text-to-Image + Generation](https://huggingface.co/papers/2301.07093). Scheduled Sampling factor is only varied for + scheduled sampling during inference for improved quality and controllability. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + gligen_normalize_constant (`float`, *optional*, defaults to 28.7): + The normalize value of the image embedding. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + gligen_images, + gligen_phrases, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5.1 Prepare GLIGEN variables + max_objs = 30 + if len(gligen_boxes) > max_objs: + warnings.warn( + f"More that {max_objs} objects found. Only first {max_objs} objects will be processed.", + FutureWarning, + ) + gligen_phrases = gligen_phrases[:max_objs] + gligen_boxes = gligen_boxes[:max_objs] + gligen_images = gligen_images[:max_objs] + + repeat_batch = batch_size * num_images_per_prompt + + if do_classifier_free_guidance: + repeat_batch = repeat_batch * 2 + + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + hidden_size = prompt_embeds.shape[2] + + cross_attention_kwargs["gligen"] = self.get_cross_attention_kwargs_with_grounded( + hidden_size=hidden_size, + gligen_phrases=gligen_phrases, + gligen_images=gligen_images, + gligen_boxes=gligen_boxes, + input_phrases_mask=input_phrases_mask, + input_images_mask=input_images_mask, + repeat_batch=repeat_batch, + normalize_constant=gligen_normalize_constant, + max_objs=max_objs, + device=device, + ) + + cross_attention_kwargs_without_grounded = {} + cross_attention_kwargs_without_grounded["gligen"] = self.get_cross_attention_kwargs_without_grounded( + hidden_size=hidden_size, repeat_batch=repeat_batch, max_objs=max_objs, device=device + ) + + # Prepare latent variables for GLIGEN inpainting + if gligen_inpaint_image is not None: + # if the given input image is not of the same size as expected by VAE + # center crop and resize the input image to expected shape + if gligen_inpaint_image.size != (self.vae.sample_size, self.vae.sample_size): + gligen_inpaint_image = self.target_size_center_crop(gligen_inpaint_image, self.vae.sample_size) + # Convert a single image into a batch of images with a batch size of 1 + # The resulting shape becomes (1, C, H, W), where C is the number of channels, + # and H and W are the height and width of the image. + # scales the pixel values to a range [-1, 1] + gligen_inpaint_image = self.image_processor.preprocess(gligen_inpaint_image) + gligen_inpaint_image = gligen_inpaint_image.to(dtype=self.vae.dtype, device=self.vae.device) + # Run AutoEncoder to get corresponding latents + gligen_inpaint_latent = self.vae.encode(gligen_inpaint_image).latent_dist.sample() + gligen_inpaint_latent = self.vae.config.scaling_factor * gligen_inpaint_latent + # Generate an inpainting mask + # pixel value = 0, where the object is present (defined by bounding boxes above) + # 1, everywhere else + gligen_inpaint_mask = self.draw_inpaint_mask_from_boxes(gligen_boxes, gligen_inpaint_latent.shape[2:]) + gligen_inpaint_mask = gligen_inpaint_mask.to( + dtype=gligen_inpaint_latent.dtype, device=gligen_inpaint_latent.device + ) + gligen_inpaint_mask = gligen_inpaint_mask[None, None] + gligen_inpaint_mask_addition = torch.cat( + (gligen_inpaint_latent * gligen_inpaint_mask, gligen_inpaint_mask), dim=1 + ) + # Convert a single mask into a batch of masks with a batch size of 1 + gligen_inpaint_mask_addition = gligen_inpaint_mask_addition.expand(repeat_batch, -1, -1, -1).clone() + + int(gligen_scheduled_sampling_beta * len(timesteps)) + self.enable_fuser(True) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if latents.shape[1] != 4: + latents = torch.randn_like(latents[:, :4]) + + if gligen_inpaint_image is not None: + gligen_inpaint_latent_with_noise = ( + self.scheduler.add_noise( + gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t]) + ) + .expand(latents.shape[0], -1, -1, -1) + .clone() + ) + latents = gligen_inpaint_latent_with_noise * gligen_inpaint_mask + latents * ( + 1 - gligen_inpaint_mask + ) + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if gligen_inpaint_image is not None: + latent_model_input = torch.cat((latent_model_input, gligen_inpaint_mask_addition), dim=1) + + # predict the noise residual with grounded information + noise_pred_with_grounding = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # predict the noise residual without grounded information + noise_pred_without_grounding = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs_without_grounded, + ).sample + + # perform guidance + if do_classifier_free_guidance: + # Using noise_pred_text from noise residual with grounded information and noise_pred_uncond from noise residual without grounded information + _, noise_pred_text = noise_pred_with_grounding.chunk(2) + noise_pred_uncond, _ = noise_pred_without_grounding.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + else: + noise_pred = noise_pred_with_grounding + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_ldm3d/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_ldm3d/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a2fcf3ab8369684d8b8c9b6e09fef2207ed124e0 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_ldm3d/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_ldm3d"] = ["StableDiffusionLDM3DPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_ldm3d import StableDiffusionLDM3DPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py b/diffusers/pipelines/deprecated/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py new file mode 100644 index 0000000000000000000000000000000000000000..70a16f5d522f6ed2d78a3b5ac04444d3cc9c2cf8 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py @@ -0,0 +1,1049 @@ +# Copyright 2025 The Intel Labs Team Authors and the HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ....image_processor import PipelineImageInput, VaeImageProcessorLDM3D +from ....loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ....models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> from diffusers import StableDiffusionLDM3DPipeline + + >>> pipe = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c") + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> output = pipe(prompt) + >>> rgb_image, depth_image = output.rgb, output.depth + >>> rgb_image[0].save("astronaut_ldm3d_rgb.jpg") + >>> depth_image[0].save("astronaut_ldm3d_depth.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +@dataclass +class LDM3DPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + rgb (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + depth (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + nsfw_content_detected (`list[bool]`) + list indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or + `None` if safety checking could not be performed. + """ + + rgb: list[PIL.Image.Image] | np.ndarray + depth: list[PIL.Image.Image] | np.ndarray + nsfw_content_detected: list[bool] | None + + +class StableDiffusionLDM3DPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + _last_supported_version = "0.33.1" + + r""" + Pipeline for text-to-image and 3D generation using LDM3D. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection | None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessorLDM3D(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + rgb_feature_extractor_input = feature_extractor_input[0] + safety_checker_input = self.feature_extractor(rgb_feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 49, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + rgb, depth = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return ((rgb, depth), has_nsfw_concept) + + return LDM3DPipelineOutput(rgb=rgb, depth=depth, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_panorama/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_panorama/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ce0601ed2649a89d7fbbfc67f5ea5f561d80cbd3 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_panorama/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_panorama"] = ["StableDiffusionPanoramaPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_panorama import StableDiffusionPanoramaPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py b/diffusers/pipelines/deprecated/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py new file mode 100644 index 0000000000000000000000000000000000000000..481c9c93dddec826d0c0754c9d831e412a937e86 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py @@ -0,0 +1,1201 @@ +# Copyright 2025 MultiDiffusion Authors and The HuggingFace Team. All rights reserved." +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ....image_processor import PipelineImageInput, VaeImageProcessor +from ....loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import DDIMScheduler +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionPanoramaPipeline, DDIMScheduler + + >>> model_ckpt = "stabilityai/stable-diffusion-2-base" + >>> scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler") + >>> pipe = StableDiffusionPanoramaPipeline.from_pretrained( + ... model_ckpt, scheduler=scheduler, torch_dtype=torch.float16 + ... ) + + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of the dolomites" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPanoramaPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, +): + _last_supported_version = "0.33.1" + + r""" + Pipeline for text-to-image generation using MultiDiffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection | None = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def decode_latents_with_padding(self, latents: torch.Tensor, padding: int = 8) -> torch.Tensor: + """ + Decode the given latents with padding for circular inference. + + Args: + latents (torch.Tensor): The input latents to decode. + padding (int, optional): The number of latents to add on each side for padding. Defaults to 8. + + Returns: + torch.Tensor: The decoded image with padding removed. + + Notes: + - The padding is added to remove boundary artifacts and improve the output quality. + - This would slightly increase the memory usage. + - The padding pixels are then removed from the decoded image. + + """ + latents = 1 / self.vae.config.scaling_factor * latents + latents_left = latents[..., :padding] + latents_right = latents[..., -padding:] + latents = torch.cat((latents_right, latents, latents_left), axis=-1) + image = self.vae.decode(latents, return_dict=False)[0] + padding_pix = self.vae_scale_factor * padding + image = image[..., padding_pix:-padding_pix] + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + def get_views( + self, + panorama_height: int, + panorama_width: int, + window_size: int = 64, + stride: int = 8, + circular_padding: bool = False, + ) -> list[tuple[int, int, int, int]]: + """ + Generates a list of views based on the given parameters. Here, we define the mappings F_i (see Eq. 7 in the + MultiDiffusion paper https://huggingface.co/papers/2302.08113). If panorama's height/width < window_size, + num_blocks of height/width should return 1. + + Args: + panorama_height (int): The height of the panorama. + panorama_width (int): The width of the panorama. + window_size (int, optional): The size of the window. Defaults to 64. + stride (int, optional): The stride value. Defaults to 8. + circular_padding (bool, optional): Whether to apply circular padding. Defaults to False. + + Returns: + list[tuple[int, int, int, int]]: A list of tuples representing the views. Each tuple contains four integers + representing the start and end coordinates of the window in the panorama. + + """ + panorama_height /= 8 + panorama_width /= 8 + num_blocks_height = (panorama_height - window_size) // stride + 1 if panorama_height > window_size else 1 + if circular_padding: + num_blocks_width = panorama_width // stride if panorama_width > window_size else 1 + else: + num_blocks_width = (panorama_width - window_size) // stride + 1 if panorama_width > window_size else 1 + total_num_blocks = int(num_blocks_height * num_blocks_width) + views = [] + for i in range(total_num_blocks): + h_start = int((i // num_blocks_width) * stride) + h_end = h_start + window_size + w_start = int((i % num_blocks_width) * stride) + w_end = w_start + window_size + views.append((h_start, h_end, w_start, w_end)) + return views + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def clip_skip(self): + return self._clip_skip + + @property + def do_classifier_free_guidance(self): + return False + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = 512, + width: int | None = 2048, + num_inference_steps: int = 50, + timesteps: list[int] = None, + guidance_scale: float = 7.5, + view_batch_size: int = 1, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + circular_padding: bool = False, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs: Any, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 2048): + The width in pixels of the generated image. The width is kept high because the pipeline is supposed + generate panorama-like images. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + The timesteps at which to generate the images. If not specified, then the default timestep spacing + strategy of the scheduler is used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + view_batch_size (`int`, *optional*, defaults to 1): + The batch size to denoise split views. For some GPUs with high performance, higher view batch size can + speedup the generation and increase the VRAM usage. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescaling factor for the guidance embeddings. A value of 0.0 means no rescaling is applied. + circular_padding (`bool`, *optional*, defaults to `False`): + If set to `True`, circular padding is applied to ensure there are no stitching artifacts. Circular + padding allows the model to seamlessly generate a transition from the rightmost part of the image to + the leftmost part, maintaining consistency in a 360-degree sense. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Define panorama grid and initialize views for synthesis. + # prepare batch grid + views = self.get_views(height, width, circular_padding=circular_padding) + views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)] + views_scheduler_status = [copy.deepcopy(self.scheduler.__dict__)] * len(views_batch) + count = torch.zeros_like(latents) + value = torch.zeros_like(latents) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 7.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 8. Denoising loop + # Each denoising step also includes refinement of the latents with respect to the + # views. + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + count.zero_() + value.zero_() + + # generate views + # Here, we iterate through different spatial crops of the latents and denoise them. These + # denoised (latent) crops are then averaged to produce the final latent + # for the current timestep via MultiDiffusion. Please see Sec. 4.1 in the + # MultiDiffusion paper for more details: https://huggingface.co/papers/2302.08113 + # Batch views denoise + for j, batch_view in enumerate(views_batch): + vb_size = len(batch_view) + # get the latents corresponding to the current view coordinates + if circular_padding: + latents_for_view = [] + for h_start, h_end, w_start, w_end in batch_view: + if w_end > latents.shape[3]: + # Add circular horizontal padding + latent_view = torch.cat( + ( + latents[:, :, h_start:h_end, w_start:], + latents[:, :, h_start:h_end, : w_end - latents.shape[3]], + ), + axis=-1, + ) + else: + latent_view = latents[:, :, h_start:h_end, w_start:w_end] + latents_for_view.append(latent_view) + latents_for_view = torch.cat(latents_for_view) + else: + latents_for_view = torch.cat( + [ + latents[:, :, h_start:h_end, w_start:w_end] + for h_start, h_end, w_start, w_end in batch_view + ] + ) + + # rematch block's scheduler status + self.scheduler.__dict__.update(views_scheduler_status[j]) + + # expand the latents if we are doing classifier free guidance + latent_model_input = ( + latents_for_view.repeat_interleave(2, dim=0) + if do_classifier_free_guidance + else latents_for_view + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # repeat prompt_embeds for batch + prompt_embeds_input = torch.cat([prompt_embeds] * vb_size) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds_input, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2] + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg( + noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents_denoised_batch = self.scheduler.step( + noise_pred, t, latents_for_view, **extra_step_kwargs + ).prev_sample + + # save views scheduler status after sample + views_scheduler_status[j] = copy.deepcopy(self.scheduler.__dict__) + + # extract value from batch + for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip( + latents_denoised_batch.chunk(vb_size), batch_view + ): + if circular_padding and w_end > latents.shape[3]: + # Case for circular padding + value[:, :, h_start:h_end, w_start:] += latents_view_denoised[ + :, :, h_start:h_end, : latents.shape[3] - w_start + ] + value[:, :, h_start:h_end, : w_end - latents.shape[3]] += latents_view_denoised[ + :, :, h_start:h_end, latents.shape[3] - w_start : + ] + count[:, :, h_start:h_end, w_start:] += 1 + count[:, :, h_start:h_end, : w_end - latents.shape[3]] += 1 + else: + value[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised + count[:, :, h_start:h_end, w_start:w_end] += 1 + + # take the MultiDiffusion step. Eq. 5 in MultiDiffusion paper: https://huggingface.co/papers/2302.08113 + latents = torch.where(count > 0, value / count, value) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type != "latent": + if circular_padding: + image = self.decode_latents_with_padding(latents) + else: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_safe/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_safe/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e911109b2e6e4ec3babaa919b2a1e9769a613ca5 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_safe/__init__.py @@ -0,0 +1,99 @@ +from dataclasses import dataclass +from enum import Enum +from typing import TYPE_CHECKING + +import numpy as np +import PIL +from PIL import Image + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + BaseOutput, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +@dataclass +class SafetyConfig(object): + WEAK = { + "sld_warmup_steps": 15, + "sld_guidance_scale": 20, + "sld_threshold": 0.0, + "sld_momentum_scale": 0.0, + "sld_mom_beta": 0.0, + } + MEDIUM = { + "sld_warmup_steps": 10, + "sld_guidance_scale": 1000, + "sld_threshold": 0.01, + "sld_momentum_scale": 0.3, + "sld_mom_beta": 0.4, + } + STRONG = { + "sld_warmup_steps": 7, + "sld_guidance_scale": 2000, + "sld_threshold": 0.025, + "sld_momentum_scale": 0.5, + "sld_mom_beta": 0.7, + } + MAX = { + "sld_warmup_steps": 0, + "sld_guidance_scale": 5000, + "sld_threshold": 1.0, + "sld_momentum_scale": 0.5, + "sld_mom_beta": 0.7, + } + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {} + +_additional_imports.update({"SafetyConfig": SafetyConfig}) + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure.update( + { + "pipeline_output": ["StableDiffusionSafePipelineOutput"], + "pipeline_stable_diffusion_safe": ["StableDiffusionPipelineSafe"], + "safety_checker": ["StableDiffusionSafetyChecker"], + } + ) + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_output import StableDiffusionSafePipelineOutput + from .pipeline_stable_diffusion_safe import StableDiffusionPipelineSafe + from .safety_checker import SafeStableDiffusionSafetyChecker + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_safe/pipeline_output.py b/diffusers/pipelines/deprecated/stable_diffusion_safe/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..21fc7fec07d102dd57988076678b3de9f4c3c15b --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_safe/pipeline_output.py @@ -0,0 +1,33 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ....utils import ( + BaseOutput, +) + + +@dataclass +class StableDiffusionSafePipelineOutput(BaseOutput): + """ + Output class for Safe Stable Diffusion pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + nsfw_content_detected (`list[bool]`) + list of flags denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, or `None` if safety checking could not be performed. + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images that were flagged by the safety checker any may contain "not-safe-for-work" + (nsfw) content, or `None` if no safety check was performed or no images were flagged. + applied_safety_concept (`str`) + The safety concept that was applied for safety guidance, or `None` if safety guidance was disabled + """ + + images: list[PIL.Image.Image] | np.ndarray + nsfw_content_detected: list[bool] | None + unsafe_images: list[PIL.Image.Image] | np.ndarray | None + applied_safety_concept: str | None diff --git a/diffusers/pipelines/deprecated/stable_diffusion_safe/pipeline_stable_diffusion_safe.py b/diffusers/pipelines/deprecated/stable_diffusion_safe/pipeline_stable_diffusion_safe.py new file mode 100644 index 0000000000000000000000000000000000000000..35c7f9b970b9665e3cebd031a5738239c0521969 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_safe/pipeline_stable_diffusion_safe.py @@ -0,0 +1,784 @@ +import inspect +import warnings +from typing import Callable + +import numpy as np +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ....configuration_utils import FrozenDict +from ....image_processor import PipelineImageInput +from ....loaders import IPAdapterMixin +from ....models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import deprecate, is_torch_xla_available, logging +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionSafePipelineOutput +from .safety_checker import SafeStableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class StableDiffusionPipelineSafe(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, IPAdapterMixin): + _last_supported_version = "0.33.1" + + r""" + Pipeline based on the [`StableDiffusionPipeline`] for text-to-image generation using Safe Latent Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: SafeStableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection | None = None, + requires_safety_checker: bool = True, + ): + super().__init__() + safety_concept: str | None = ( + "an image showing hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity," + " bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, child" + " abuse, brutality, cruelty" + ) + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self._safety_text_concept = safety_concept + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.register_to_config(requires_safety_checker=requires_safety_checker) + + @property + def safety_concept(self): + r""" + Getter method for the safety concept used with SLD + + Returns: + `str`: The text describing the safety concept + """ + return self._safety_text_concept + + @safety_concept.setter + def safety_concept(self, concept): + r""" + Setter method for the safety concept used with SLD + + Args: + concept (`str`): + The text of the new safety concept + """ + self._safety_text_concept = concept + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + enable_safety_guidance, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # Encode the safety concept text + if enable_safety_guidance: + safety_concept_input = self.tokenizer( + [self._safety_text_concept], + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + safety_embeddings = self.text_encoder(safety_concept_input.input_ids.to(self.device))[0] + + # duplicate safety embeddings for each generation per prompt, using mps friendly method + seq_len = safety_embeddings.shape[1] + safety_embeddings = safety_embeddings.repeat(batch_size, num_images_per_prompt, 1) + safety_embeddings = safety_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance + sld, we need to do three forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing three forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, safety_embeddings]) + + else: + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def run_safety_checker(self, image, device, dtype, enable_safety_guidance): + if self.safety_checker is not None: + images = image.copy() + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + flagged_images = np.zeros((2, *image.shape[1:])) + if any(has_nsfw_concept): + logger.warning( + "Potential NSFW content was detected in one or more images. A black image will be returned" + " instead." + f"{'You may look at this images in the `unsafe_images` variable of the output at your own discretion.' if enable_safety_guidance else 'Try again with a different prompt and/or seed.'}" + ) + for idx, has_nsfw_concept in enumerate(has_nsfw_concept): + if has_nsfw_concept: + flagged_images[idx] = images[idx] + image[idx] = np.zeros(image[idx].shape) # black image + else: + has_nsfw_concept = None + flagged_images = None + return image, has_nsfw_concept, flagged_images + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def perform_safety_guidance( + self, + enable_safety_guidance, + safety_momentum, + noise_guidance, + noise_pred_out, + i, + sld_guidance_scale, + sld_warmup_steps, + sld_threshold, + sld_momentum_scale, + sld_mom_beta, + ): + # Perform SLD guidance + if enable_safety_guidance: + if safety_momentum is None: + safety_momentum = torch.zeros_like(noise_guidance) + noise_pred_text, noise_pred_uncond = noise_pred_out[0], noise_pred_out[1] + noise_pred_safety_concept = noise_pred_out[2] + + # Equation 6 + scale = torch.clamp(torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0) + + # Equation 6 + safety_concept_scale = torch.where( + (noise_pred_text - noise_pred_safety_concept) >= sld_threshold, torch.zeros_like(scale), scale + ) + + # Equation 4 + noise_guidance_safety = torch.mul((noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale) + + # Equation 7 + noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum + + # Equation 8 + safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety + + if i >= sld_warmup_steps: # Warmup + # Equation 3 + noise_guidance = noise_guidance - noise_guidance_safety + return noise_guidance, safety_momentum + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + sld_guidance_scale: float | None = 1000, + sld_warmup_steps: int | None = 10, + sld_threshold: float | None = 0.01, + sld_momentum_scale: float | None = 0.3, + sld_mom_beta: float | None = 0.4, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + sld_guidance_scale (`float`, *optional*, defaults to 1000): + If `sld_guidance_scale < 1`, safety guidance is disabled. + sld_warmup_steps (`int`, *optional*, defaults to 10): + Number of warmup steps for safety guidance. SLD is only be applied for diffusion steps greater than + `sld_warmup_steps`. + sld_threshold (`float`, *optional*, defaults to 0.01): + Threshold that separates the hyperplane between appropriate and inappropriate images. + sld_momentum_scale (`float`, *optional*, defaults to 0.3): + Scale of the SLD momentum to be added to the safety guidance at each diffusion step. If set to 0.0, + momentum is disabled. Momentum is built up during warmup for diffusion steps smaller than + `sld_warmup_steps`. + sld_mom_beta (`float`, *optional*, defaults to 0.4): + Defines how safety guidance momentum builds up. `sld_mom_beta` indicates how much of the previous + momentum is kept. Momentum is built up during warmup for diffusion steps smaller than + `sld_warmup_steps`. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + + Examples: + + ```py + import torch + from diffusers import StableDiffusionPipelineSafe + from diffusers.pipelines.stable_diffusion_safe import SafetyConfig + + pipeline = StableDiffusionPipelineSafe.from_pretrained( + "AIML-TUDA/stable-diffusion-safe", torch_dtype=torch.float16 + ).to("cuda") + prompt = "the four horsewomen of the apocalypse, painting by tom of finland, gaston bussiere, craig mullins, j. c. leyendecker" + image = pipeline(prompt=prompt, **SafetyConfig.MEDIUM).images[0] + ``` + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + enable_safety_guidance = sld_guidance_scale > 1.0 and do_classifier_free_guidance + if not enable_safety_guidance: + warnings.warn("Safety checker disabled!") + + if ip_adapter_image is not None: + output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True + image_embeds, negative_image_embeds = self.encode_image( + ip_adapter_image, device, num_images_per_prompt, output_hidden_state + ) + if do_classifier_free_guidance: + if enable_safety_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds, image_embeds]) + else: + image_embeds = torch.cat([negative_image_embeds, image_embeds]) + + # 3. Encode input prompt + prompt_embeds = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt, enable_safety_guidance + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + safety_momentum = None + + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = ( + torch.cat([latents] * (3 if enable_safety_guidance else 2)) + if do_classifier_free_guidance + else latents + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, t, encoder_hidden_states=prompt_embeds, added_cond_kwargs=added_cond_kwargs + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_out = noise_pred.chunk((3 if enable_safety_guidance else 2)) + noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1] + + # default classifier free guidance + noise_guidance = noise_pred_text - noise_pred_uncond + + # Perform SLD guidance + if enable_safety_guidance: + if safety_momentum is None: + safety_momentum = torch.zeros_like(noise_guidance) + noise_pred_safety_concept = noise_pred_out[2] + + # Equation 6 + scale = torch.clamp( + torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0 + ) + + # Equation 6 + safety_concept_scale = torch.where( + (noise_pred_text - noise_pred_safety_concept) >= sld_threshold, + torch.zeros_like(scale), + scale, + ) + + # Equation 4 + noise_guidance_safety = torch.mul( + (noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale + ) + + # Equation 7 + noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum + + # Equation 8 + safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety + + if i >= sld_warmup_steps: # Warmup + # Equation 3 + noise_guidance = noise_guidance - noise_guidance_safety + + noise_pred = noise_pred_uncond + guidance_scale * noise_guidance + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept, flagged_images = self.run_safety_checker( + image, device, prompt_embeds.dtype, enable_safety_guidance + ) + + # 10. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + if flagged_images is not None: + flagged_images = self.numpy_to_pil(flagged_images) + + if not return_dict: + return ( + image, + has_nsfw_concept, + self._safety_text_concept if enable_safety_guidance else None, + flagged_images, + ) + + return StableDiffusionSafePipelineOutput( + images=image, + nsfw_content_detected=has_nsfw_concept, + applied_safety_concept=self._safety_text_concept if enable_safety_guidance else None, + unsafe_images=flagged_images, + ) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_safe/safety_checker.py b/diffusers/pipelines/deprecated/stable_diffusion_safe/safety_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..c07efc3cc68e4092e72453131d3b5ac24b96a2c0 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_safe/safety_checker.py @@ -0,0 +1,109 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel + +from ....utils import logging + + +logger = logging.get_logger(__name__) + + +def cosine_distance(image_embeds, text_embeds): + normalized_image_embeds = nn.functional.normalize(image_embeds) + normalized_text_embeds = nn.functional.normalize(text_embeds) + return torch.mm(normalized_image_embeds, normalized_text_embeds.t()) + + +class SafeStableDiffusionSafetyChecker(PreTrainedModel): + config_class = CLIPConfig + + _no_split_modules = ["CLIPEncoderLayer"] + + def __init__(self, config: CLIPConfig): + super().__init__(config) + + self.vision_model = CLIPVisionModel(config.vision_config) + self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False) + + self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False) + self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False) + + self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False) + self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False) + + @torch.no_grad() + def forward(self, clip_input, images): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy() + cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy() + + result = [] + batch_size = image_embeds.shape[0] + for i in range(batch_size): + result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []} + + # increase this value to create a stronger `nfsw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + for concept_idx in range(len(special_cos_dist[0])): + concept_cos = special_cos_dist[i][concept_idx] + concept_threshold = self.special_care_embeds_weights[concept_idx].item() + result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["special_scores"][concept_idx] > 0: + result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]}) + adjustment = 0.01 + + for concept_idx in range(len(cos_dist[0])): + concept_cos = cos_dist[i][concept_idx] + concept_threshold = self.concept_embeds_weights[concept_idx].item() + result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["concept_scores"][concept_idx] > 0: + result_img["bad_concepts"].append(concept_idx) + + result.append(result_img) + + has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result] + + return images, has_nsfw_concepts + + @torch.no_grad() + def forward_onnx(self, clip_input: torch.Tensor, images: torch.Tensor): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds) + cos_dist = cosine_distance(image_embeds, self.concept_embeds) + + # increase this value to create a stronger `nsfw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment + # special_scores = special_scores.round(decimals=3) + special_care = torch.any(special_scores > 0, dim=1) + special_adjustment = special_care * 0.01 + special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1]) + + concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment + # concept_scores = concept_scores.round(decimals=3) + has_nsfw_concepts = torch.any(concept_scores > 0, dim=1) + + return images, has_nsfw_concepts diff --git a/diffusers/pipelines/deprecated/stable_diffusion_sag/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_sag/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8cdd1ec6bdf0cf4a8ff9e8e1eea801fef97ff672 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_sag/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_sag/pipeline_stable_diffusion_sag.py b/diffusers/pipelines/deprecated/stable_diffusion_sag/pipeline_stable_diffusion_sag.py new file mode 100644 index 0000000000000000000000000000000000000000..678ef74f387cbc4d05caf5af42c921814723fca2 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_sag/pipeline_stable_diffusion_sag.py @@ -0,0 +1,969 @@ +# Copyright 2025 Susung Hong and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +import torch.nn.functional as F +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ....image_processor import PipelineImageInput, VaeImageProcessor +from ....loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionSAGPipeline + + >>> pipe = StableDiffusionSAGPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, sag_scale=0.75).images[0] + ``` +""" + + +# processes and stores attention probabilities +class CrossAttnStoreProcessor: + def __init__(self): + self.attention_probs = None + + def __call__( + self, + attn, + hidden_states, + encoder_hidden_states=None, + attention_mask=None, + ): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + self.attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(self.attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +# Modified to get self-attention guidance scale in this paper (https://huggingface.co/papers/2210.00939) as an input +class StableDiffusionSAGPipeline( + DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, IPAdapterMixin +): + _last_supported_version = "0.33.1" + + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection | None = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + image_embeds = [] + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + single_negative_image_embeds = torch.stack( + [single_negative_image_embeds] * num_images_per_prompt, dim=0 + ) + + if do_classifier_free_guidance: + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds]) + single_image_embeds = single_image_embeds.to(device) + + image_embeds.append(single_image_embeds) + else: + image_embeds = ip_adapter_image_embeds + return image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + sag_scale: float = 0.75, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int | None = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + sag_scale (`float`, *optional*, defaults to 0.75): + Chosen between [0, 1.0] for better quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. If not provided, embeddings are computed from the + `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + # and `sag_scale` is` `s` of equation (16) + # of the self-attention guidance paper: https://huggingface.co/papers/2210.00939 + # `sag_scale = 0` means no self-attention guidance + do_self_attention_guidance = sag_scale > 0.0 + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + do_classifier_free_guidance, + ) + + if do_classifier_free_guidance: + image_embeds = [] + negative_image_embeds = [] + for tmp_image_embeds in ip_adapter_image_embeds: + single_negative_image_embeds, single_image_embeds = tmp_image_embeds.chunk(2) + image_embeds.append(single_image_embeds) + negative_image_embeds.append(single_negative_image_embeds) + else: + image_embeds = ip_adapter_image_embeds + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + if timesteps.dtype not in [torch.int16, torch.int32, torch.int64]: + raise ValueError( + f"{self.__class__.__name__} does not support using a scheduler of type {self.scheduler.__class__.__name__}. Please make sure to use one of 'DDIMScheduler, PNDMScheduler, DDPMScheduler, DEISMultistepScheduler, UniPCMultistepScheduler, DPMSolverMultistepScheduler, DPMSolverSinglestepScheduler'." + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + if do_classifier_free_guidance: + added_uncond_kwargs = ( + {"image_embeds": negative_image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 7. Denoising loop + original_attn_proc = self.unet.attn_processors + store_processor = CrossAttnStoreProcessor() + self.unet.mid_block.attentions[0].transformer_blocks[0].attn1.processor = store_processor + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + map_size = None + + def get_map_size(module, input, output): + nonlocal map_size + map_size = output[0].shape[-2:] + + with self.unet.mid_block.attentions[0].register_forward_hook(get_map_size): + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # perform self-attention guidance with the stored self-attention map + if do_self_attention_guidance: + # classifier-free guidance produces two chunks of attention map + # and we only use unconditional one according to equation (25) + # in https://huggingface.co/papers/2210.00939 + if do_classifier_free_guidance: + # DDIM-like prediction of x0 + pred_x0 = self.pred_x0(latents, noise_pred_uncond, t) + # get the stored attention maps + uncond_attn, cond_attn = store_processor.attention_probs.chunk(2) + # self-attention-based degrading of latents + degraded_latents = self.sag_masking( + pred_x0, uncond_attn, map_size, t, self.pred_epsilon(latents, noise_pred_uncond, t) + ) + uncond_emb, _ = prompt_embeds.chunk(2) + # forward and give guidance + degraded_pred = self.unet( + degraded_latents, + t, + encoder_hidden_states=uncond_emb, + added_cond_kwargs=added_uncond_kwargs, + ).sample + noise_pred += sag_scale * (noise_pred_uncond - degraded_pred) + else: + # DDIM-like prediction of x0 + pred_x0 = self.pred_x0(latents, noise_pred, t) + # get the stored attention maps + cond_attn = store_processor.attention_probs + # self-attention-based degrading of latents + degraded_latents = self.sag_masking( + pred_x0, cond_attn, map_size, t, self.pred_epsilon(latents, noise_pred, t) + ) + # forward and give guidance + degraded_pred = self.unet( + degraded_latents, + t, + encoder_hidden_states=prompt_embeds, + added_cond_kwargs=added_cond_kwargs, + ).sample + noise_pred += sag_scale * (noise_pred - degraded_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + def sag_masking(self, original_latents, attn_map, map_size, t, eps): + # Same masking process as in SAG paper: https://huggingface.co/papers/2210.00939 + bh, hw1, hw2 = attn_map.shape + b, latent_channel, latent_h, latent_w = original_latents.shape + h = self.unet.config.attention_head_dim + if isinstance(h, list): + h = h[-1] + + # Produce attention mask + attn_map = attn_map.reshape(b, h, hw1, hw2) + attn_mask = attn_map.mean(1, keepdim=False).sum(1, keepdim=False) > 1.0 + attn_mask = ( + attn_mask.reshape(b, map_size[0], map_size[1]) + .unsqueeze(1) + .repeat(1, latent_channel, 1, 1) + .type(attn_map.dtype) + ) + attn_mask = F.interpolate(attn_mask, (latent_h, latent_w)) + + # Blur according to the self-attention mask + degraded_latents = gaussian_blur_2d(original_latents, kernel_size=9, sigma=1.0) + degraded_latents = degraded_latents * attn_mask + original_latents * (1 - attn_mask) + + # Noise it again to match the noise level + degraded_latents = self.scheduler.add_noise(degraded_latents, noise=eps, timesteps=t[None]) + + return degraded_latents + + # Modified from diffusers.schedulers.scheduling_ddim.DDIMScheduler.step + # Note: there are some schedulers that clip or do not return x_0 (PNDMScheduler, DDIMScheduler, etc.) + def pred_x0(self, sample, model_output, timestep): + alpha_prod_t = self.scheduler.alphas_cumprod[timestep].to(sample.device) + + beta_prod_t = 1 - alpha_prod_t + if self.scheduler.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.scheduler.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.scheduler.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + # predict V + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`," + " or `v_prediction`" + ) + + return pred_original_sample + + def pred_epsilon(self, sample, model_output, timestep): + alpha_prod_t = self.scheduler.alphas_cumprod[timestep] + + beta_prod_t = 1 - alpha_prod_t + if self.scheduler.config.prediction_type == "epsilon": + pred_eps = model_output + elif self.scheduler.config.prediction_type == "sample": + pred_eps = (sample - (alpha_prod_t**0.5) * model_output) / (beta_prod_t**0.5) + elif self.scheduler.config.prediction_type == "v_prediction": + pred_eps = (beta_prod_t**0.5) * sample + (alpha_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`," + " or `v_prediction`" + ) + + return pred_eps + + +# Gaussian blur +def gaussian_blur_2d(img, kernel_size, sigma): + ksize_half = (kernel_size - 1) * 0.5 + + x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size) + + pdf = torch.exp(-0.5 * (x / sigma).pow(2)) + + x_kernel = pdf / pdf.sum() + x_kernel = x_kernel.to(device=img.device, dtype=img.dtype) + + kernel2d = torch.mm(x_kernel[:, None], x_kernel[None, :]) + kernel2d = kernel2d.expand(img.shape[-3], 1, kernel2d.shape[0], kernel2d.shape[1]) + + padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2] + + img = F.pad(img, padding, mode="reflect") + img = F.conv2d(img, kernel2d, groups=img.shape[-3]) + + return img diff --git a/diffusers/pipelines/deprecated/stable_diffusion_variants/__init__.py b/diffusers/pipelines/deprecated/stable_diffusion_variants/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..36cf1a33ce6ada8e718aabadb9a706737aee30bd --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_variants/__init__.py @@ -0,0 +1,55 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_cycle_diffusion"] = ["CycleDiffusionPipeline"] + _import_structure["pipeline_stable_diffusion_inpaint_legacy"] = ["StableDiffusionInpaintPipelineLegacy"] + _import_structure["pipeline_stable_diffusion_model_editing"] = ["StableDiffusionModelEditingPipeline"] + + _import_structure["pipeline_stable_diffusion_paradigms"] = ["StableDiffusionParadigmsPipeline"] + _import_structure["pipeline_stable_diffusion_pix2pix_zero"] = ["StableDiffusionPix2PixZeroPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * + + else: + from .pipeline_cycle_diffusion import CycleDiffusionPipeline + from .pipeline_stable_diffusion_inpaint_legacy import StableDiffusionInpaintPipelineLegacy + from .pipeline_stable_diffusion_model_editing import StableDiffusionModelEditingPipeline + from .pipeline_stable_diffusion_paradigms import StableDiffusionParadigmsPipeline + from .pipeline_stable_diffusion_pix2pix_zero import StableDiffusionPix2PixZeroPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..c5a1127c2908787eb38350874f5a63953553bbd0 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py @@ -0,0 +1,953 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....configuration_utils import FrozenDict +from ....image_processor import PipelineImageInput, VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import DDIMScheduler +from ....utils import PIL_INTERPOLATION, USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline +from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def posterior_sample(scheduler, latents, timestep, clean_latents, generator, eta): + # 1. get previous step value (=t-1) + prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps + + if prev_timestep <= 0: + return clean_latents + + # 2. compute alphas, betas + alpha_prod_t = scheduler.alphas_cumprod[timestep] + alpha_prod_t_prev = ( + scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod + ) + + variance = scheduler._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + # direction pointing to x_t + e_t = (latents - alpha_prod_t ** (0.5) * clean_latents) / (1 - alpha_prod_t) ** (0.5) + dir_xt = (1.0 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * e_t + noise = std_dev_t * randn_tensor( + clean_latents.shape, dtype=clean_latents.dtype, device=clean_latents.device, generator=generator + ) + prev_latents = alpha_prod_t_prev ** (0.5) * clean_latents + dir_xt + noise + + return prev_latents + + +def compute_noise(scheduler, prev_latents, latents, timestep, noise_pred, eta): + # 1. get previous step value (=t-1) + prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = scheduler.alphas_cumprod[timestep] + alpha_prod_t_prev = ( + scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod + ) + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5) + + # 4. Clip "predicted x_0" + if scheduler.config.clip_sample: + pred_original_sample = torch.clamp(pred_original_sample, -1, 1) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = scheduler._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * noise_pred + + noise = (prev_latents - (alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction)) / ( + variance ** (0.5) * eta + ) + return noise + + +class CycleDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin): + r""" + Pipeline for text-guided image to image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can only be an + instance of [`DDIMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + image = image.to(device=device, dtype=dtype) + + batch_size = image.shape[0] + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if isinstance(generator, list): + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt * num_images_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0) + + # add noise to latents using the timestep + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + clean_latents = init_latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents, clean_latents + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + source_prompt: str | list[str], + image: PipelineImageInput = None, + strength: float = 0.8, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 7.5, + source_guidance_scale: float | None = 1, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.1, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + source_prompt (`str` or `list[str]`): + The prompt or prompts describing the input `image`. Used together with `prompt` to guide the + cycle-diffusion editing process. + image (`torch.Tensor` `np.ndarray`, `PIL.Image.Image`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image` or tensor representing an image batch to be used as the starting point. Can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + source_guidance_scale (`float`, *optional*, defaults to 1): + Guidance scale for the source prompt. This is useful to control the amount of influence the source + prompt has for encoding. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Example: + + ```py + import requests + import torch + from PIL import Image + from io import BytesIO + + from diffusers import CycleDiffusionPipeline, DDIMScheduler + + # load the pipeline + # make sure you're logged in with `hf auth login` + model_id_or_path = "CompVis/stable-diffusion-v1-4" + scheduler = DDIMScheduler.from_pretrained(model_id_or_path, subfolder="scheduler") + pipe = CycleDiffusionPipeline.from_pretrained(model_id_or_path, scheduler=scheduler).to("cuda") + + # let's download an initial image + url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/An%20astronaut%20riding%20a%20horse.png" + response = requests.get(url) + init_image = Image.open(BytesIO(response.content)).convert("RGB") + init_image = init_image.resize((512, 512)) + init_image.save("horse.png") + + # let's specify a prompt + source_prompt = "An astronaut riding a horse" + prompt = "An astronaut riding an elephant" + + # call the pipeline + image = pipe( + prompt=prompt, + source_prompt=source_prompt, + image=init_image, + num_inference_steps=100, + eta=0.1, + strength=0.8, + guidance_scale=2, + source_guidance_scale=1, + ).images[0] + + image.save("horse_to_elephant.png") + + # let's try another example + # See more samples at the original repo: https://github.com/ChenWu98/cycle-diffusion + url = ( + "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/A%20black%20colored%20car.png" + ) + response = requests.get(url) + init_image = Image.open(BytesIO(response.content)).convert("RGB") + init_image = init_image.resize((512, 512)) + init_image.save("black.png") + + source_prompt = "A black colored car" + prompt = "A blue colored car" + + # call the pipeline + torch.manual_seed(0) + image = pipe( + prompt=prompt, + source_prompt=source_prompt, + image=init_image, + num_inference_steps=100, + eta=0.1, + strength=0.85, + guidance_scale=3, + source_guidance_scale=1, + ).images[0] + + image.save("black_to_blue.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 1. Check inputs + self.check_inputs(prompt, strength, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds_tuple = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + source_prompt_embeds_tuple = self.encode_prompt( + source_prompt, device, num_images_per_prompt, do_classifier_free_guidance, None, clip_skip=clip_skip + ) + if prompt_embeds_tuple[1] is not None: + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + else: + prompt_embeds = prompt_embeds_tuple[0] + if source_prompt_embeds_tuple[1] is not None: + source_prompt_embeds = torch.cat([source_prompt_embeds_tuple[1], source_prompt_embeds_tuple[0]]) + else: + source_prompt_embeds = source_prompt_embeds_tuple[0] + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents, clean_latents = self.prepare_latents( + image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator + ) + source_latents = latents + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + generator = extra_step_kwargs.pop("generator", None) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + source_latent_model_input = ( + torch.cat([source_latents] * 2) if do_classifier_free_guidance else source_latents + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + source_latent_model_input = self.scheduler.scale_model_input(source_latent_model_input, t) + + # predict the noise residual + if do_classifier_free_guidance: + concat_latent_model_input = torch.stack( + [ + source_latent_model_input[0], + latent_model_input[0], + source_latent_model_input[1], + latent_model_input[1], + ], + dim=0, + ) + concat_prompt_embeds = torch.stack( + [ + source_prompt_embeds[0], + prompt_embeds[0], + source_prompt_embeds[1], + prompt_embeds[1], + ], + dim=0, + ) + else: + concat_latent_model_input = torch.cat( + [ + source_latent_model_input, + latent_model_input, + ], + dim=0, + ) + concat_prompt_embeds = torch.cat( + [ + source_prompt_embeds, + prompt_embeds, + ], + dim=0, + ) + + concat_noise_pred = self.unet( + concat_latent_model_input, + t, + cross_attention_kwargs=cross_attention_kwargs, + encoder_hidden_states=concat_prompt_embeds, + ).sample + + # perform guidance + if do_classifier_free_guidance: + ( + source_noise_pred_uncond, + noise_pred_uncond, + source_noise_pred_text, + noise_pred_text, + ) = concat_noise_pred.chunk(4, dim=0) + + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + source_noise_pred = source_noise_pred_uncond + source_guidance_scale * ( + source_noise_pred_text - source_noise_pred_uncond + ) + + else: + (source_noise_pred, noise_pred) = concat_noise_pred.chunk(2, dim=0) + + # Sample source_latents from the posterior distribution. + prev_source_latents = posterior_sample( + self.scheduler, source_latents, t, clean_latents, generator=generator, **extra_step_kwargs + ) + # Compute noise. + noise = compute_noise( + self.scheduler, prev_source_latents, source_latents, t, source_noise_pred, **extra_step_kwargs + ) + source_latents = prev_source_latents + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, t, latents, variance_noise=noise, **extra_step_kwargs + ).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + # 9. Post-processing + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_onnx_stable_diffusion_inpaint_legacy.py b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_onnx_stable_diffusion_inpaint_legacy.py new file mode 100644 index 0000000000000000000000000000000000000000..f526dc419cea63d5e60fadae61a10c07f266dc1a --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_onnx_stable_diffusion_inpaint_legacy.py @@ -0,0 +1,543 @@ +import inspect +from typing import Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ....configuration_utils import FrozenDict +from ....schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ....utils import deprecate, logging +from ...onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ...pipeline_utils import DiffusionPipeline +from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def preprocess(image): + w, h = image.size + w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL.Image.LANCZOS) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask, scale_factor=8): + mask = mask.convert("L") + w, h = mask.size + w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32 + mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL.Image.NEAREST) + mask = np.array(mask).astype(np.float32) / 255.0 + mask = np.tile(mask, (4, 1, 1)) + mask = mask[None].transpose(0, 1, 2, 3) # what does this step do? + mask = 1 - mask # repaint white, keep black + return mask + + +class OnnxStableDiffusionInpaintPipelineLegacy(DiffusionPipeline): + r""" + Pipeline for text-guided image inpainting using Stable Diffusion. This is a *legacy feature* for Onnx pipelines to + provide compatibility with StableDiffusionInpaintPipelineLegacy and may be removed in the future. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler, + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int | None, + do_classifier_free_guidance: bool, + negative_prompt: str | None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def check_inputs( + self, + prompt, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def __call__( + self, + prompt: str | list[str], + image: np.ndarray | PIL.Image.Image = None, + mask_image: np.ndarray | PIL.Image.Image = None, + strength: float = 0.8, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: np.random.RandomState | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, np.ndarray], None] | None = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`nd.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. This is the image whose masked region will be inpainted. + mask_image (`nd.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.uu + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (?) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # check inputs. Raise error if not correct + self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + # define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if generator is None: + generator = np.random + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + if isinstance(image, PIL.Image.Image): + image = preprocess(image) + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + latents_dtype = prompt_embeds.dtype + image = image.astype(latents_dtype) + + # encode the init image into latents and scale the latents + init_latents = self.vae_encoder(sample=image)[0] + init_latents = 0.18215 * init_latents + + # Expand init_latents for batch_size and num_images_per_prompt + init_latents = np.concatenate([init_latents] * num_images_per_prompt, axis=0) + init_latents_orig = init_latents + + # preprocess mask + if not isinstance(mask_image, np.ndarray): + mask_image = preprocess_mask(mask_image, 8) + mask_image = mask_image.astype(latents_dtype) + mask = np.concatenate([mask_image] * num_images_per_prompt, axis=0) + + # check sizes + if not mask.shape == init_latents.shape: + raise ValueError("The mask and image should be the same size!") + + # get the original timestep using init_timestep + offset = self.scheduler.config.get("steps_offset", 0) + init_timestep = int(num_inference_steps * strength) + offset + init_timestep = min(init_timestep, num_inference_steps) + + timesteps = self.scheduler.timesteps.numpy()[-init_timestep] + timesteps = np.array([timesteps] * batch_size * num_images_per_prompt) + + # add noise to latents using the timesteps + noise = generator.randn(*init_latents.shape).astype(latents_dtype) + init_latents = self.scheduler.add_noise( + torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(timesteps) + ) + init_latents = init_latents.numpy() + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (?) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to ? in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + latents = init_latents + + t_start = max(num_inference_steps - init_timestep + offset, 0) + timesteps = self.scheduler.timesteps[t_start:].numpy() + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + timestep = np.array([t], dtype=timestep_dtype) + noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[ + 0 + ] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ).prev_sample + + latents = latents.numpy() + + init_latents_proper = self.scheduler.add_noise( + torch.from_numpy(init_latents_orig), torch.from_numpy(noise), torch.from_numpy(np.array([t])) + ) + + init_latents_proper = init_latents_proper.numpy() + + latents = (init_latents_proper * mask) + (latents * (1 - mask)) + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + # There will throw an error if use safety_checker batchsize>1 + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py new file mode 100644 index 0000000000000000000000000000000000000000..00d2a77269fdfa1d653077af2eabb86f893a42c0 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py @@ -0,0 +1,789 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....configuration_utils import FrozenDict +from ....image_processor import VaeImageProcessor +from ....loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import PIL_INTERPOLATION, USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline +from ...stable_diffusion import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) + + +def preprocess_image(image, batch_size): + w, h = image.size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]) + image = np.array(image).astype(np.float32) / 255.0 + image = np.vstack([image[None].transpose(0, 3, 1, 2)] * batch_size) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask, batch_size, scale_factor=8): + if not isinstance(mask, torch.Tensor): + mask = mask.convert("L") + w, h = mask.size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"]) + mask = np.array(mask).astype(np.float32) / 255.0 + mask = np.tile(mask, (4, 1, 1)) + mask = np.vstack([mask[None]] * batch_size) + mask = 1 - mask # repaint white, keep black + mask = torch.from_numpy(mask) + return mask + + else: + valid_mask_channel_sizes = [1, 3] + # if mask channel is fourth tensor dimension, permute dimensions to pytorch standard (B, C, H, W) + if mask.shape[3] in valid_mask_channel_sizes: + mask = mask.permute(0, 3, 1, 2) + elif mask.shape[1] not in valid_mask_channel_sizes: + raise ValueError( + f"Mask channel dimension of size in {valid_mask_channel_sizes} should be second or fourth dimension," + f" but received mask of shape {tuple(mask.shape)}" + ) + # (potentially) reduce mask channel dimension from 3 to 1 for broadcasting to latent shape + mask = mask.mean(dim=1, keepdim=True) + h, w = mask.shape[-2:] + h, w = (x - x % 8 for x in (h, w)) # resize to integer multiple of 8 + mask = torch.nn.functional.interpolate(mask, (h // scale_factor, w // scale_factor)) + return mask + + +class StableDiffusionInpaintPipelineLegacy( + DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin, FromSingleFileMixin +): + r""" + Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + In addition the pipeline inherits the following loading methods: + - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`] + - *LoRA*: [`loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] + - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`] + + as well as the following saving methods: + - *LoRA*: [`loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + deprecation_message = ( + f"The class {self.__class__} is deprecated and will be removed in v1.0.0. You can achieve exactly the same functionality" + "by loading your model into `StableDiffusionInpaintPipeline` instead. See https://github.com/huggingface/diffusers/pull/3533" + "for more information." + ) + deprecate("legacy is outdated", "1.0.0", deprecation_message, standard_warn=False) + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, image, timestep, num_images_per_prompt, dtype, device, generator): + image = image.to(device=device, dtype=dtype) + init_latent_dist = self.vae.encode(image).latent_dist + init_latents = init_latent_dist.sample(generator=generator) + init_latents = self.vae.config.scaling_factor * init_latents + + # Expand init_latents for batch_size and num_images_per_prompt + init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0) + init_latents_orig = init_latents + + # add noise to latents using the timesteps + noise = randn_tensor(init_latents.shape, generator=generator, device=device, dtype=dtype) + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + return latents, init_latents_orig, noise + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + image: torch.Tensor | PIL.Image.Image = None, + mask_image: torch.Tensor | PIL.Image.Image = None, + strength: float = 0.8, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + add_predicted_noise: bool | None = False, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. This is the image whose masked region will be inpainted. + mask_image (`torch.Tensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If mask is a tensor, the + expected shape should be either `(B, H, W, C)` or `(B, C, H, W)`, where C is 1 or 3. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength` + is 1, the denoising process will be run on the masked area for the full number of iterations specified + in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more noise to + that region the larger the `strength`. If `strength` is 0, no inpainting will occur. + num_inference_steps (`int`, *optional*, defaults to 50): + The reference number of denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. This parameter will be modulated by `strength`, as explained above. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` + is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + add_predicted_noise (`bool`, *optional*, defaults to True): + Use predicted noise instead of random noise when constructing noisy versions of the original image in + the reverse diffusion process + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + # 1. Check inputs + self.check_inputs(prompt, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Preprocess image and mask + if not isinstance(image, torch.Tensor): + image = preprocess_image(image, batch_size) + + mask_image = preprocess_mask(mask_image, batch_size, self.vae_scale_factor) + + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + # encode the init image into latents and scale the latents + latents, init_latents_orig, noise = self.prepare_latents( + image, latent_timestep, num_images_per_prompt, prompt_embeds.dtype, device, generator + ) + + # 7. Prepare mask latent + mask = mask_image.to(device=device, dtype=latents.dtype) + mask = torch.cat([mask] * num_images_per_prompt) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + # masking + if add_predicted_noise: + init_latents_proper = self.scheduler.add_noise( + init_latents_orig, noise_pred_uncond, torch.tensor([t]) + ) + else: + init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t])) + + latents = (init_latents_proper * mask) + (latents * (1 - mask)) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + # use original latents corresponding to unmasked portions of the image + latents = (init_latents_orig * mask) + (latents * (1 - mask)) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py new file mode 100644 index 0000000000000000000000000000000000000000..851820c00aedf6089aa11c1c59d4d663fa41435b --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py @@ -0,0 +1,832 @@ +# Copyright 2025 TIME Authors and The HuggingFace Team. All rights reserved." +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....image_processor import VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import PNDMScheduler +from ....schedulers.scheduling_utils import SchedulerMixin +from ....utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +AUGS_CONST = ["A photo of ", "An image of ", "A picture of "] + + +class StableDiffusionModelEditingPipeline( + DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin +): + r""" + Pipeline for text-to-image model editing. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + with_to_k ([`bool`]): + Whether to edit the key projection matrices along with the value projection matrices. + with_augs ([`list`]): + Textual augmentations to apply while editing the text-to-image model. Set to `[]` for no augmentations. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: SchedulerMixin, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + with_to_k: bool = True, + with_augs: list = AUGS_CONST, + ): + super().__init__() + + if isinstance(scheduler, PNDMScheduler): + logger.error("PNDMScheduler for this pipeline is currently not supported.") + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + self.with_to_k = with_to_k + self.with_augs = with_augs + + # get cross-attention layers + ca_layers = [] + + def append_ca(net_): + if net_.__class__.__name__ == "CrossAttention": + ca_layers.append(net_) + elif hasattr(net_, "children"): + for net__ in net_.children(): + append_ca(net__) + + # recursively find all cross-attention layers in unet + for net in self.unet.named_children(): + if "down" in net[0]: + append_ca(net[1]) + elif "up" in net[0]: + append_ca(net[1]) + elif "mid" in net[0]: + append_ca(net[1]) + + # get projection matrices + self.ca_clip_layers = [l for l in ca_layers if l.to_v.in_features == 768] + self.projection_matrices = [l.to_v for l in self.ca_clip_layers] + self.og_matrices = [copy.deepcopy(l.to_v) for l in self.ca_clip_layers] + if self.with_to_k: + self.projection_matrices = self.projection_matrices + [l.to_k for l in self.ca_clip_layers] + self.og_matrices = self.og_matrices + [copy.deepcopy(l.to_k) for l in self.ca_clip_layers] + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def edit_model( + self, + source_prompt: str, + destination_prompt: str, + lamb: float = 0.1, + restart_params: bool = True, + ): + r""" + Apply model editing via closed-form solution (see Eq. 5 in the TIME + [paper](https://huggingface.co/papers/2303.08084)). + + Args: + source_prompt (`str`): + The source prompt containing the concept to be edited. + destination_prompt (`str`): + The destination prompt. Must contain all words from `source_prompt` with additional ones to specify the + target edit. + lamb (`float`, *optional*, defaults to 0.1): + The lambda parameter specifying the regularization intensity. Smaller values increase the editing + power. + restart_params (`bool`, *optional*, defaults to True): + Restart the model parameters to their pre-trained version before editing. This is done to avoid edit + compounding. When it is `False`, edits accumulate. + """ + + # restart LDM parameters + if restart_params: + num_ca_clip_layers = len(self.ca_clip_layers) + for idx_, l in enumerate(self.ca_clip_layers): + l.to_v = copy.deepcopy(self.og_matrices[idx_]) + self.projection_matrices[idx_] = l.to_v + if self.with_to_k: + l.to_k = copy.deepcopy(self.og_matrices[num_ca_clip_layers + idx_]) + self.projection_matrices[num_ca_clip_layers + idx_] = l.to_k + + # set up sentences + old_texts = [source_prompt] + new_texts = [destination_prompt] + # add augmentations + base = old_texts[0] if old_texts[0][0:1] != "A" else "a" + old_texts[0][1:] + for aug in self.with_augs: + old_texts.append(aug + base) + base = new_texts[0] if new_texts[0][0:1] != "A" else "a" + new_texts[0][1:] + for aug in self.with_augs: + new_texts.append(aug + base) + + # prepare input k* and v* + old_embs, new_embs = [], [] + for old_text, new_text in zip(old_texts, new_texts): + text_input = self.tokenizer( + [old_text, new_text], + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0] + old_emb, new_emb = text_embeddings + old_embs.append(old_emb) + new_embs.append(new_emb) + + # identify corresponding destinations for each token in old_emb + idxs_replaces = [] + for old_text, new_text in zip(old_texts, new_texts): + tokens_a = self.tokenizer(old_text).input_ids + tokens_b = self.tokenizer(new_text).input_ids + tokens_a = [self.tokenizer.encode("a ")[1] if self.tokenizer.decode(t) == "an" else t for t in tokens_a] + tokens_b = [self.tokenizer.encode("a ")[1] if self.tokenizer.decode(t) == "an" else t for t in tokens_b] + num_orig_tokens = len(tokens_a) + idxs_replace = [] + j = 0 + for i in range(num_orig_tokens): + curr_token = tokens_a[i] + while tokens_b[j] != curr_token: + j += 1 + idxs_replace.append(j) + j += 1 + while j < 77: + idxs_replace.append(j) + j += 1 + while len(idxs_replace) < 77: + idxs_replace.append(76) + idxs_replaces.append(idxs_replace) + + # prepare batch: for each pair of sentences, old context and new values + contexts, valuess = [], [] + for old_emb, new_emb, idxs_replace in zip(old_embs, new_embs, idxs_replaces): + context = old_emb.detach() + values = [] + with torch.no_grad(): + for layer in self.projection_matrices: + values.append(layer(new_emb[idxs_replace]).detach()) + contexts.append(context) + valuess.append(values) + + # edit the model + for layer_num in range(len(self.projection_matrices)): + # mat1 = \lambda W + \sum{v k^T} + mat1 = lamb * self.projection_matrices[layer_num].weight + + # mat2 = \lambda I + \sum{k k^T} + mat2 = lamb * torch.eye( + self.projection_matrices[layer_num].weight.shape[1], + device=self.projection_matrices[layer_num].weight.device, + ) + + # aggregate sums for mat1, mat2 + for context, values in zip(contexts, valuess): + context_vector = context.reshape(context.shape[0], context.shape[1], 1) + context_vector_T = context.reshape(context.shape[0], 1, context.shape[1]) + value_vector = values[layer_num].reshape(values[layer_num].shape[0], values[layer_num].shape[1], 1) + for_mat1 = (value_vector @ context_vector_T).sum(dim=0) + for_mat2 = (context_vector @ context_vector_T).sum(dim=0) + mat1 += for_mat1 + mat2 += for_mat2 + + # update projection matrix + self.projection_matrices[layer_num].weight = torch.nn.Parameter(mat1 @ torch.inverse(mat2)) + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + ```py + >>> import torch + >>> from diffusers import StableDiffusionModelEditingPipeline + + >>> model_ckpt = "CompVis/stable-diffusion-v1-4" + >>> pipe = StableDiffusionModelEditingPipeline.from_pretrained(model_ckpt) + + >>> pipe = pipe.to("cuda") + + >>> source_prompt = "A pack of roses" + >>> destination_prompt = "A pack of blue roses" + >>> pipe.edit_model(source_prompt, destination_prompt) + + >>> prompt = "A field of roses" + >>> image = pipe(prompt).images[0] + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_paradigms.py b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_paradigms.py new file mode 100644 index 0000000000000000000000000000000000000000..ea81be87a0f4b246a0b328e050da5a07357141e9 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_paradigms.py @@ -0,0 +1,798 @@ +# Copyright 2025 ParaDiGMS authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....image_processor import VaeImageProcessor +from ....loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import DDPMParallelScheduler + >>> from diffusers import StableDiffusionParadigmsPipeline + + >>> scheduler = DDPMParallelScheduler.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", subfolder="scheduler" + ... ) + + >>> pipe = StableDiffusionParadigmsPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", scheduler=scheduler, torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> ngpu, batch_per_device = torch.cuda.device_count(), 5 + >>> pipe.wrapped_unet = torch.nn.DataParallel(pipe.unet, device_ids=[d for d in range(ngpu)]) + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, parallel=ngpu * batch_per_device, num_inference_steps=1000).images[0] + ``` +""" + + +class StableDiffusionParadigmsPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-to-image generation using a parallelized version of Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # attribute to wrap the unet with torch.nn.DataParallel when running multiple denoising steps on multiple GPUs + self.wrapped_unet = self.unet + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _cumsum(self, input, dim, debug=False): + if debug: + # cumsum_cuda_kernel does not have a deterministic implementation + # so perform cumsum on cpu for debugging purposes + return torch.cumsum(input.cpu().float(), dim=dim).to(input.device) + else: + return torch.cumsum(input, dim=dim) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + parallel: int = 10, + tolerance: float = 0.1, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + debug: bool = False, + clip_skip: int = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + parallel (`int`, *optional*, defaults to 10): + The batch size to use when doing parallel sampling. More parallelism may lead to faster inference but + requires higher memory usage and can also require more total FLOPs. + tolerance (`float`, *optional*, defaults to 0.1): + The error tolerance for determining when to slide the batch window forward for parallel sampling. Lower + tolerance usually leads to less or no degradation. Higher tolerance is faster but can risk degradation + of sample quality. The tolerance is specified as a ratio of the scheduler's noise magnitude. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + debug (`bool`, *optional*, defaults to `False`): + Whether or not to run in debug mode. In debug mode, `torch.cumsum` is evaluated using the CPU. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + extra_step_kwargs.pop("generator", None) + + # # 7. Denoising loop + scheduler = self.scheduler + parallel = min(parallel, len(scheduler.timesteps)) + + begin_idx = 0 + end_idx = parallel + latents_time_evolution_buffer = torch.stack([latents] * (len(scheduler.timesteps) + 1)) + + # We must make sure the noise of stochastic schedulers such as DDPM is sampled only once per timestep. + # Sampling inside the parallel denoising loop will mess this up, so we pre-sample the noise vectors outside the denoising loop. + noise_array = torch.zeros_like(latents_time_evolution_buffer) + for j in range(len(scheduler.timesteps)): + base_noise = randn_tensor( + shape=latents.shape, generator=generator, device=latents.device, dtype=prompt_embeds.dtype + ) + noise = (self.scheduler._get_variance(scheduler.timesteps[j]) ** 0.5) * base_noise + noise_array[j] = noise.clone() + + # We specify the error tolerance as a ratio of the scheduler's noise magnitude. We similarly compute the error tolerance + # outside of the denoising loop to avoid recomputing it at every step. + # We will be dividing the norm of the noise, so we store its inverse here to avoid a division at every step. + inverse_variance_norm = 1.0 / torch.tensor( + [scheduler._get_variance(scheduler.timesteps[j]) for j in range(len(scheduler.timesteps))] + [0] + ).to(noise_array.device) + latent_dim = noise_array[0, 0].numel() + inverse_variance_norm = inverse_variance_norm[:, None] / latent_dim + + scaled_tolerance = tolerance**2 + + with self.progress_bar(total=num_inference_steps) as progress_bar: + steps = 0 + while begin_idx < len(scheduler.timesteps): + # these have shape (parallel_dim, 2*batch_size, ...) + # parallel_len is at most parallel, but could be less if we are at the end of the timesteps + # we are processing batch window of timesteps spanning [begin_idx, end_idx) + parallel_len = end_idx - begin_idx + + block_prompt_embeds = torch.stack([prompt_embeds] * parallel_len) + block_latents = latents_time_evolution_buffer[begin_idx:end_idx] + block_t = scheduler.timesteps[begin_idx:end_idx, None].repeat(1, batch_size * num_images_per_prompt) + t_vec = block_t + if do_classifier_free_guidance: + t_vec = t_vec.repeat(1, 2) + + # expand the latents if we are doing classifier free guidance + latent_model_input = ( + torch.cat([block_latents] * 2, dim=1) if do_classifier_free_guidance else block_latents + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t_vec) + + # if parallel_len is small, no need to use multiple GPUs + net = self.wrapped_unet if parallel_len > 3 else self.unet + # predict the noise residual, shape is now [parallel_len * 2 * batch_size * num_images_per_prompt, ...] + model_output = net( + latent_model_input.flatten(0, 1), + t_vec.flatten(0, 1), + encoder_hidden_states=block_prompt_embeds.flatten(0, 1), + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + per_latent_shape = model_output.shape[1:] + if do_classifier_free_guidance: + model_output = model_output.reshape( + parallel_len, 2, batch_size * num_images_per_prompt, *per_latent_shape + ) + noise_pred_uncond, noise_pred_text = model_output[:, 0], model_output[:, 1] + model_output = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + model_output = model_output.reshape( + parallel_len * batch_size * num_images_per_prompt, *per_latent_shape + ) + + block_latents_denoise = scheduler.batch_step_no_noise( + model_output=model_output, + timesteps=block_t.flatten(0, 1), + sample=block_latents.flatten(0, 1), + **extra_step_kwargs, + ).reshape(block_latents.shape) + + # back to shape (parallel_dim, batch_size, ...) + # now we want to add the pre-sampled noise + # parallel sampling algorithm requires computing the cumulative drift from the beginning + # of the window, so we need to compute cumulative sum of the deltas and the pre-sampled noises. + delta = block_latents_denoise - block_latents + cumulative_delta = self._cumsum(delta, dim=0, debug=debug) + cumulative_noise = self._cumsum(noise_array[begin_idx:end_idx], dim=0, debug=debug) + + # if we are using an ODE-like scheduler (like DDIM), we don't want to add noise + if scheduler._is_ode_scheduler: + cumulative_noise = 0 + + block_latents_new = ( + latents_time_evolution_buffer[begin_idx][None,] + cumulative_delta + cumulative_noise + ) + cur_error = torch.linalg.norm( + (block_latents_new - latents_time_evolution_buffer[begin_idx + 1 : end_idx + 1]).reshape( + parallel_len, batch_size * num_images_per_prompt, -1 + ), + dim=-1, + ).pow(2) + error_ratio = cur_error * inverse_variance_norm[begin_idx + 1 : end_idx + 1] + + # find the first index of the vector error_ratio that is greater than error tolerance + # we can shift the window for the next iteration up to this index + error_ratio = torch.nn.functional.pad( + error_ratio, (0, 0, 0, 1), value=1e9 + ) # handle the case when everything is below ratio, by padding the end of parallel_len dimension + any_error_at_time = torch.max(error_ratio > scaled_tolerance, dim=1).values.int() + ind = torch.argmax(any_error_at_time).item() + + # compute the new begin and end idxs for the window + new_begin_idx = begin_idx + min(1 + ind, parallel) + new_end_idx = min(new_begin_idx + parallel, len(scheduler.timesteps)) + + # store the computed latents for the current window in the global buffer + latents_time_evolution_buffer[begin_idx + 1 : end_idx + 1] = block_latents_new + # initialize the new sliding window latents with the end of the current window, + # should be better than random initialization + latents_time_evolution_buffer[end_idx : new_end_idx + 1] = latents_time_evolution_buffer[end_idx][ + None, + ] + + steps += 1 + + progress_bar.update(new_begin_idx - begin_idx) + if callback is not None and steps % callback_steps == 0: + callback(begin_idx, block_t[begin_idx], latents_time_evolution_buffer[begin_idx]) + + begin_idx = new_begin_idx + end_idx = new_end_idx + + latents = latents_time_evolution_buffer[-1] + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py new file mode 100644 index 0000000000000000000000000000000000000000..f88c6d8fbc3028c0612fb78c4a48fce898976628 --- /dev/null +++ b/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py @@ -0,0 +1,1314 @@ +# Copyright 2025 Pix2Pix Zero Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import ( + BlipForConditionalGeneration, + BlipProcessor, + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, +) + +from ....image_processor import PipelineImageInput, VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.attention_processor import Attention +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import DDIMScheduler, DDPMScheduler, EulerAncestralDiscreteScheduler, LMSDiscreteScheduler +from ....schedulers.scheduling_ddim_inverse import DDIMInverseScheduler +from ....utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class Pix2PixInversionPipelineOutput(BaseOutput, TextualInversionLoaderMixin): + """ + Output class for Stable Diffusion pipelines. + + Args: + latents (`torch.Tensor`) + inverted latents tensor + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + latents: torch.Tensor + images: list[PIL.Image.Image] | np.ndarray + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import requests + >>> import torch + + >>> from diffusers import DDIMScheduler, StableDiffusionPix2PixZeroPipeline + + + >>> def download(embedding_url, local_filepath): + ... r = requests.get(embedding_url) + ... with open(local_filepath, "wb") as f: + ... f.write(r.content) + + + >>> model_ckpt = "CompVis/stable-diffusion-v1-4" + >>> pipeline = StableDiffusionPix2PixZeroPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16) + >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.to("cuda") + + >>> prompt = "a high resolution painting of a cat in the style of van gough" + >>> source_emb_url = "https://hf.co/datasets/sayakpaul/sample-datasets/resolve/main/cat.pt" + >>> target_emb_url = "https://hf.co/datasets/sayakpaul/sample-datasets/resolve/main/dog.pt" + + >>> for url in [source_emb_url, target_emb_url]: + ... download(url, url.split("/")[-1]) + + >>> src_embeds = torch.load(source_emb_url.split("/")[-1]) + >>> target_embeds = torch.load(target_emb_url.split("/")[-1]) + >>> images = pipeline( + ... prompt, + ... source_embeds=src_embeds, + ... target_embeds=target_embeds, + ... num_inference_steps=50, + ... cross_attention_guidance_amount=0.15, + ... ).images + + >>> images[0].save("edited_image_dog.png") + ``` +""" + +EXAMPLE_INVERT_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from transformers import BlipForConditionalGeneration, BlipProcessor + >>> from diffusers import DDIMScheduler, DDIMInverseScheduler, StableDiffusionPix2PixZeroPipeline + + >>> import requests + >>> from PIL import Image + + >>> captioner_id = "Salesforce/blip-image-captioning-base" + >>> processor = BlipProcessor.from_pretrained(captioner_id) + >>> model = BlipForConditionalGeneration.from_pretrained( + ... captioner_id, torch_dtype=torch.float16, low_cpu_mem_usage=True + ... ) + + >>> sd_model_ckpt = "CompVis/stable-diffusion-v1-4" + >>> pipeline = StableDiffusionPix2PixZeroPipeline.from_pretrained( + ... sd_model_ckpt, + ... caption_generator=model, + ... caption_processor=processor, + ... torch_dtype=torch.float16, + ... safety_checker=None, + ... ) + + >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.enable_model_cpu_offload() + + >>> img_url = "https://github.com/pix2pixzero/pix2pix-zero/raw/main/assets/test_images/cats/cat_6.png" + + >>> raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB").resize((512, 512)) + >>> # generate caption + >>> caption = pipeline.generate_caption(raw_image) + + >>> # "a photography of a cat with flowers and dai dai daie - daie - daie kasaii" + >>> inv_latents = pipeline.invert(caption, image=raw_image).latents + >>> # we need to generate source and target embeds + + >>> source_prompts = ["a cat sitting on the street", "a cat playing in the field", "a face of a cat"] + + >>> target_prompts = ["a dog sitting on the street", "a dog playing in the field", "a face of a dog"] + + >>> source_embeds = pipeline.get_embeds(source_prompts) + >>> target_embeds = pipeline.get_embeds(target_prompts) + >>> # the latents can then be used to edit a real image + >>> # when using Stable Diffusion 2 or other models that use v-prediction + >>> # set `cross_attention_guidance_amount` to 0.01 or less to avoid input latent gradient explosion + + >>> image = pipeline( + ... caption, + ... source_embeds=source_embeds, + ... target_embeds=target_embeds, + ... num_inference_steps=50, + ... cross_attention_guidance_amount=0.15, + ... generator=generator, + ... latents=inv_latents, + ... negative_prompt=caption, + ... ).images[0] + >>> image.save("edited_image.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +def prepare_unet(unet: UNet2DConditionModel): + """Modifies the UNet (`unet`) to perform Pix2Pix Zero optimizations.""" + pix2pix_zero_attn_procs = {} + for name in unet.attn_processors.keys(): + module_name = name.replace(".processor", "") + module = unet.get_submodule(module_name) + if "attn2" in name: + pix2pix_zero_attn_procs[name] = Pix2PixZeroAttnProcessor(is_pix2pix_zero=True) + module.requires_grad_(True) + else: + pix2pix_zero_attn_procs[name] = Pix2PixZeroAttnProcessor(is_pix2pix_zero=False) + module.requires_grad_(False) + + unet.set_attn_processor(pix2pix_zero_attn_procs) + return unet + + +class Pix2PixZeroL2Loss: + def __init__(self): + self.loss = 0.0 + + def compute_loss(self, predictions, targets): + self.loss += ((predictions - targets) ** 2).sum((1, 2)).mean(0) + + +class Pix2PixZeroAttnProcessor: + """An attention processor class to store the attention weights. + In Pix2Pix Zero, it happens during computations in the cross-attention blocks.""" + + def __init__(self, is_pix2pix_zero=False): + self.is_pix2pix_zero = is_pix2pix_zero + if self.is_pix2pix_zero: + self.reference_cross_attn_map = {} + + def __call__( + self, + attn: Attention, + hidden_states, + encoder_hidden_states=None, + attention_mask=None, + timestep=None, + loss=None, + ): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + if self.is_pix2pix_zero and timestep is not None: + # new bookkeeping to save the attention weights. + if loss is None: + self.reference_cross_attn_map[timestep.item()] = attention_probs.detach().cpu() + # compute loss + elif loss is not None: + prev_attn_probs = self.reference_cross_attn_map.pop(timestep.item()) + loss.compute_loss(attention_probs, prev_attn_probs.to(attention_probs.device)) + + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +class StableDiffusionPix2PixZeroPipeline(DiffusionPipeline, StableDiffusionMixin): + r""" + Pipeline for pixel-level image editing using Pix2Pix Zero. Based on Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`], or [`DDPMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + requires_safety_checker (bool): + Whether the pipeline requires a safety checker. We recommend setting it to True if you're using the + pipeline publicly. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = [ + "safety_checker", + "feature_extractor", + "caption_generator", + "caption_processor", + "inverse_scheduler", + ] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler | DDIMScheduler | EulerAncestralDiscreteScheduler | LMSDiscreteScheduler, + feature_extractor: CLIPImageProcessor, + safety_checker: StableDiffusionSafetyChecker, + inverse_scheduler: DDIMInverseScheduler, + caption_generator: BlipForConditionalGeneration, + caption_processor: BlipProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + caption_processor=caption_processor, + caption_generator=caption_generator, + inverse_scheduler=inverse_scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + source_embeds, + target_embeds, + callback_steps, + prompt_embeds=None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if source_embeds is None and target_embeds is None: + raise ValueError("`source_embeds` and `target_embeds` cannot be undefined.") + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def generate_caption(self, images): + """Generates caption for a given image.""" + text = "a photography of" + + prev_device = self.caption_generator.device + + device = self._execution_device + inputs = self.caption_processor(images, text, return_tensors="pt").to( + device=device, dtype=self.caption_generator.dtype + ) + self.caption_generator.to(device) + outputs = self.caption_generator.generate(**inputs, max_new_tokens=128) + + # offload caption generator + self.caption_generator.to(prev_device) + + caption = self.caption_processor.batch_decode(outputs, skip_special_tokens=True)[0] + return caption + + def construct_direction(self, embs_source: torch.Tensor, embs_target: torch.Tensor): + """Constructs the edit direction to steer the image generation process semantically.""" + return (embs_target.mean(0) - embs_source.mean(0)).unsqueeze(0) + + @torch.no_grad() + def get_embeds(self, prompt: list[str], batch_size: int = 16) -> torch.Tensor: + num_prompts = len(prompt) + embeds = [] + for i in range(0, num_prompts, batch_size): + prompt_slice = prompt[i : i + batch_size] + + input_ids = self.tokenizer( + prompt_slice, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ).input_ids + + input_ids = input_ids.to(self.text_encoder.device) + embeds.append(self.text_encoder(input_ids)[0]) + + return torch.cat(embeds, dim=0).mean(0)[None] + + def prepare_image_latents(self, image, batch_size, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if isinstance(generator, list): + latents = [ + self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + latents = torch.cat(latents, dim=0) + else: + latents = self.vae.encode(image).latent_dist.sample(generator) + + latents = self.vae.config.scaling_factor * latents + + if batch_size != latents.shape[0]: + if batch_size % latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_latents_per_image = batch_size // latents.shape[0] + latents = torch.cat([latents] * additional_latents_per_image, dim=0) + else: + raise ValueError( + f"Cannot duplicate `image` of batch size {latents.shape[0]} to {batch_size} text prompts." + ) + else: + latents = torch.cat([latents], dim=0) + + return latents + + def get_epsilon(self, model_output: torch.Tensor, sample: torch.Tensor, timestep: int): + pred_type = self.inverse_scheduler.config.prediction_type + alpha_prod_t = self.inverse_scheduler.alphas_cumprod[timestep] + + beta_prod_t = 1 - alpha_prod_t + + if pred_type == "epsilon": + return model_output + elif pred_type == "sample": + return (sample - alpha_prod_t ** (0.5) * model_output) / beta_prod_t ** (0.5) + elif pred_type == "v_prediction": + return (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {pred_type} must be one of `epsilon`, `sample`, or `v_prediction`" + ) + + def auto_corr_loss(self, hidden_states, generator=None): + reg_loss = 0.0 + for i in range(hidden_states.shape[0]): + for j in range(hidden_states.shape[1]): + noise = hidden_states[i : i + 1, j : j + 1, :, :] + while True: + roll_amount = torch.randint(noise.shape[2] // 2, (1,), generator=generator).item() + reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=2)).mean() ** 2 + reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=3)).mean() ** 2 + + if noise.shape[2] <= 8: + break + noise = F.avg_pool2d(noise, kernel_size=2) + return reg_loss + + def kl_divergence(self, hidden_states): + mean = hidden_states.mean() + var = hidden_states.var() + return var + mean**2 - 1 - torch.log(var + 1e-7) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + source_embeds: torch.Tensor = None, + target_embeds: torch.Tensor = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + cross_attention_guidance_amount: float = 0.1, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int | None = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + source_embeds (`torch.Tensor`): + Source concept embeddings. Generation of the embeddings as per the [original + paper](https://huggingface.co/papers/2302.03027). Used in discovering the edit direction. + target_embeds (`torch.Tensor`): + Target concept embeddings. Generation of the embeddings as per the [original + paper](https://huggingface.co/papers/2302.03027). Used in discovering the edit direction. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + cross_attention_guidance_amount (`float`, defaults to 0.1): + Amount of guidance needed from the reference cross-attention maps. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + # 0. Define the spatial resolutions. + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + source_embeds, + target_embeds, + callback_steps, + prompt_embeds, + ) + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Generate the inverted noise from the input image or any other image + # generated from the input prompt. + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + latents_init = latents.clone() + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Rejig the UNet so that we can obtain the cross-attenion maps and + # use them for guiding the subsequent image generation. + self.unet = prepare_unet(self.unet) + + # 7. Denoising loop where we obtain the cross-attention maps. + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs={"timestep": t}, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + # 8. Compute the edit directions. + edit_direction = self.construct_direction(source_embeds, target_embeds).to(prompt_embeds.device) + + # 9. Edit the prompt embeddings as per the edit directions discovered. + prompt_embeds_edit = prompt_embeds.clone() + prompt_embeds_edit[1:2] += edit_direction + + # 10. Second denoising loop to generate the edited image. + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + latents = latents_init + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # we want to learn the latent such that it steers the generation + # process towards the edited direction, so make the make initial + # noise learnable + x_in = latent_model_input.detach().clone() + x_in.requires_grad = True + + # optimizer + opt = torch.optim.SGD([x_in], lr=cross_attention_guidance_amount) + + with torch.enable_grad(): + # initialize loss + loss = Pix2PixZeroL2Loss() + + # predict the noise residual + noise_pred = self.unet( + x_in, + t, + encoder_hidden_states=prompt_embeds_edit.detach(), + cross_attention_kwargs={"timestep": t, "loss": loss}, + ).sample + + loss.loss.backward(retain_graph=False) + opt.step() + + # recompute the noise + noise_pred = self.unet( + x_in.detach(), + t, + encoder_hidden_states=prompt_embeds_edit, + cross_attention_kwargs={"timestep": None}, + ).sample + + latents = x_in.detach().chunk(2)[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_INVERT_DOC_STRING) + def invert( + self, + prompt: str | None = None, + image: PipelineImageInput = None, + num_inference_steps: int = 50, + guidance_scale: float = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + cross_attention_guidance_amount: float = 0.1, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int | None = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + lambda_auto_corr: float = 20.0, + lambda_kl: float = 20.0, + num_reg_steps: int = 5, + num_auto_corr_rolls: int = 5, + ): + r""" + Function used to generate inverted latents given a prompt and image. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor` `np.ndarray`, `PIL.Image.Image`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch which will be used for conditioning. Can also accept + image latents as `image`, if passing latents directly, it will not be encoded again. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 1): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + cross_attention_guidance_amount (`float`, defaults to 0.1): + Amount of guidance needed from the reference cross-attention maps. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + lambda_auto_corr (`float`, *optional*, defaults to 20.0): + Lambda parameter to control auto correction + lambda_kl (`float`, *optional*, defaults to 20.0): + Lambda parameter to control Kullback–Leibler divergence output + num_reg_steps (`int`, *optional*, defaults to 5): + Number of regularization loss steps + num_auto_corr_rolls (`int`, *optional*, defaults to 5): + Number of auto correction roll steps + + Examples: + + Returns: + [`~pipelines.stable_diffusion.pipeline_stable_diffusion_pix2pix_zero.Pix2PixInversionPipelineOutput`] or + `tuple`: + [`~pipelines.stable_diffusion.pipeline_stable_diffusion_pix2pix_zero.Pix2PixInversionPipelineOutput`] if + `return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is the inverted + latents tensor and then second is the corresponding decoded image. + """ + # 1. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Preprocess image + image = self.image_processor.preprocess(image) + + # 4. Prepare latent variables + latents = self.prepare_image_latents(image, batch_size, self.vae.dtype, device, generator) + + # 5. Encode input prompt + num_images_per_prompt = 1 + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.inverse_scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.inverse_scheduler.timesteps + + # 6. Rejig the UNet so that we can obtain the cross-attenion maps and + # use them for guiding the subsequent image generation. + self.unet = prepare_unet(self.unet) + + # 7. Denoising loop where we obtain the cross-attention maps. + num_warmup_steps = len(timesteps) - num_inference_steps * self.inverse_scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.inverse_scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs={"timestep": t}, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # regularization of the noise prediction + with torch.enable_grad(): + for _ in range(num_reg_steps): + if lambda_auto_corr > 0: + for _ in range(num_auto_corr_rolls): + var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True) + + # Derive epsilon from model output before regularizing to IID standard normal + var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t) + + l_ac = self.auto_corr_loss(var_epsilon, generator=generator) + l_ac.backward() + + grad = var.grad.detach() / num_auto_corr_rolls + noise_pred = noise_pred - lambda_auto_corr * grad + + if lambda_kl > 0: + var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True) + + # Derive epsilon from model output before regularizing to IID standard normal + var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t) + + l_kld = self.kl_divergence(var_epsilon) + l_kld.backward() + + grad = var.grad.detach() + noise_pred = noise_pred - lambda_kl * grad + + noise_pred = noise_pred.detach() + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.inverse_scheduler.step(noise_pred, t, latents).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.inverse_scheduler.order == 0 + ): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + inverted_latents = latents.detach().clone() + + # 8. Post-processing + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (inverted_latents, image) + + return Pix2PixInversionPipelineOutput(latents=inverted_latents, images=image) diff --git a/diffusers/pipelines/deprecated/stochastic_karras_ve/__init__.py b/diffusers/pipelines/deprecated/stochastic_karras_ve/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..15c9a8c27f98dd7e1913bd57dfd5e8dae71172b4 --- /dev/null +++ b/diffusers/pipelines/deprecated/stochastic_karras_ve/__init__.py @@ -0,0 +1,19 @@ +from typing import TYPE_CHECKING + +from ....utils import DIFFUSERS_SLOW_IMPORT, _LazyModule + + +_import_structure = {"pipeline_stochastic_karras_ve": ["KarrasVePipeline"]} + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + from .pipeline_stochastic_karras_ve import KarrasVePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) diff --git a/diffusers/pipelines/deprecated/stochastic_karras_ve/pipeline_stochastic_karras_ve.py b/diffusers/pipelines/deprecated/stochastic_karras_ve/pipeline_stochastic_karras_ve.py new file mode 100644 index 0000000000000000000000000000000000000000..16f1265d0f69c37ef9a71241729e99c106de70a1 --- /dev/null +++ b/diffusers/pipelines/deprecated/stochastic_karras_ve/pipeline_stochastic_karras_ve.py @@ -0,0 +1,126 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch + +from ....models import UNet2DModel +from ....schedulers import KarrasVeScheduler +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +class KarrasVePipeline(DiffusionPipeline): + r""" + Pipeline for unconditional image generation. + + Parameters: + unet ([`UNet2DModel`]): + A `UNet2DModel` to denoise the encoded image. + scheduler ([`KarrasVeScheduler`]): + A scheduler to be used in combination with `unet` to denoise the encoded image. + """ + + # add type hints for linting + unet: UNet2DModel + scheduler: KarrasVeScheduler + + def __init__(self, unet: UNet2DModel, scheduler: KarrasVeScheduler): + super().__init__() + self.register_modules(unet=unet, scheduler=scheduler) + + @torch.no_grad() + def __call__( + self, + batch_size: int = 1, + num_inference_steps: int = 50, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + **kwargs, + ) -> tuple | ImagePipelineOutput: + r""" + The call function to the pipeline for generation. + + Args: + batch_size (`int`, *optional*, defaults to 1): + The number of images to generate. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple. + + Example: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + img_size = self.unet.config.sample_size + shape = (batch_size, 3, img_size, img_size) + + model = self.unet + + # sample x_0 ~ N(0, sigma_0^2 * I) + sample = randn_tensor(shape, generator=generator, device=self.device) * self.scheduler.init_noise_sigma + + self.scheduler.set_timesteps(num_inference_steps) + + for t in self.progress_bar(self.scheduler.timesteps): + # here sigma_t == t_i from the paper + sigma = self.scheduler.schedule[t] + sigma_prev = self.scheduler.schedule[t - 1] if t > 0 else 0 + + # 1. Select temporarily increased noise level sigma_hat + # 2. Add new noise to move from sample_i to sample_hat + sample_hat, sigma_hat = self.scheduler.add_noise_to_input(sample, sigma, generator=generator) + + # 3. Predict the noise residual given the noise magnitude `sigma_hat` + # The model inputs and output are adjusted by following eq. (213) in [1]. + model_output = (sigma_hat / 2) * model((sample_hat + 1) / 2, sigma_hat / 2).sample + + # 4. Evaluate dx/dt at sigma_hat + # 5. Take Euler step from sigma to sigma_prev + step_output = self.scheduler.step(model_output, sigma_hat, sigma_prev, sample_hat) + + if sigma_prev != 0: + # 6. Apply 2nd order correction + # The model inputs and output are adjusted by following eq. (213) in [1]. + model_output = (sigma_prev / 2) * model((step_output.prev_sample + 1) / 2, sigma_prev / 2).sample + step_output = self.scheduler.step_correct( + model_output, + sigma_hat, + sigma_prev, + sample_hat, + step_output.prev_sample, + step_output["derivative"], + ) + sample = step_output.prev_sample + + sample = (sample / 2 + 0.5).clamp(0, 1) + image = sample.cpu().permute(0, 2, 3, 1).numpy() + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/deprecated/text_to_video_synthesis/__init__.py b/diffusers/pipelines/deprecated/text_to_video_synthesis/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6c32f5f16cae8d121d82d296efcc5e109a7ea836 --- /dev/null +++ b/diffusers/pipelines/deprecated/text_to_video_synthesis/__init__.py @@ -0,0 +1,54 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_output"] = ["TextToVideoSDPipelineOutput"] + _import_structure["pipeline_text_to_video_synth"] = ["TextToVideoSDPipeline"] + _import_structure["pipeline_text_to_video_synth_img2img"] = ["VideoToVideoSDPipeline"] + _import_structure["pipeline_text_to_video_zero"] = ["TextToVideoZeroPipeline"] + _import_structure["pipeline_text_to_video_zero_sdxl"] = ["TextToVideoZeroSDXLPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_output import TextToVideoSDPipelineOutput + from .pipeline_text_to_video_synth import TextToVideoSDPipeline + from .pipeline_text_to_video_synth_img2img import VideoToVideoSDPipeline + from .pipeline_text_to_video_zero import TextToVideoZeroPipeline + from .pipeline_text_to_video_zero_sdxl import TextToVideoZeroSDXLPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_output.py b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..c93609a2dd6f7b0f0d9f8cdd7bc2988b98858458 --- /dev/null +++ b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_output.py @@ -0,0 +1,25 @@ +from dataclasses import dataclass + +import numpy as np +import PIL +import torch + +from ....utils import ( + BaseOutput, +) + + +@dataclass +class TextToVideoSDPipelineOutput(BaseOutput): + """ + Output class for text-to-video pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised + PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)` + """ + + frames: torch.Tensor | np.ndarray | list[list[PIL.Image.Image]] diff --git a/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_synth.py b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_synth.py new file mode 100644 index 0000000000000000000000000000000000000000..b3d52f1de27984361a9a87f9daa91f8c1a192a22 --- /dev/null +++ b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_synth.py @@ -0,0 +1,657 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet3DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ....video_processor import VideoProcessor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from . import TextToVideoSDPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import TextToVideoSDPipeline + >>> from diffusers.utils import export_to_video + + >>> pipe = TextToVideoSDPipeline.from_pretrained( + ... "damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16" + ... ) + >>> pipe.enable_model_cpu_offload() + + >>> prompt = "Spiderman is surfing" + >>> video_frames = pipe(prompt).frames[0] + >>> video_path = export_to_video(video_frames) + >>> video_path + ``` +""" + + +class TextToVideoSDPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for text-to-video generation. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`] to tokenize text. + unet ([`UNet3DConditionModel`]): + A [`UNet3DConditionModel`] to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet3DConditionModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def decode_latents(self, latents): + latents = 1 / self.vae.config.scaling_factor * latents + + batch_size, channels, num_frames, height, width = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width) + + image = self.vae.decode(latents).sample + video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + video = video.float() + return video + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents( + self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None + ): + shape = ( + batch_size, + num_channels_latents, + num_frames, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_frames: int = 16, + num_inference_steps: int = 50, + guidance_scale: float = 9.0, + negative_prompt: str | list[str] | None = None, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated video. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated video. + num_frames (`int`, *optional*, defaults to 16): + The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds + amounts to 2 seconds of video. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality videos at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. Latents should be of shape + `(batch_size, num_channel, num_frames, height, width)`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between `torch.Tensor` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated frames. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + num_images_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + num_frames, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # reshape latents + bsz, channel, frames, width, height = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # reshape latents back + latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post processing + if output_type == "latent": + video = latents + else: + video_tensor = self.decode_latents(latents) + video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type) + + # 9. Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return TextToVideoSDPipelineOutput(frames=video) diff --git a/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..c0d71c6ffd885692c96127125e4d8f904d63540f --- /dev/null +++ b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py @@ -0,0 +1,716 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet3DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ....video_processor import VideoProcessor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from . import TextToVideoSDPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler + >>> from diffusers.utils import export_to_video + + >>> pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16) + >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config) + >>> pipe.to("cuda") + + >>> prompt = "spiderman running in the desert" + >>> video_frames = pipe(prompt, num_inference_steps=40, height=320, width=576, num_frames=24).frames[0] + >>> # safe low-res video + >>> video_path = export_to_video(video_frames, output_video_path="./video_576_spiderman.mp4") + + >>> # let's offload the text-to-image model + >>> pipe.to("cpu") + + >>> # and load the image-to-image model + >>> pipe = DiffusionPipeline.from_pretrained( + ... "cerspense/zeroscope_v2_XL", torch_dtype=torch.float16, revision="refs/pr/15" + ... ) + >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config) + >>> pipe.enable_model_cpu_offload() + + >>> # The VAE consumes A LOT of memory, let's make sure we run it in sliced mode + >>> pipe.vae.enable_slicing() + + >>> # now let's upscale it + >>> video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames] + + >>> # and denoise it + >>> video_frames = pipe(prompt, video=video, strength=0.6).frames[0] + >>> video_path = export_to_video(video_frames, output_video_path="./video_1024_spiderman.mp4") + >>> video_path + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class VideoToVideoSDPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for text-guided video-to-video generation. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`] to tokenize text. + unet ([`UNet3DConditionModel`]): + A [`UNet3DConditionModel`] to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet3DConditionModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents + def decode_latents(self, latents): + latents = 1 / self.vae.config.scaling_factor * latents + + batch_size, channels, num_frames, height, width = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width) + + image = self.vae.decode(latents).sample + video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + video = video.float() + return video + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, video, timestep, batch_size, dtype, device, generator=None): + video = video.to(device=device, dtype=dtype) + + # change from (b, c, f, h, w) -> (b * f, c, w, h) + bsz, channel, frames, width, height = video.shape + video = video.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + + if video.shape[1] == 4: + init_latents = video + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + elif isinstance(generator, list): + init_latents = [ + retrieve_latents(self.vae.encode(video[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(video), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `video` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + latents = latents[None, :].reshape((bsz, frames, latents.shape[1]) + latents.shape[2:]).permute(0, 2, 1, 3, 4) + + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + video: list[np.ndarray] | torch.Tensor = None, + strength: float = 0.6, + num_inference_steps: int = 50, + guidance_scale: float = 15.0, + negative_prompt: str | list[str] | None = None, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + video (`list[np.ndarray]` or `torch.Tensor`): + `video` frames or tensor representing a video batch to be used as the starting point for the process. + Can also accept video latents as `image`, if passing latents directly, it will not be encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `video`. Must be between 0 and 1. `video` is used as a + starting point, adding more noise to it the larger the `strength`. The number of denoising steps + depends on the amount of noise initially added. When `strength` is 1, added noise is maximum and the + denoising process runs for the full number of iterations specified in `num_inference_steps`. A value of + 1 essentially ignores `video`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality videos at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in video generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. Latents should be of shape + `(batch_size, num_channel, num_frames, height, width)`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between `torch.Tensor` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated frames. + """ + # 0. Default height and width to unet + num_images_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Preprocess video + video = self.video_processor.preprocess_video(video) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents = self.prepare_latents(video, latent_timestep, batch_size, prompt_embeds.dtype, device, generator) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # reshape latents + bsz, channel, frames, width, height = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # reshape latents back + latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + + # 9. Post processing + if output_type == "latent": + video = latents + else: + video_tensor = self.decode_latents(latents) + video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type) + + # 10. Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return TextToVideoSDPipelineOutput(frames=video) diff --git a/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_zero.py b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_zero.py new file mode 100644 index 0000000000000000000000000000000000000000..6ea24ae2c81711315ce92a16b85f9877d432b6c9 --- /dev/null +++ b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_zero.py @@ -0,0 +1,1004 @@ +import copy +import inspect +from dataclasses import dataclass +from typing import Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from torch.nn.functional import grid_sample +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ....image_processor import VaeImageProcessor +from ....loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + BaseOutput, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import empty_device_cache, randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ...stable_diffusion import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def rearrange_0(tensor, f): + F, C, H, W = tensor.size() + tensor = torch.permute(torch.reshape(tensor, (F // f, f, C, H, W)), (0, 2, 1, 3, 4)) + return tensor + + +def rearrange_1(tensor): + B, C, F, H, W = tensor.size() + return torch.reshape(torch.permute(tensor, (0, 2, 1, 3, 4)), (B * F, C, H, W)) + + +def rearrange_3(tensor, f): + F, D, C = tensor.size() + return torch.reshape(tensor, (F // f, f, D, C)) + + +def rearrange_4(tensor): + B, F, D, C = tensor.size() + return torch.reshape(tensor, (B * F, D, C)) + + +class CrossFrameAttnProcessor: + """ + Cross frame attention processor. Each frame attends the first frame. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = key.size()[0] // self.batch_size + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +class CrossFrameAttnProcessor2_0: + """ + Cross frame attention processor with scaled_dot_product attention of Pytorch 2.0. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + inner_dim = hidden_states.shape[-1] + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = max(1, key.size()[0] // self.batch_size) + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + head_dim = inner_dim // attn.heads + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + return hidden_states + + +@dataclass +class TextToVideoPipelineOutput(BaseOutput): + r""" + Output class for zero-shot text-to-video pipeline. + + Args: + images (`[list[PIL.Image.Image]`, `np.ndarray`]): + list of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + nsfw_content_detected (`[list[bool]]`): + list indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or + `None` if safety checking could not be performed. + """ + + images: list[PIL.Image.Image] | np.ndarray + nsfw_content_detected: list[bool] | None + + +def coords_grid(batch, ht, wd, device): + # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py + coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device)) + coords = torch.stack(coords[::-1], dim=0).float() + return coords[None].repeat(batch, 1, 1, 1) + + +def warp_single_latent(latent, reference_flow): + """ + Warp latent of a single frame with given flow + + Args: + latent: latent code of a single frame + reference_flow: flow which to warp the latent with + + Returns: + warped: warped latent + """ + _, _, H, W = reference_flow.size() + _, _, h, w = latent.size() + coords0 = coords_grid(1, H, W, device=latent.device).to(latent.dtype) + + coords_t0 = coords0 + reference_flow + coords_t0[:, 0] /= W + coords_t0[:, 1] /= H + + coords_t0 = coords_t0 * 2.0 - 1.0 + coords_t0 = F.interpolate(coords_t0, size=(h, w), mode="bilinear") + coords_t0 = torch.permute(coords_t0, (0, 2, 3, 1)) + + warped = grid_sample(latent, coords_t0, mode="nearest", padding_mode="reflection") + return warped + + +def create_motion_field(motion_field_strength_x, motion_field_strength_y, frame_ids, device, dtype): + """ + Create translation motion field + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + device: device + dtype: dtype + + Returns: + + """ + seq_length = len(frame_ids) + reference_flow = torch.zeros((seq_length, 2, 512, 512), device=device, dtype=dtype) + for fr_idx in range(seq_length): + reference_flow[fr_idx, 0, :, :] = motion_field_strength_x * (frame_ids[fr_idx]) + reference_flow[fr_idx, 1, :, :] = motion_field_strength_y * (frame_ids[fr_idx]) + return reference_flow + + +def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_strength_y, frame_ids, latents): + """ + Creates translation motion and warps the latents accordingly + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + latents: latent codes of frames + + Returns: + warped_latents: warped latents + """ + motion_field = create_motion_field( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + frame_ids=frame_ids, + device=latents.device, + dtype=latents.dtype, + ) + warped_latents = latents.clone().detach() + for i in range(len(warped_latents)): + warped_latents[i] = warp_single_latent(latents[i][None], motion_field[i][None]) + return warped_latents + + +class TextToVideoZeroPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for zero-shot text-to-video generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`] to tokenize text. + unet ([`UNet2DConditionModel`]): + A [`UNet3DConditionModel`] to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`CLIPImageProcessor`]): + A [`CLIPImageProcessor`] to extract features from generated images; used as inputs to the `safety_checker`. + """ + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + def forward_loop(self, x_t0, t0, t1, generator): + """ + Perform DDPM forward process from time t0 to t1. This is the same as adding noise with corresponding variance. + + Args: + x_t0: + Latent code at time t0. + t0: + Timestep at t0. + t1: + Timestamp at t1. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + + Returns: + x_t1: + Forward process applied to x_t0 from time t0 to t1. + """ + eps = randn_tensor(x_t0.size(), generator=generator, dtype=x_t0.dtype, device=x_t0.device) + alpha_vec = torch.prod(self.scheduler.alphas[t0:t1]) + x_t1 = torch.sqrt(alpha_vec) * x_t0 + torch.sqrt(1 - alpha_vec) * eps + return x_t1 + + def backward_loop( + self, + latents, + timesteps, + prompt_embeds, + guidance_scale, + callback, + callback_steps, + num_warmup_steps, + extra_step_kwargs, + cross_attention_kwargs=None, + ): + """ + Perform backward process given list of time steps. + + Args: + latents: + Latents at time timesteps[0]. + timesteps: + Time steps along which to perform backward process. + prompt_embeds: + Pre-generated text embeddings. + guidance_scale: + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + extra_step_kwargs: + Extra_step_kwargs. + cross_attention_kwargs: + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + num_warmup_steps: + number of warmup steps. + + Returns: + latents: + Latents of backward process output at time timesteps[-1]. + """ + do_classifier_free_guidance = guidance_scale > 1.0 + num_steps = (len(timesteps) - num_warmup_steps) // self.scheduler.order + with self.progress_bar(total=num_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + return latents.clone().detach() + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + video_length: int | None = 8, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_videos_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + motion_field_strength_x: float = 12, + motion_field_strength_y: float = 12, + output_type: str | None = "tensor", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int | None = 1, + t0: int = 44, + t1: int = 47, + frame_ids: list[int] | None = None, + ): + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + video_length (`int`, *optional*, defaults to 8): + The number of generated video frames. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in video generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between `"latent"` and `"np"`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a + [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`] instead of + a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + motion_field_strength_x (`float`, *optional*, defaults to 12): + Strength of motion in generated video along x-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + motion_field_strength_y (`float`, *optional*, defaults to 12): + Strength of motion in generated video along y-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + t0 (`int`, *optional*, defaults to 44): + Timestep t0. Should be in the range [0, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + t1 (`int`, *optional*, defaults to 47): + Timestep t0. Should be in the range [t0 + 1, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + frame_ids (`list[int]`, *optional*): + Indexes of the frames that are being generated. This is used when generating longer videos + chunk-by-chunk. + + Returns: + [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`]: + The output contains a `ndarray` of the generated video, when `output_type` != `"latent"`, otherwise a + latent code of generated videos and a list of `bool`s indicating whether the corresponding generated + video contains "not-safe-for-work" (nsfw) content.. + """ + assert video_length > 0 + if frame_ids is None: + frame_ids = list(range(video_length)) + assert len(frame_ids) == video_length + + assert num_videos_per_prompt == 1 + + # set the processor + original_attn_proc = self.unet.attn_processors + processor = ( + CrossFrameAttnProcessor2_0(batch_size=2) + if hasattr(F, "scaled_dot_product_attention") + else CrossFrameAttnProcessor(batch_size=2) + ) + self.unet.set_attn_processor(processor) + + if isinstance(prompt, str): + prompt = [prompt] + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + + # Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # Encode input prompt + prompt_embeds_tuple = self.encode_prompt( + prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt + ) + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + # Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + # Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + # Perform the first backward process up to time T_1 + x_1_t1 = self.backward_loop( + timesteps=timesteps[: -t1 - 1], + prompt_embeds=prompt_embeds, + latents=latents, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=num_warmup_steps, + ) + scheduler_copy = copy.deepcopy(self.scheduler) + + # Perform the second backward process up to time T_0 + x_1_t0 = self.backward_loop( + timesteps=timesteps[-t1 - 1 : -t0 - 1], + prompt_embeds=prompt_embeds, + latents=x_1_t1, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + ) + + # Propagate first frame latents at time T_0 to remaining frames + x_2k_t0 = x_1_t0.repeat(video_length - 1, 1, 1, 1) + + # Add motion in latents at time T_0 + x_2k_t0 = create_motion_field_and_warp_latents( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + latents=x_2k_t0, + frame_ids=frame_ids[1:], + ) + + # Perform forward process up to time T_1 + x_2k_t1 = self.forward_loop( + x_t0=x_2k_t0, + t0=timesteps[-t0 - 1].item(), + t1=timesteps[-t1 - 1].item(), + generator=generator, + ) + + # Perform backward process from time T_1 to 0 + x_1k_t1 = torch.cat([x_1_t1, x_2k_t1]) + b, l, d = prompt_embeds.size() + prompt_embeds = prompt_embeds[:, None].repeat(1, video_length, 1, 1).reshape(b * video_length, l, d) + + self.scheduler = scheduler_copy + x_1k_0 = self.backward_loop( + timesteps=timesteps[-t1 - 1 :], + prompt_embeds=prompt_embeds, + latents=x_1k_t1, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + ) + latents = x_1k_0 + + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + empty_device_cache() + + if output_type == "latent": + image = latents + has_nsfw_concept = None + else: + image = self.decode_latents(latents) + # Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # Offload all models + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return TextToVideoPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def decode_latents(self, latents): + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image diff --git a/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py new file mode 100644 index 0000000000000000000000000000000000000000..9af63e5044bdf3d564c014644668fac903e53380 --- /dev/null +++ b/diffusers/pipelines/deprecated/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py @@ -0,0 +1,1340 @@ +import copy +import inspect +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL +import torch +import torch.nn.functional as F +from torch.nn.functional import grid_sample +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ....image_processor import VaeImageProcessor +from ....loaders import StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL, UNet2DConditionModel +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + is_invisible_watermark_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin + + +if is_invisible_watermark_available(): + from ...stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + + +from ....utils import is_torch_xla_available + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_0 +def rearrange_0(tensor, f): + F, C, H, W = tensor.size() + tensor = torch.permute(torch.reshape(tensor, (F // f, f, C, H, W)), (0, 2, 1, 3, 4)) + return tensor + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_1 +def rearrange_1(tensor): + B, C, F, H, W = tensor.size() + return torch.reshape(torch.permute(tensor, (0, 2, 1, 3, 4)), (B * F, C, H, W)) + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_3 +def rearrange_3(tensor, f): + F, D, C = tensor.size() + return torch.reshape(tensor, (F // f, f, D, C)) + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_4 +def rearrange_4(tensor): + B, F, D, C = tensor.size() + return torch.reshape(tensor, (B * F, D, C)) + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor +class CrossFrameAttnProcessor: + """ + Cross frame attention processor. Each frame attends the first frame. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = key.size()[0] // self.batch_size + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor2_0 +class CrossFrameAttnProcessor2_0: + """ + Cross frame attention processor with scaled_dot_product attention of Pytorch 2.0. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + inner_dim = hidden_states.shape[-1] + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = max(1, key.size()[0] // self.batch_size) + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + head_dim = inner_dim // attn.heads + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + return hidden_states + + +@dataclass +class TextToVideoSDXLPipelineOutput(BaseOutput): + """ + Output class for zero-shot text-to-video pipeline. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.coords_grid +def coords_grid(batch, ht, wd, device): + # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py + coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device)) + coords = torch.stack(coords[::-1], dim=0).float() + return coords[None].repeat(batch, 1, 1, 1) + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.warp_single_latent +def warp_single_latent(latent, reference_flow): + """ + Warp latent of a single frame with given flow + + Args: + latent: latent code of a single frame + reference_flow: flow which to warp the latent with + + Returns: + warped: warped latent + """ + _, _, H, W = reference_flow.size() + _, _, h, w = latent.size() + coords0 = coords_grid(1, H, W, device=latent.device).to(latent.dtype) + + coords_t0 = coords0 + reference_flow + coords_t0[:, 0] /= W + coords_t0[:, 1] /= H + + coords_t0 = coords_t0 * 2.0 - 1.0 + coords_t0 = F.interpolate(coords_t0, size=(h, w), mode="bilinear") + coords_t0 = torch.permute(coords_t0, (0, 2, 3, 1)) + + warped = grid_sample(latent, coords_t0, mode="nearest", padding_mode="reflection") + return warped + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.create_motion_field +def create_motion_field(motion_field_strength_x, motion_field_strength_y, frame_ids, device, dtype): + """ + Create translation motion field + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + device: device + dtype: dtype + + Returns: + + """ + seq_length = len(frame_ids) + reference_flow = torch.zeros((seq_length, 2, 512, 512), device=device, dtype=dtype) + for fr_idx in range(seq_length): + reference_flow[fr_idx, 0, :, :] = motion_field_strength_x * (frame_ids[fr_idx]) + reference_flow[fr_idx, 1, :, :] = motion_field_strength_y * (frame_ids[fr_idx]) + return reference_flow + + +# Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.create_motion_field_and_warp_latents +def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_strength_y, frame_ids, latents): + """ + Creates translation motion and warps the latents accordingly + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + latents: latent codes of frames + + Returns: + warped_latents: warped latents + """ + motion_field = create_motion_field( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + frame_ids=frame_ids, + device=latents.device, + dtype=latents.dtype, + ) + warped_latents = latents.clone().detach() + for i in range(len(warped_latents)): + warped_latents[i] = warp_single_latent(latents[i][None], motion_field[i][None]) + return warped_latents + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class TextToVideoZeroSDXLPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for zero-shot text-to-video generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + ): + super().__init__() + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoZeroPipeline.forward_loop + def forward_loop(self, x_t0, t0, t1, generator): + """ + Perform DDPM forward process from time t0 to t1. This is the same as adding noise with corresponding variance. + + Args: + x_t0: + Latent code at time t0. + t0: + Timestep at t0. + t1: + Timestamp at t1. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + + Returns: + x_t1: + Forward process applied to x_t0 from time t0 to t1. + """ + eps = randn_tensor(x_t0.size(), generator=generator, dtype=x_t0.dtype, device=x_t0.device) + alpha_vec = torch.prod(self.scheduler.alphas[t0:t1]) + x_t1 = torch.sqrt(alpha_vec) * x_t0 + torch.sqrt(1 - alpha_vec) * eps + return x_t1 + + def backward_loop( + self, + latents, + timesteps, + prompt_embeds, + guidance_scale, + callback, + callback_steps, + num_warmup_steps, + extra_step_kwargs, + add_text_embeds, + add_time_ids, + cross_attention_kwargs=None, + guidance_rescale: float = 0.0, + ): + """ + Perform backward process given list of time steps + + Args: + latents: + Latents at time timesteps[0]. + timesteps: + Time steps along which to perform backward process. + prompt_embeds: + Pre-generated text embeddings. + guidance_scale: + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + extra_step_kwargs: + Extra_step_kwargs. + cross_attention_kwargs: + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + num_warmup_steps: + number of warmup steps. + + Returns: + latents: latents of backward process output at time timesteps[-1] + """ + + do_classifier_free_guidance = guidance_scale > 1.0 + num_steps = (len(timesteps) - num_warmup_steps) // self.scheduler.order + + with self.progress_bar(total=num_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + return latents.clone().detach() + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + video_length: int | None = 8, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + denoising_end: float | None = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_videos_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + frame_ids: list[int] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + latents: torch.Tensor | None = None, + motion_field_strength_x: float = 12, + motion_field_strength_y: float = 12, + output_type: str | None = "tensor", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + t0: int = 44, + t1: int = 47, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + video_length (`int`, *optional*, defaults to 8): + The number of generated video frames. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + frame_ids (`list[int]`, *optional*): + Indexes of the frames that are being generated. This is used when generating longer videos + chunk-by-chunk. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + motion_field_strength_x (`float`, *optional*, defaults to 12): + Strength of motion in generated video along x-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + motion_field_strength_y (`float`, *optional*, defaults to 12): + Strength of motion in generated video along y-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py). + guidance_rescale (`float`, *optional*, defaults to 0.7): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + t0 (`int`, *optional*, defaults to 44): + Timestep t0. Should be in the range [0, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + t1 (`int`, *optional*, defaults to 47): + Timestep t0. Should be in the range [t0 + 1, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + + Returns: + [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoSDXLPipelineOutput`] or + `tuple`: [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoSDXLPipelineOutput`] + if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + assert video_length > 0 + if frame_ids is None: + frame_ids = list(range(video_length)) + assert len(frame_ids) == video_length + + assert num_videos_per_prompt == 1 + + # set the processor + original_attn_proc = self.unet.attn_processors + processor = ( + CrossFrameAttnProcessor2_0(batch_size=2) + if hasattr(F, "scaled_dot_product_attention") + else CrossFrameAttnProcessor(batch_size=2) + ) + self.unet.set_attn_processor(processor) + + if isinstance(prompt, str): + prompt = [prompt] + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + # 2. Define call parameters + batch_size = ( + 1 if isinstance(prompt, str) else len(prompt) if isinstance(prompt, list) else prompt_embeds.shape[0] + ) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_videos_per_prompt, 1) + + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + # Perform the first backward process up to time T_1 + x_1_t1 = self.backward_loop( + timesteps=timesteps[: -t1 - 1], + prompt_embeds=prompt_embeds, + latents=latents, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=num_warmup_steps, + add_text_embeds=add_text_embeds, + add_time_ids=add_time_ids, + ) + + scheduler_copy = copy.deepcopy(self.scheduler) + + # Perform the second backward process up to time T_0 + x_1_t0 = self.backward_loop( + timesteps=timesteps[-t1 - 1 : -t0 - 1], + prompt_embeds=prompt_embeds, + latents=x_1_t1, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + add_text_embeds=add_text_embeds, + add_time_ids=add_time_ids, + ) + + # Propagate first frame latents at time T_0 to remaining frames + x_2k_t0 = x_1_t0.repeat(video_length - 1, 1, 1, 1) + + # Add motion in latents at time T_0 + x_2k_t0 = create_motion_field_and_warp_latents( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + latents=x_2k_t0, + frame_ids=frame_ids[1:], + ) + + # Perform forward process up to time T_1 + x_2k_t1 = self.forward_loop( + x_t0=x_2k_t0, + t0=timesteps[-t0 - 1].to(torch.long), + t1=timesteps[-t1 - 1].to(torch.long), + generator=generator, + ) + + # Perform backward process from time T_1 to 0 + latents = torch.cat([x_1_t1, x_2k_t1]) + + self.scheduler = scheduler_copy + timesteps = timesteps[-t1 - 1 :] + + b, l, d = prompt_embeds.size() + prompt_embeds = prompt_embeds[:, None].repeat(1, video_length, 1, 1).reshape(b * video_length, l, d) + + b, k = add_text_embeds.size() + add_text_embeds = add_text_embeds[:, None].repeat(1, video_length, 1).reshape(b * video_length, k) + + b, k = add_time_ids.size() + add_time_ids = add_time_ids[:, None].repeat(1, video_length, 1).reshape(b * video_length, k) + + # 7.1 Apply denoising_end + if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + x_1k_0 = self.backward_loop( + timesteps=timesteps, + prompt_embeds=prompt_embeds, + latents=latents, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + add_text_embeds=add_text_embeds, + add_time_ids=add_time_ids, + ) + + latents = x_1k_0 + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + return TextToVideoSDXLPipelineOutput(images=image) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return TextToVideoSDXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/deprecated/unclip/__init__.py b/diffusers/pipelines/deprecated/unclip/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7444df4912734949b25e10ad93426119a9a588cf --- /dev/null +++ b/diffusers/pipelines/deprecated/unclip/__init__.py @@ -0,0 +1,52 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, + is_transformers_available, + is_transformers_version, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import UnCLIPImageVariationPipeline, UnCLIPPipeline + + _dummy_objects.update( + {"UnCLIPImageVariationPipeline": UnCLIPImageVariationPipeline, "UnCLIPPipeline": UnCLIPPipeline} + ) +else: + _import_structure["pipeline_unclip"] = ["UnCLIPPipeline"] + _import_structure["pipeline_unclip_image_variation"] = ["UnCLIPImageVariationPipeline"] + _import_structure["text_proj"] = ["UnCLIPTextProjModel"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_unclip import UnCLIPPipeline + from .pipeline_unclip_image_variation import UnCLIPImageVariationPipeline + from .text_proj import UnCLIPTextProjModel + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/unclip/pipeline_unclip.py b/diffusers/pipelines/deprecated/unclip/pipeline_unclip.py new file mode 100644 index 0000000000000000000000000000000000000000..cf3697c6354f494c1b1ebc7f67c53b69a2fadbab --- /dev/null +++ b/diffusers/pipelines/deprecated/unclip/pipeline_unclip.py @@ -0,0 +1,503 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect + +import torch +from torch.nn import functional as F +from transformers import CLIPTextModelWithProjection, CLIPTokenizer +from transformers.models.clip.modeling_clip import CLIPTextModelOutput + +from ....models import PriorTransformer, UNet2DConditionModel, UNet2DModel +from ....schedulers import UnCLIPScheduler +from ....utils import is_torch_xla_available, logging +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput +from .text_proj import UnCLIPTextProjModel + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class UnCLIPPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline for text-to-image generation using unCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`~transformers.CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + text_proj ([`UnCLIPTextProjModel`]): + Utility class to prepare and combine the embeddings before they are passed to the decoder. + decoder ([`UNet2DConditionModel`]): + The decoder to invert the image embedding into an image. + super_res_first ([`UNet2DModel`]): + Super resolution UNet. Used in all but the last step of the super resolution diffusion process. + super_res_last ([`UNet2DModel`]): + Super resolution UNet. Used in the last step of the super resolution diffusion process. + prior_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the prior denoising process (a modified [`DDPMScheduler`]). + decoder_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the decoder denoising process (a modified [`DDPMScheduler`]). + super_res_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the super resolution denoising process (a modified [`DDPMScheduler`]). + + """ + + _last_supported_version = "0.33.1" + _exclude_from_cpu_offload = ["prior"] + + prior: PriorTransformer + decoder: UNet2DConditionModel + text_proj: UnCLIPTextProjModel + text_encoder: CLIPTextModelWithProjection + tokenizer: CLIPTokenizer + super_res_first: UNet2DModel + super_res_last: UNet2DModel + + prior_scheduler: UnCLIPScheduler + decoder_scheduler: UnCLIPScheduler + super_res_scheduler: UnCLIPScheduler + + model_cpu_offload_seq = "text_encoder->text_proj->decoder->super_res_first->super_res_last" + + def __init__( + self, + prior: PriorTransformer, + decoder: UNet2DConditionModel, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_proj: UnCLIPTextProjModel, + super_res_first: UNet2DModel, + super_res_last: UNet2DModel, + prior_scheduler: UnCLIPScheduler, + decoder_scheduler: UnCLIPScheduler, + super_res_scheduler: UnCLIPScheduler, + ): + super().__init__() + + self.register_modules( + prior=prior, + decoder=decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_proj=text_proj, + super_res_first=super_res_first, + super_res_last=super_res_last, + prior_scheduler=prior_scheduler, + decoder_scheduler=decoder_scheduler, + super_res_scheduler=super_res_scheduler, + ) + + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + text_model_output: CLIPTextModelOutput | tuple | None = None, + text_attention_mask: torch.Tensor | None = None, + ): + if text_model_output is None: + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + prompt_embeds = text_encoder_output.text_embeds + text_enc_hid_states = text_encoder_output.last_hidden_state + + else: + batch_size = text_model_output[0].shape[0] + prompt_embeds, text_enc_hid_states = text_model_output[0], text_model_output[1] + text_mask = text_attention_mask + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_enc_hid_states = text_enc_hid_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds + uncond_text_enc_hid_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_enc_hid_states.shape[1] + uncond_text_enc_hid_states = uncond_text_enc_hid_states.repeat(1, num_images_per_prompt, 1) + uncond_text_enc_hid_states = uncond_text_enc_hid_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_enc_hid_states = torch.cat([uncond_text_enc_hid_states, text_enc_hid_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_enc_hid_states, text_mask + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + prior_num_inference_steps: int = 25, + decoder_num_inference_steps: int = 25, + super_res_num_inference_steps: int = 7, + generator: torch.Generator | list[torch.Generator] | None = None, + prior_latents: torch.Tensor | None = None, + decoder_latents: torch.Tensor | None = None, + super_res_latents: torch.Tensor | None = None, + text_model_output: CLIPTextModelOutput | tuple | None = None, + text_attention_mask: torch.Tensor | None = None, + prior_guidance_scale: float = 4.0, + decoder_guidance_scale: float = 8.0, + output_type: str | None = "pil", + return_dict: bool = True, + ): + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. This can only be left undefined if `text_model_output` + and `text_attention_mask` is passed. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + prior_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the prior. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + decoder_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + super_res_num_inference_steps (`int`, *optional*, defaults to 7): + The number of denoising steps for super resolution. More denoising steps usually lead to a higher + quality image at the expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prior_latents (`torch.Tensor` of shape (batch size, embeddings dimension), *optional*): + Pre-generated noisy latents to be used as inputs for the prior. + decoder_latents (`torch.Tensor` of shape (batch size, channels, height, width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + super_res_latents (`torch.Tensor` of shape (batch size, channels, super res height, super res width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + decoder_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + text_model_output (`CLIPTextModelOutput`, *optional*): + Pre-defined [`CLIPTextModel`] outputs that can be derived from the text encoder. Pre-defined text + outputs can be passed for tasks like text embedding interpolations. Make sure to also pass + `text_attention_mask` in this case. `prompt` can the be left `None`. + text_attention_mask (`torch.Tensor`, *optional*): + Pre-defined CLIP text attention mask that can be derived from the tokenizer. Pre-defined text attention + masks are necessary when passing `text_model_output`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + if prompt is not None: + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + else: + batch_size = text_model_output[0].shape[0] + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = prior_guidance_scale > 1.0 or decoder_guidance_scale > 1.0 + + prompt_embeds, text_enc_hid_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, text_model_output, text_attention_mask + ) + + # prior + + self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device) + prior_timesteps_tensor = self.prior_scheduler.timesteps + + embedding_dim = self.prior.config.embedding_dim + + prior_latents = self.prepare_latents( + (batch_size, embedding_dim), + prompt_embeds.dtype, + device, + generator, + prior_latents, + self.prior_scheduler, + ) + + for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([prior_latents] * 2) if do_classifier_free_guidance else prior_latents + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=prompt_embeds, + encoder_hidden_states=text_enc_hid_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + if i + 1 == prior_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = prior_timesteps_tensor[i + 1] + + prior_latents = self.prior_scheduler.step( + predicted_image_embedding, + timestep=t, + sample=prior_latents, + generator=generator, + prev_timestep=prev_timestep, + ).prev_sample + + prior_latents = self.prior.post_process_latents(prior_latents) + + image_embeddings = prior_latents + + # done prior + + # decoder + + text_enc_hid_states, additive_clip_time_embeddings = self.text_proj( + image_embeddings=image_embeddings, + prompt_embeds=prompt_embeds, + text_encoder_hidden_states=text_enc_hid_states, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + + if device.type == "mps": + # HACK: MPS: There is a panic when padding bool tensors, + # so cast to int tensor for the pad and back to bool afterwards + text_mask = text_mask.type(torch.int) + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1) + decoder_text_mask = decoder_text_mask.type(torch.bool) + else: + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=True) + + self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device) + decoder_timesteps_tensor = self.decoder_scheduler.timesteps + + num_channels_latents = self.decoder.config.in_channels + height = self.decoder.config.sample_size + width = self.decoder.config.sample_size + + decoder_latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + text_enc_hid_states.dtype, + device, + generator, + decoder_latents, + self.decoder_scheduler, + ) + + for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents + + noise_pred = self.decoder( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_enc_hid_states, + class_labels=additive_clip_time_embeddings, + attention_mask=decoder_text_mask, + ).sample + + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1) + noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1) + noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, predicted_variance], dim=1) + + if i + 1 == decoder_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = decoder_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + decoder_latents = self.decoder_scheduler.step( + noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + decoder_latents = decoder_latents.clamp(-1, 1) + + image_small = decoder_latents + + # done decoder + + # super res + + self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device) + super_res_timesteps_tensor = self.super_res_scheduler.timesteps + + channels = self.super_res_first.config.in_channels // 2 + height = self.super_res_first.config.sample_size + width = self.super_res_first.config.sample_size + + super_res_latents = self.prepare_latents( + (batch_size, channels, height, width), + image_small.dtype, + device, + generator, + super_res_latents, + self.super_res_scheduler, + ) + + if device.type == "mps": + # MPS does not support many interpolations + image_upscaled = F.interpolate(image_small, size=[height, width]) + else: + interpolate_antialias = {} + if "antialias" in inspect.signature(F.interpolate).parameters: + interpolate_antialias["antialias"] = True + + image_upscaled = F.interpolate( + image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias + ) + + for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)): + # no classifier free guidance + + if i == super_res_timesteps_tensor.shape[0] - 1: + unet = self.super_res_last + else: + unet = self.super_res_first + + latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1) + + noise_pred = unet( + sample=latent_model_input, + timestep=t, + ).sample + + if i + 1 == super_res_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = super_res_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + super_res_latents = self.super_res_scheduler.step( + noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + image = super_res_latents + # done super res + + self.maybe_free_model_hooks() + + # post processing + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/deprecated/unclip/pipeline_unclip_image_variation.py b/diffusers/pipelines/deprecated/unclip/pipeline_unclip_image_variation.py new file mode 100644 index 0000000000000000000000000000000000000000..3ea5ca75949c10db7259bfbd1d13ff0e5ae65948 --- /dev/null +++ b/diffusers/pipelines/deprecated/unclip/pipeline_unclip_image_variation.py @@ -0,0 +1,430 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect + +import PIL.Image +import torch +from torch.nn import functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ....models import UNet2DConditionModel, UNet2DModel +from ....schedulers import UnCLIPScheduler +from ....utils import is_torch_xla_available, logging +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput +from .text_proj import UnCLIPTextProjModel + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class UnCLIPImageVariationPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline to generate image variations from an input image using UnCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`~transformers.CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `image_encoder`. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + text_proj ([`UnCLIPTextProjModel`]): + Utility class to prepare and combine the embeddings before they are passed to the decoder. + decoder ([`UNet2DConditionModel`]): + The decoder to invert the image embedding into an image. + super_res_first ([`UNet2DModel`]): + Super resolution UNet. Used in all but the last step of the super resolution diffusion process. + super_res_last ([`UNet2DModel`]): + Super resolution UNet. Used in the last step of the super resolution diffusion process. + decoder_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the decoder denoising process (a modified [`DDPMScheduler`]). + super_res_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the super resolution denoising process (a modified [`DDPMScheduler`]). + """ + + _last_supported_version = "0.33.1" + decoder: UNet2DConditionModel + text_proj: UnCLIPTextProjModel + text_encoder: CLIPTextModelWithProjection + tokenizer: CLIPTokenizer + feature_extractor: CLIPImageProcessor + image_encoder: CLIPVisionModelWithProjection + super_res_first: UNet2DModel + super_res_last: UNet2DModel + + decoder_scheduler: UnCLIPScheduler + super_res_scheduler: UnCLIPScheduler + model_cpu_offload_seq = "text_encoder->image_encoder->text_proj->decoder->super_res_first->super_res_last" + + def __init__( + self, + decoder: UNet2DConditionModel, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_proj: UnCLIPTextProjModel, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection, + super_res_first: UNet2DModel, + super_res_last: UNet2DModel, + decoder_scheduler: UnCLIPScheduler, + super_res_scheduler: UnCLIPScheduler, + ): + super().__init__() + + self.register_modules( + decoder=decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_proj=text_proj, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + super_res_first=super_res_first, + super_res_last=super_res_last, + decoder_scheduler=decoder_scheduler, + super_res_scheduler=super_res_scheduler, + ) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + prompt_embeds = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds + uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + def _encode_image(self, image, device, num_images_per_prompt, image_embeddings: torch.Tensor | None = None): + dtype = next(self.image_encoder.parameters()).dtype + + if image_embeddings is None: + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(images=image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeddings = self.image_encoder(image).image_embeds + + image_embeddings = image_embeddings.repeat_interleave(num_images_per_prompt, dim=0) + + return image_embeddings + + @torch.no_grad() + def __call__( + self, + image: PIL.Image.Image | list[PIL.Image.Image] | torch.Tensor | None = None, + num_images_per_prompt: int = 1, + decoder_num_inference_steps: int = 25, + super_res_num_inference_steps: int = 7, + generator: torch.Generator | None = None, + decoder_latents: torch.Tensor | None = None, + super_res_latents: torch.Tensor | None = None, + image_embeddings: torch.Tensor | None = None, + decoder_guidance_scale: float = 8.0, + output_type: str | None = "pil", + return_dict: bool = True, + ): + """ + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.Tensor`): + `Image` or tensor representing an image batch to be used as the starting point. If you provide a + tensor, it needs to be compatible with the [`CLIPImageProcessor`] + [configuration](https://huggingface.co/fusing/karlo-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json). + Can be left as `None` only when `image_embeddings` are passed. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + decoder_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + super_res_num_inference_steps (`int`, *optional*, defaults to 7): + The number of denoising steps for super resolution. More denoising steps usually lead to a higher + quality image at the expense of slower inference. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + decoder_latents (`torch.Tensor` of shape (batch size, channels, height, width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + super_res_latents (`torch.Tensor` of shape (batch size, channels, super res height, super res width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + decoder_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + image_embeddings (`torch.Tensor`, *optional*): + Pre-defined image embeddings that can be derived from the image encoder. Pre-defined image embeddings + can be passed for tasks like image interpolations. `image` can be left as `None`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + if image is not None: + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + else: + batch_size = image_embeddings.shape[0] + + prompt = [""] * batch_size + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = decoder_guidance_scale > 1.0 + + prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance + ) + + image_embeddings = self._encode_image(image, device, num_images_per_prompt, image_embeddings) + + # decoder + text_encoder_hidden_states, additive_clip_time_embeddings = self.text_proj( + image_embeddings=image_embeddings, + prompt_embeds=prompt_embeds, + text_encoder_hidden_states=text_encoder_hidden_states, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + + if device.type == "mps": + # HACK: MPS: There is a panic when padding bool tensors, + # so cast to int tensor for the pad and back to bool afterwards + text_mask = text_mask.type(torch.int) + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1) + decoder_text_mask = decoder_text_mask.type(torch.bool) + else: + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=True) + + self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device) + decoder_timesteps_tensor = self.decoder_scheduler.timesteps + + num_channels_latents = self.decoder.config.in_channels + height = self.decoder.config.sample_size + width = self.decoder.config.sample_size + + if decoder_latents is None: + decoder_latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + text_encoder_hidden_states.dtype, + device, + generator, + decoder_latents, + self.decoder_scheduler, + ) + + for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents + + noise_pred = self.decoder( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_encoder_hidden_states, + class_labels=additive_clip_time_embeddings, + attention_mask=decoder_text_mask, + ).sample + + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1) + noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1) + noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, predicted_variance], dim=1) + + if i + 1 == decoder_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = decoder_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + decoder_latents = self.decoder_scheduler.step( + noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + decoder_latents = decoder_latents.clamp(-1, 1) + + image_small = decoder_latents + + # done decoder + + # super res + + self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device) + super_res_timesteps_tensor = self.super_res_scheduler.timesteps + + channels = self.super_res_first.config.in_channels // 2 + height = self.super_res_first.config.sample_size + width = self.super_res_first.config.sample_size + + if super_res_latents is None: + super_res_latents = self.prepare_latents( + (batch_size, channels, height, width), + image_small.dtype, + device, + generator, + super_res_latents, + self.super_res_scheduler, + ) + + if device.type == "mps": + # MPS does not support many interpolations + image_upscaled = F.interpolate(image_small, size=[height, width]) + else: + interpolate_antialias = {} + if "antialias" in inspect.signature(F.interpolate).parameters: + interpolate_antialias["antialias"] = True + + image_upscaled = F.interpolate( + image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias + ) + + for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)): + # no classifier free guidance + + if i == super_res_timesteps_tensor.shape[0] - 1: + unet = self.super_res_last + else: + unet = self.super_res_first + + latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1) + + noise_pred = unet( + sample=latent_model_input, + timestep=t, + ).sample + + if i + 1 == super_res_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = super_res_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + super_res_latents = self.super_res_scheduler.step( + noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + image = super_res_latents + + # done super res + self.maybe_free_model_hooks() + + # post processing + + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/deprecated/unclip/text_proj.py b/diffusers/pipelines/deprecated/unclip/text_proj.py new file mode 100644 index 0000000000000000000000000000000000000000..5493df794acc0a93ef0dba2ae03e8a114a76d4fd --- /dev/null +++ b/diffusers/pipelines/deprecated/unclip/text_proj.py @@ -0,0 +1,86 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +from torch import nn + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models import ModelMixin + + +class UnCLIPTextProjModel(ModelMixin, ConfigMixin): + """ + Utility class for CLIP embeddings. Used to combine the image and text embeddings into a format usable by the + decoder. + + For more details, see the original paper: https://huggingface.co/papers/2204.06125 section 2.1 + """ + + @register_to_config + def __init__( + self, + *, + clip_extra_context_tokens: int = 4, + clip_embeddings_dim: int = 768, + time_embed_dim: int, + cross_attention_dim, + ): + super().__init__() + + self.learned_classifier_free_guidance_embeddings = nn.Parameter(torch.zeros(clip_embeddings_dim)) + + # parameters for additional clip time embeddings + self.embedding_proj = nn.Linear(clip_embeddings_dim, time_embed_dim) + self.clip_image_embeddings_project_to_time_embeddings = nn.Linear(clip_embeddings_dim, time_embed_dim) + + # parameters for encoder hidden states + self.clip_extra_context_tokens = clip_extra_context_tokens + self.clip_extra_context_tokens_proj = nn.Linear( + clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim + ) + self.encoder_hidden_states_proj = nn.Linear(clip_embeddings_dim, cross_attention_dim) + self.text_encoder_hidden_states_norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, *, image_embeddings, prompt_embeds, text_encoder_hidden_states, do_classifier_free_guidance): + if do_classifier_free_guidance: + # Add the classifier free guidance embeddings to the image embeddings + image_embeddings_batch_size = image_embeddings.shape[0] + classifier_free_guidance_embeddings = self.learned_classifier_free_guidance_embeddings.unsqueeze(0) + classifier_free_guidance_embeddings = classifier_free_guidance_embeddings.expand( + image_embeddings_batch_size, -1 + ) + image_embeddings = torch.cat([classifier_free_guidance_embeddings, image_embeddings], dim=0) + + # The image embeddings batch size and the text embeddings batch size are equal + assert image_embeddings.shape[0] == prompt_embeds.shape[0] + + batch_size = prompt_embeds.shape[0] + + # "Specifically, we modify the architecture described in Nichol et al. (2021) by projecting and + # adding CLIP embeddings to the existing timestep embedding, ... + time_projected_prompt_embeds = self.embedding_proj(prompt_embeds) + time_projected_image_embeddings = self.clip_image_embeddings_project_to_time_embeddings(image_embeddings) + additive_clip_time_embeddings = time_projected_image_embeddings + time_projected_prompt_embeds + + # ... and by projecting CLIP embeddings into four + # extra tokens of context that are concatenated to the sequence of outputs from the GLIDE text encoder" + clip_extra_context_tokens = self.clip_extra_context_tokens_proj(image_embeddings) + clip_extra_context_tokens = clip_extra_context_tokens.reshape(batch_size, -1, self.clip_extra_context_tokens) + clip_extra_context_tokens = clip_extra_context_tokens.permute(0, 2, 1) + + text_encoder_hidden_states = self.encoder_hidden_states_proj(text_encoder_hidden_states) + text_encoder_hidden_states = self.text_encoder_hidden_states_norm(text_encoder_hidden_states) + text_encoder_hidden_states = torch.cat([clip_extra_context_tokens, text_encoder_hidden_states], dim=1) + + return text_encoder_hidden_states, additive_clip_time_embeddings diff --git a/diffusers/pipelines/deprecated/unidiffuser/__init__.py b/diffusers/pipelines/deprecated/unidiffuser/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..aeaba167dc11204e2afe44e7906c0827b56f53f2 --- /dev/null +++ b/diffusers/pipelines/deprecated/unidiffuser/__init__.py @@ -0,0 +1,58 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import ( + ImageTextPipelineOutput, + UniDiffuserPipeline, + ) + + _dummy_objects.update( + {"ImageTextPipelineOutput": ImageTextPipelineOutput, "UniDiffuserPipeline": UniDiffuserPipeline} + ) +else: + _import_structure["modeling_text_decoder"] = ["UniDiffuserTextDecoder"] + _import_structure["modeling_uvit"] = ["UniDiffuserModel", "UTransformer2DModel"] + _import_structure["pipeline_unidiffuser"] = ["ImageTextPipelineOutput", "UniDiffuserPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import ( + ImageTextPipelineOutput, + UniDiffuserPipeline, + ) + else: + from .modeling_text_decoder import UniDiffuserTextDecoder + from .modeling_uvit import UniDiffuserModel, UTransformer2DModel + from .pipeline_unidiffuser import ImageTextPipelineOutput, UniDiffuserPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/unidiffuser/modeling_text_decoder.py b/diffusers/pipelines/deprecated/unidiffuser/modeling_text_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..a068f99c636848cdf541abf8e6b3357601caf124 --- /dev/null +++ b/diffusers/pipelines/deprecated/unidiffuser/modeling_text_decoder.py @@ -0,0 +1,294 @@ +import numpy as np +import torch +from torch import nn +from transformers import GPT2Config, GPT2LMHeadModel +from transformers.modeling_utils import ModuleUtilsMixin + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models import ModelMixin + + +# Modified from ClipCaptionModel in https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py +class UniDiffuserTextDecoder(ModelMixin, ConfigMixin, ModuleUtilsMixin): + """ + Text decoder model for a image-text [UniDiffuser](https://huggingface.co/papers/2303.06555) model. This is used to + generate text from the UniDiffuser image-text embedding. + + Parameters: + prefix_length (`int`): + Max number of prefix tokens that will be supplied to the model. + prefix_inner_dim (`int`): + The hidden size of the incoming prefix embeddings. For UniDiffuser, this would be the hidden dim of the + CLIP text encoder. + prefix_hidden_dim (`int`, *optional*): + Hidden dim of the MLP if we encode the prefix. + vocab_size (`int`, *optional*, defaults to 50257): + Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`GPT2Model`] or [`TFGPT2Model`]. + n_positions (`int`, *optional*, defaults to 1024): + The maximum sequence length that this model might ever be used with. Typically set this to something large + just in case (e.g., 512 or 1024 or 2048). + n_embd (`int`, *optional*, defaults to 768): + Dimensionality of the embeddings and hidden states. + n_layer (`int`, *optional*, defaults to 12): + Number of hidden layers in the Transformer encoder. + n_head (`int`, *optional*, defaults to 12): + Number of attention heads for each attention layer in the Transformer encoder. + n_inner (`int`, *optional*, defaults to None): + Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd + activation_function (`str`, *optional*, defaults to `"gelu"`): + Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`. + resid_pdrop (`float`, *optional*, defaults to 0.1): + The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. + embd_pdrop (`float`, *optional*, defaults to 0.1): + The dropout ratio for the embeddings. + attn_pdrop (`float`, *optional*, defaults to 0.1): + The dropout ratio for the attention. + layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): + The epsilon to use in the layer normalization layers. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + scale_attn_weights (`bool`, *optional*, defaults to `True`): + Scale attention weights by dividing by sqrt(hidden_size).. + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). + scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`): + Whether to additionally scale attention weights by `1 / layer_idx + 1`. + reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`): + Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention + dot-product/softmax to float() when training with mixed precision. + """ + + _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"] + + @register_to_config + def __init__( + self, + prefix_length: int, + prefix_inner_dim: int, + prefix_hidden_dim: int | None = None, + vocab_size: int = 50257, # Start of GPT2 config args + n_positions: int = 1024, + n_embd: int = 768, + n_layer: int = 12, + n_head: int = 12, + n_inner: int | None = None, + activation_function: str = "gelu_new", + resid_pdrop: float = 0.1, + embd_pdrop: float = 0.1, + attn_pdrop: float = 0.1, + layer_norm_epsilon: float = 1e-5, + initializer_range: float = 0.02, + scale_attn_weights: bool = True, + use_cache: bool = True, + scale_attn_by_inverse_layer_idx: bool = False, + reorder_and_upcast_attn: bool = False, + ): + super().__init__() + + self.prefix_length = prefix_length + + if prefix_inner_dim != n_embd and prefix_hidden_dim is None: + raise ValueError( + f"`prefix_hidden_dim` cannot be `None` when `prefix_inner_dim`: {prefix_hidden_dim} and" + f" `n_embd`: {n_embd} are not equal." + ) + + self.prefix_inner_dim = prefix_inner_dim + self.prefix_hidden_dim = prefix_hidden_dim + + self.encode_prefix = ( + nn.Linear(self.prefix_inner_dim, self.prefix_hidden_dim) + if self.prefix_hidden_dim is not None + else nn.Identity() + ) + self.decode_prefix = ( + nn.Linear(self.prefix_hidden_dim, n_embd) if self.prefix_hidden_dim is not None else nn.Identity() + ) + + gpt_config = GPT2Config( + vocab_size=vocab_size, + n_positions=n_positions, + n_embd=n_embd, + n_layer=n_layer, + n_head=n_head, + n_inner=n_inner, + activation_function=activation_function, + resid_pdrop=resid_pdrop, + embd_pdrop=embd_pdrop, + attn_pdrop=attn_pdrop, + layer_norm_epsilon=layer_norm_epsilon, + initializer_range=initializer_range, + scale_attn_weights=scale_attn_weights, + use_cache=use_cache, + scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx, + reorder_and_upcast_attn=reorder_and_upcast_attn, + ) + self.transformer = GPT2LMHeadModel(gpt_config) + + def forward( + self, + input_ids: torch.Tensor, + prefix_embeds: torch.Tensor, + attention_mask: torch.Tensor | None = None, + labels: torch.Tensor | None = None, + ): + """ + Args: + input_ids (`torch.Tensor` of shape `(N, max_seq_len)`): + Text tokens to use for inference. + prefix_embeds (`torch.Tensor` of shape `(N, prefix_length, 768)`): + Prefix embedding to prepend to the embedded tokens. + attention_mask (`torch.Tensor` of shape `(N, prefix_length + max_seq_len, 768)`, *optional*): + Attention mask for the prefix embedding. + labels (`torch.Tensor`, *optional*): + Labels to use for language modeling. + """ + embedding_text = self.transformer.transformer.wte(input_ids) + hidden = self.encode_prefix(prefix_embeds) + prefix_embeds = self.decode_prefix(hidden) + embedding_cat = torch.cat((prefix_embeds, embedding_text), dim=1) + + if labels is not None: + dummy_token = self.get_dummy_token(input_ids.shape[0], input_ids.device) + labels = torch.cat((dummy_token, input_ids), dim=1) + out = self.transformer(inputs_embeds=embedding_cat, labels=labels, attention_mask=attention_mask) + if self.prefix_hidden_dim is not None: + return out, hidden + else: + return out + + def get_dummy_token(self, batch_size: int, device: torch.device) -> torch.Tensor: + return torch.zeros(batch_size, self.prefix_length, dtype=torch.int64, device=device) + + def encode(self, prefix): + return self.encode_prefix(prefix) + + @torch.no_grad() + def generate_captions(self, features, eos_token_id, device): + """ + Generate captions given text embedding features. Returns list[L]. + + Args: + features (`torch.Tensor` of shape `(B, L, D)`): + Text embedding features to generate captions from. + eos_token_id (`int`): + The token ID of the EOS token for the text decoder model. + device: + Device to perform text generation on. + + Returns: + `list[str]`: A list of strings generated from the decoder model. + """ + + features = torch.split(features, 1, dim=0) + generated_tokens = [] + generated_seq_lengths = [] + for feature in features: + feature = self.decode_prefix(feature.to(device)) # back to the clip feature + # Only support beam search for now + output_tokens, seq_lengths = self.generate_beam( + input_embeds=feature, device=device, eos_token_id=eos_token_id + ) + generated_tokens.append(output_tokens[0]) + generated_seq_lengths.append(seq_lengths[0]) + generated_tokens = torch.stack(generated_tokens) + generated_seq_lengths = torch.stack(generated_seq_lengths) + return generated_tokens, generated_seq_lengths + + @torch.no_grad() + def generate_beam( + self, + input_ids=None, + input_embeds=None, + device=None, + beam_size: int = 5, + entry_length: int = 67, + temperature: float = 1.0, + eos_token_id: int | None = None, + ): + """ + Generates text using the given tokenizer and text prompt or token embedding via beam search. This + implementation is based on the beam search implementation from the [original UniDiffuser + code](https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py#L89). + + Args: + eos_token_id (`int`, *optional*): + The token ID of the EOS token for the text decoder model. + input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*): + Tokenizer indices of input sequence tokens in the vocabulary. One of `input_ids` and `input_embeds` + must be supplied. + input_embeds (`torch.Tensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*): + An embedded representation to directly pass to the transformer as a prefix for beam search. One of + `input_ids` and `input_embeds` must be supplied. + device: + The device to perform beam search on. + beam_size (`int`, *optional*, defaults to `5`): + The number of best states to store during beam search. + entry_length (`int`, *optional*, defaults to `67`): + The number of iterations to run beam search. + temperature (`float`, *optional*, defaults to 1.0): + The temperature to use when performing the softmax over logits from the decoding model. + + Returns: + `tuple(torch.Tensor, torch.Tensor)`: A tuple of tensors where the first element is a tensor of generated + token sequences sorted by score in descending order, and the second element is the sequence lengths + corresponding to those sequences. + """ + # Generates text until stop_token is reached using beam search with the desired beam size. + stop_token_index = eos_token_id + tokens = None + scores = None + seq_lengths = torch.ones(beam_size, device=device, dtype=torch.int) + is_stopped = torch.zeros(beam_size, device=device, dtype=torch.bool) + + if input_embeds is not None: + generated = input_embeds + else: + generated = self.transformer.transformer.wte(input_ids) + + for i in range(entry_length): + outputs = self.transformer(inputs_embeds=generated) + logits = outputs.logits + logits = logits[:, -1, :] / (temperature if temperature > 0 else 1.0) + logits = logits.softmax(-1).log() + + if scores is None: + scores, next_tokens = logits.topk(beam_size, -1) + generated = generated.expand(beam_size, *generated.shape[1:]) + next_tokens, scores = next_tokens.permute(1, 0), scores.squeeze(0) + if tokens is None: + tokens = next_tokens + else: + tokens = tokens.expand(beam_size, *tokens.shape[1:]) + tokens = torch.cat((tokens, next_tokens), dim=1) + else: + logits[is_stopped] = -float(np.inf) + logits[is_stopped, 0] = 0 + scores_sum = scores[:, None] + logits + seq_lengths[~is_stopped] += 1 + scores_sum_average = scores_sum / seq_lengths[:, None] + scores_sum_average, next_tokens = scores_sum_average.view(-1).topk(beam_size, -1) + next_tokens_source = next_tokens // scores_sum.shape[1] + seq_lengths = seq_lengths[next_tokens_source] + next_tokens = next_tokens % scores_sum.shape[1] + next_tokens = next_tokens.unsqueeze(1) + tokens = tokens[next_tokens_source] + tokens = torch.cat((tokens, next_tokens), dim=1) + generated = generated[next_tokens_source] + scores = scores_sum_average * seq_lengths + is_stopped = is_stopped[next_tokens_source] + + next_token_embed = self.transformer.transformer.wte(next_tokens.squeeze()).view(generated.shape[0], 1, -1) + generated = torch.cat((generated, next_token_embed), dim=1) + is_stopped = is_stopped + next_tokens.eq(stop_token_index).squeeze() + if is_stopped.all(): + break + + scores = scores / seq_lengths + order = scores.argsort(descending=True) + # tokens tensors are already padded to max_seq_length + output_texts = [tokens[i] for i in order] + output_texts = torch.stack(output_texts, dim=0) + seq_lengths = torch.tensor([seq_lengths[i] for i in order], dtype=seq_lengths.dtype) + return output_texts, seq_lengths diff --git a/diffusers/pipelines/deprecated/unidiffuser/modeling_uvit.py b/diffusers/pipelines/deprecated/unidiffuser/modeling_uvit.py new file mode 100644 index 0000000000000000000000000000000000000000..6fd4ff50285fcd3dbc3a78adc16fc498f59df3ea --- /dev/null +++ b/diffusers/pipelines/deprecated/unidiffuser/modeling_uvit.py @@ -0,0 +1,1196 @@ +import math + +import torch +from torch import nn + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models import ModelMixin +from ....models.attention import FeedForward +from ....models.attention_processor import Attention +from ....models.embeddings import TimestepEmbedding, Timesteps, get_2d_sincos_pos_embed +from ....models.modeling_outputs import Transformer2DModelOutput +from ....models.normalization import AdaLayerNorm +from ....utils import logging + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def _no_grad_trunc_normal_(tensor, mean, std, a, b): + # Cut & paste from PyTorch official master until it's in a few official releases - RW + # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf + def norm_cdf(x): + # Computes standard normal cumulative distribution function + return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0 + + if (mean < a - 2 * std) or (mean > b + 2 * std): + logger.warning( + "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. " + "The distribution of values may be incorrect." + ) + + with torch.no_grad(): + # Values are generated by using a truncated uniform distribution and + # then using the inverse CDF for the normal distribution. + # Get upper and lower cdf values + l = norm_cdf((a - mean) / std) + u = norm_cdf((b - mean) / std) + + # Uniformly fill tensor with values from [l, u], then translate to + # [2l-1, 2u-1]. + tensor.uniform_(2 * l - 1, 2 * u - 1) + + # Use inverse cdf transform for normal distribution to get truncated + # standard normal + tensor.erfinv_() + + # Transform to proper mean, std + tensor.mul_(std * math.sqrt(2.0)) + tensor.add_(mean) + + # Clamp to ensure it's in the proper range + tensor.clamp_(min=a, max=b) + return tensor + + +def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0): + # type: (torch.Tensor, float, float, float, float) -> torch.Tensor + r"""Fills the input Tensor with values drawn from a truncated + normal distribution. The values are effectively drawn from the normal distribution :math:`\mathcal{N}(\text{mean}, + \text{std}^2)` with values outside :math:`[a, b]` redrawn until they are within the bounds. The method used for + generating the random values works best when :math:`a \leq \text{mean} \leq b`. + + Args: + tensor: an n-dimensional `torch.Tensor` + mean: the mean of the normal distribution + std: the standard deviation of the normal distribution + a: the minimum cutoff value + b: the maximum cutoff value + Examples: + >>> w = torch.empty(3, 5) >>> nn.init.trunc_normal_(w) + """ + return _no_grad_trunc_normal_(tensor, mean, std, a, b) + + +class PatchEmbed(nn.Module): + """2D Image to Patch Embedding""" + + def __init__( + self, + height=224, + width=224, + patch_size=16, + in_channels=3, + embed_dim=768, + layer_norm=False, + flatten=True, + bias=True, + use_pos_embed=True, + ): + super().__init__() + + num_patches = (height // patch_size) * (width // patch_size) + self.flatten = flatten + self.layer_norm = layer_norm + + self.proj = nn.Conv2d( + in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias + ) + if layer_norm: + self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6) + else: + self.norm = None + + self.use_pos_embed = use_pos_embed + if self.use_pos_embed: + pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5), output_type="pt") + self.register_buffer("pos_embed", pos_embed.float().unsqueeze(0), persistent=False) + + def forward(self, latent): + latent = self.proj(latent) + if self.flatten: + latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC + if self.layer_norm: + latent = self.norm(latent) + if self.use_pos_embed: + return latent + self.pos_embed + else: + return latent + + +class SkipBlock(nn.Module): + def __init__(self, dim: int): + super().__init__() + + self.skip_linear = nn.Linear(2 * dim, dim) + + # Use torch.nn.LayerNorm for now, following the original code + self.norm = nn.LayerNorm(dim) + + def forward(self, x, skip): + x = self.skip_linear(torch.cat([x, skip], dim=-1)) + x = self.norm(x) + + return x + + +# Modified to support both pre-LayerNorm and post-LayerNorm configurations +# Don't support AdaLayerNormZero for now +# Modified from diffusers.models.attention.BasicTransformerBlock +class UTransformerBlock(nn.Module): + r""" + A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): + Activation function to be used in feed-forward. + num_embeds_ada_norm (:obj: `int`, *optional*): + The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (:obj: `bool`, *optional*, defaults to `False`): + Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float32 when performing the attention calculation. + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + norm_type (`str`, defaults to `"layer_norm"`): + The layer norm implementation to use. + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). Note that `BasicTransformerBlock` uses pre-LayerNorm, e.g. + `pre_layer_norm = True`. + final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: int | None = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: int | None = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", + pre_layer_norm: bool = True, + final_dropout: bool = False, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + + self.pre_layer_norm = pre_layer_norm + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + # 1. Self-Attn + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + ) # is self-attn if encoder_hidden_states is none + else: + self.attn2 = None + + if self.use_ada_layer_norm: + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = ( + AdaLayerNorm(dim, num_embeds_ada_norm) + if self.use_ada_layer_norm + else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + ) + else: + self.norm2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout) + + def forward( + self, + hidden_states, + attention_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + timestep=None, + cross_attention_kwargs=None, + class_labels=None, + ): + # Pre-LayerNorm + if self.pre_layer_norm: + if self.use_ada_layer_norm: + norm_hidden_states = self.norm1(hidden_states, timestep) + else: + norm_hidden_states = self.norm1(hidden_states) + else: + norm_hidden_states = hidden_states + + # 1. Self-Attention + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + # Post-LayerNorm + if not self.pre_layer_norm: + if self.use_ada_layer_norm: + attn_output = self.norm1(attn_output, timestep) + else: + attn_output = self.norm1(attn_output) + + hidden_states = attn_output + hidden_states + + if self.attn2 is not None: + # Pre-LayerNorm + if self.pre_layer_norm: + norm_hidden_states = ( + self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states) + ) + else: + norm_hidden_states = hidden_states + # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly + # prepare attention mask here + + # 2. Cross-Attention + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + + # Post-LayerNorm + if not self.pre_layer_norm: + attn_output = self.norm2(attn_output, timestep) if self.use_ada_layer_norm else self.norm2(attn_output) + + hidden_states = attn_output + hidden_states + + # 3. Feed-forward + # Pre-LayerNorm + if self.pre_layer_norm: + norm_hidden_states = self.norm3(hidden_states) + else: + norm_hidden_states = hidden_states + + ff_output = self.ff(norm_hidden_states) + + # Post-LayerNorm + if not self.pre_layer_norm: + ff_output = self.norm3(ff_output) + + hidden_states = ff_output + hidden_states + + return hidden_states + + +# Like UTransformerBlock except with LayerNorms on the residual backbone of the block +# Modified from diffusers.models.attention.BasicTransformerBlock +class UniDiffuserBlock(nn.Module): + r""" + A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations and puts the + LayerNorms on the residual backbone of the block. This matches the transformer block in the [original UniDiffuser + implementation](https://github.com/thu-ml/unidiffuser/blob/main/libs/uvit_multi_post_ln_v1.py#L104). + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): + Activation function to be used in feed-forward. + num_embeds_ada_norm (:obj: `int`, *optional*): + The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (:obj: `bool`, *optional*, defaults to `False`): + Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float() when performing the attention calculation. + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + norm_type (`str`, defaults to `"layer_norm"`): + The layer norm implementation to use. + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm + (`pre_layer_norm = False`). + final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: int | None = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: int | None = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", + pre_layer_norm: bool = False, + final_dropout: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + + self.pre_layer_norm = pre_layer_norm + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + # 1. Self-Attn + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + ) # is self-attn if encoder_hidden_states is none + else: + self.attn2 = None + + if self.use_ada_layer_norm: + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = ( + AdaLayerNorm(dim, num_embeds_ada_norm) + if self.use_ada_layer_norm + else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + ) + else: + self.norm2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout) + + def forward( + self, + hidden_states, + attention_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + timestep=None, + cross_attention_kwargs=None, + class_labels=None, + ): + # Following the diffusers transformer block implementation, put the LayerNorm on the + # residual backbone + # Pre-LayerNorm + if self.pre_layer_norm: + if self.use_ada_layer_norm: + hidden_states = self.norm1(hidden_states, timestep) + else: + hidden_states = self.norm1(hidden_states) + + # 1. Self-Attention + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + attn_output = self.attn1( + hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + # Following the diffusers transformer block implementation, put the LayerNorm on the + # residual backbone + # Post-LayerNorm + if not self.pre_layer_norm: + if self.use_ada_layer_norm: + hidden_states = self.norm1(hidden_states, timestep) + else: + hidden_states = self.norm1(hidden_states) + + if self.attn2 is not None: + # Pre-LayerNorm + if self.pre_layer_norm: + hidden_states = ( + self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states) + ) + # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly + # prepare attention mask here + + # 2. Cross-Attention + attn_output = self.attn2( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + # Post-LayerNorm + if not self.pre_layer_norm: + hidden_states = ( + self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states) + ) + + # 3. Feed-forward + # Pre-LayerNorm + if self.pre_layer_norm: + hidden_states = self.norm3(hidden_states) + + ff_output = self.ff(hidden_states) + + hidden_states = ff_output + hidden_states + + # Post-LayerNorm + if not self.pre_layer_norm: + hidden_states = self.norm3(hidden_states) + + return hidden_states + + +# Modified from diffusers.models.transformer_2d.Transformer2DModel +# Modify the transformer block structure to be U-Net like following U-ViT +# Only supports patch-style input and torch.nn.LayerNorm currently +# https://github.com/baofff/U-ViT +class UTransformer2DModel(ModelMixin, ConfigMixin): + """ + Transformer model based on the [U-ViT](https://github.com/baofff/U-ViT) architecture for image-like data. Compared + to [`Transformer2DModel`], this model has skip connections between transformer blocks in a "U"-shaped fashion, + similar to a U-Net. Supports only continuous (actual embeddings) inputs, which are embedded via a [`PatchEmbed`] + layer and then reshaped to (b, t, d). + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input. + out_channels (`int`, *optional*): + The number of output channels; if `None`, defaults to `in_channels`. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + norm_num_groups (`int`, *optional*, defaults to `32`): + The number of groups to use when performing Group Normalization. + cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + patch_size (`int`, *optional*, defaults to 2): + The patch size to use in the patch embedding. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + use_linear_projection (int, *optional*): TODO: Not used + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used in each + transformer block. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float() when performing the attention calculation. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`. + block_type (`str`, *optional*, defaults to `"unidiffuser"`): + The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual + backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard + behavior in `diffusers`.) + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm + (`pre_layer_norm = False`). + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + use_patch_pos_embed (`bool`, *optional*): + Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`). + final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + """ + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: int | None = None, + out_channels: int | None = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: int | None = None, + attention_bias: bool = False, + sample_size: int | None = None, + num_vector_embeds: int | None = None, + patch_size: int | None = 2, + activation_fn: str = "geglu", + num_embeds_ada_norm: int | None = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + block_type: str = "unidiffuser", + pre_layer_norm: bool = False, + norm_elementwise_affine: bool = True, + use_patch_pos_embed=False, + ff_final_dropout: bool = False, + ): + super().__init__() + self.use_linear_projection = use_linear_projection + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + + # 1. Input + # Only support patch input of shape (batch_size, num_channels, height, width) for now + assert in_channels is not None and patch_size is not None, "Patch input requires in_channels and patch_size." + + assert sample_size is not None, "UTransformer2DModel over patched input must provide sample_size" + + # 2. Define input layers + self.height = sample_size + self.width = sample_size + + self.patch_size = patch_size + self.pos_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + use_pos_embed=use_patch_pos_embed, + ) + + # 3. Define transformers blocks + # Modify this to have in_blocks ("downsample" blocks, even though we don't actually downsample), a mid_block, + # and out_blocks ("upsample" blocks). Like a U-Net, there are skip connections from in_blocks to out_blocks in + # a "U"-shaped fashion (e.g. first in_block to last out_block, etc.). + # Quick hack to make the transformer block type configurable + if block_type == "unidiffuser": + block_cls = UniDiffuserBlock + else: + block_cls = UTransformerBlock + self.transformer_in_blocks = nn.ModuleList( + [ + block_cls( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + final_dropout=ff_final_dropout, + ) + for d in range(num_layers // 2) + ] + ) + + self.transformer_mid_block = block_cls( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + final_dropout=ff_final_dropout, + ) + + # For each skip connection, we use a SkipBlock (concatenation + Linear + LayerNorm) to process the inputs + # before each transformer out_block. + self.transformer_out_blocks = nn.ModuleList( + [ + nn.ModuleDict( + { + "skip": SkipBlock( + inner_dim, + ), + "block": block_cls( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + final_dropout=ff_final_dropout, + ), + } + ) + for d in range(num_layers // 2) + ] + ) + + # 4. Define output layers + self.out_channels = in_channels if out_channels is None else out_channels + + # Following the UniDiffuser U-ViT implementation, we process the transformer output with + # a LayerNorm layer with per-element affine params + self.norm_out = nn.LayerNorm(inner_dim) + + def forward( + self, + hidden_states, + encoder_hidden_states=None, + timestep=None, + class_labels=None, + cross_attention_kwargs=None, + return_dict: bool = True, + hidden_states_is_embedding: bool = False, + unpatchify: bool = True, + ): + """ + Args: + hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`. + When continuous, `torch.Tensor` of shape `(batch size, channel, height, width)`): Input hidden_states + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.long`, *optional*): + Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Optional class labels to be applied as an embedding in AdaLayerZeroNorm. Used to indicate class labels + conditioning. + cross_attention_kwargs (*optional*): + Keyword arguments to supply to the cross attention layers, if used. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + hidden_states_is_embedding (`bool`, *optional*, defaults to `False`): + Whether or not hidden_states is an embedding directly usable by the transformer. In this case we will + ignore input handling (e.g. continuous, vectorized, etc.) and directly feed hidden_states into the + transformer blocks. + unpatchify (`bool`, *optional*, defaults to `True`): + Whether to unpatchify the transformer output. + + Returns: + [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`: + [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + # 0. Check inputs + + if not unpatchify and return_dict: + raise ValueError( + f"Cannot both define `unpatchify`: {unpatchify} and `return_dict`: {return_dict} since when" + f" `unpatchify` is {unpatchify} the returned output is of shape (batch_size, seq_len, hidden_dim)" + " rather than (batch_size, num_channels, height, width)." + ) + + # 1. Input + if not hidden_states_is_embedding: + hidden_states = self.pos_embed(hidden_states) + + # 2. Blocks + + # In ("downsample") blocks + skips = [] + for in_block in self.transformer_in_blocks: + hidden_states = in_block( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + skips.append(hidden_states) + + # Mid block + hidden_states = self.transformer_mid_block(hidden_states) + + # Out ("upsample") blocks + for out_block in self.transformer_out_blocks: + hidden_states = out_block["skip"](hidden_states, skips.pop()) + hidden_states = out_block["block"]( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + + # 3. Output + # Don't support AdaLayerNorm for now, so no conditioning/scale/shift logic + hidden_states = self.norm_out(hidden_states) + # hidden_states = self.proj_out(hidden_states) + + if unpatchify: + # unpatchify + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + else: + output = hidden_states + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) + + +class UniDiffuserModel(ModelMixin, ConfigMixin): + """ + Transformer model for a image-text [UniDiffuser](https://huggingface.co/papers/2303.06555) model. This is a + modification of [`UTransformer2DModel`] with input and output heads for the VAE-embedded latent image, the + CLIP-embedded image, and the CLIP-embedded prompt (see paper for more details). + + Parameters: + text_dim (`int`): The hidden dimension of the CLIP text model used to embed images. + clip_img_dim (`int`): The hidden dimension of the CLIP vision model used to embed prompts. + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input. + out_channels (`int`, *optional*): + The number of output channels; if `None`, defaults to `in_channels`. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + norm_num_groups (`int`, *optional*, defaults to `32`): + The number of groups to use when performing Group Normalization. + cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + patch_size (`int`, *optional*, defaults to 2): + The patch size to use in the patch embedding. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + use_linear_projection (int, *optional*): TODO: Not used + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used in each + transformer block. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float32 when performing the attention calculation. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`. + block_type (`str`, *optional*, defaults to `"unidiffuser"`): + The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual + backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard + behavior in `diffusers`.) + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm + (`pre_layer_norm = False`). + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + use_patch_pos_embed (`bool`, *optional*): + Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`). + ff_final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + use_data_type_embedding (`bool`, *optional*): + Whether to use a data type embedding. This is only relevant for UniDiffuser-v1 style models; UniDiffuser-v1 + is continue-trained from UniDiffuser-v0 on non-publically-available data and accepts a `data_type` + argument, which can either be `1` to use the weights trained on non-publically-available data or `0` + otherwise. This argument is subsequently embedded by the data type embedding, if used. + """ + + @register_to_config + def __init__( + self, + text_dim: int = 768, + clip_img_dim: int = 512, + num_text_tokens: int = 77, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: int | None = None, + out_channels: int | None = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: int | None = None, + attention_bias: bool = False, + sample_size: int | None = None, + num_vector_embeds: int | None = None, + patch_size: int | None = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: int | None = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + block_type: str = "unidiffuser", + pre_layer_norm: bool = False, + use_timestep_embedding=False, + norm_elementwise_affine: bool = True, + use_patch_pos_embed=False, + ff_final_dropout: bool = True, + use_data_type_embedding: bool = False, + ): + super().__init__() + + # 0. Handle dimensions + self.inner_dim = num_attention_heads * attention_head_dim + + assert sample_size is not None, "UniDiffuserModel over patched input must provide sample_size" + self.sample_size = sample_size + self.in_channels = in_channels + self.out_channels = in_channels if out_channels is None else out_channels + + self.patch_size = patch_size + # Assume image is square... + self.num_patches = (self.sample_size // patch_size) * (self.sample_size // patch_size) + + # 1. Define input layers + # 1.1 Input layers for text and image input + # For now, only support patch input for VAE latent image input + self.vae_img_in = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=self.inner_dim, + use_pos_embed=use_patch_pos_embed, + ) + self.clip_img_in = nn.Linear(clip_img_dim, self.inner_dim) + self.text_in = nn.Linear(text_dim, self.inner_dim) + + # 1.2. Timestep embeddings for t_img, t_text + self.timestep_img_proj = Timesteps( + self.inner_dim, + flip_sin_to_cos=True, + downscale_freq_shift=0, + ) + self.timestep_img_embed = ( + TimestepEmbedding( + self.inner_dim, + 4 * self.inner_dim, + out_dim=self.inner_dim, + ) + if use_timestep_embedding + else nn.Identity() + ) + + self.timestep_text_proj = Timesteps( + self.inner_dim, + flip_sin_to_cos=True, + downscale_freq_shift=0, + ) + self.timestep_text_embed = ( + TimestepEmbedding( + self.inner_dim, + 4 * self.inner_dim, + out_dim=self.inner_dim, + ) + if use_timestep_embedding + else nn.Identity() + ) + + # 1.3. Positional embedding + self.num_text_tokens = num_text_tokens + self.num_tokens = 1 + 1 + num_text_tokens + 1 + self.num_patches + self.pos_embed = nn.Parameter(torch.zeros(1, self.num_tokens, self.inner_dim)) + self.pos_embed_drop = nn.Dropout(p=dropout) + trunc_normal_(self.pos_embed, std=0.02) + + # 1.4. Handle data type token embeddings for UniDiffuser-V1, if necessary + self.use_data_type_embedding = use_data_type_embedding + if self.use_data_type_embedding: + self.data_type_token_embedding = nn.Embedding(2, self.inner_dim) + self.data_type_pos_embed_token = nn.Parameter(torch.zeros(1, 1, self.inner_dim)) + + # 2. Define transformer blocks + self.transformer = UTransformer2DModel( + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + in_channels=in_channels, + out_channels=out_channels, + num_layers=num_layers, + dropout=dropout, + norm_num_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attention_bias=attention_bias, + sample_size=sample_size, + num_vector_embeds=num_vector_embeds, + patch_size=patch_size, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + block_type=block_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + use_patch_pos_embed=use_patch_pos_embed, + ff_final_dropout=ff_final_dropout, + ) + + # 3. Define output layers + patch_dim = (patch_size**2) * out_channels + self.vae_img_out = nn.Linear(self.inner_dim, patch_dim) + self.clip_img_out = nn.Linear(self.inner_dim, clip_img_dim) + self.text_out = nn.Linear(self.inner_dim, text_dim) + + @torch.jit.ignore + def no_weight_decay(self): + return {"pos_embed"} + + def forward( + self, + latent_image_embeds: torch.Tensor, + image_embeds: torch.Tensor, + prompt_embeds: torch.Tensor, + timestep_img: torch.Tensor | float | int, + timestep_text: torch.Tensor | float | int, + data_type: torch.Tensor | float | int | None = 1, + encoder_hidden_states=None, + cross_attention_kwargs=None, + ): + """ + Args: + latent_image_embeds (`torch.Tensor` of shape `(batch size, latent channels, height, width)`): + Latent image representation from the VAE encoder. + image_embeds (`torch.Tensor` of shape `(batch size, 1, clip_img_dim)`): + CLIP-embedded image representation (unsqueezed in the first dimension). + prompt_embeds (`torch.Tensor` of shape `(batch size, seq_len, text_dim)`): + CLIP-embedded text representation. + timestep_img (`torch.long` or `float` or `int`): + Current denoising step for the image. + timestep_text (`torch.long` or `float` or `int`): + Current denoising step for the text. + data_type: (`torch.int` or `float` or `int`, *optional*, defaults to `1`): + Only used in UniDiffuser-v1-style models. Can be either `1`, to use weights trained on nonpublic data, + or `0` otherwise. + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + cross_attention_kwargs (*optional*): + Keyword arguments to supply to the cross attention layers, if used. + + + Returns: + `tuple`: Returns relevant parts of the model's noise prediction: the first element of the tuple is tbe VAE + image embedding, the second element is the CLIP image embedding, and the third element is the CLIP text + embedding. + """ + batch_size = latent_image_embeds.shape[0] + + # 1. Input + # 1.1. Map inputs to shape (B, N, inner_dim) + vae_hidden_states = self.vae_img_in(latent_image_embeds) + clip_hidden_states = self.clip_img_in(image_embeds) + text_hidden_states = self.text_in(prompt_embeds) + + num_text_tokens, num_img_tokens = text_hidden_states.size(1), vae_hidden_states.size(1) + + # 1.2. Encode image timesteps to single token (B, 1, inner_dim) + if not torch.is_tensor(timestep_img): + timestep_img = torch.tensor([timestep_img], dtype=torch.long, device=vae_hidden_states.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep_img = timestep_img * torch.ones(batch_size, dtype=timestep_img.dtype, device=timestep_img.device) + + timestep_img_token = self.timestep_img_proj(timestep_img) + # t_img_token does not contain any weights and will always return f32 tensors + # but time_embedding might be fp16, so we need to cast here. + timestep_img_token = timestep_img_token.to(dtype=self.dtype) + timestep_img_token = self.timestep_img_embed(timestep_img_token) + timestep_img_token = timestep_img_token.unsqueeze(dim=1) + + # 1.3. Encode text timesteps to single token (B, 1, inner_dim) + if not torch.is_tensor(timestep_text): + timestep_text = torch.tensor([timestep_text], dtype=torch.long, device=vae_hidden_states.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep_text = timestep_text * torch.ones(batch_size, dtype=timestep_text.dtype, device=timestep_text.device) + + timestep_text_token = self.timestep_text_proj(timestep_text) + # t_text_token does not contain any weights and will always return f32 tensors + # but time_embedding might be fp16, so we need to cast here. + timestep_text_token = timestep_text_token.to(dtype=self.dtype) + timestep_text_token = self.timestep_text_embed(timestep_text_token) + timestep_text_token = timestep_text_token.unsqueeze(dim=1) + + # 1.4. Concatenate all of the embeddings together. + if self.use_data_type_embedding: + assert data_type is not None, "data_type must be supplied if the model uses a data type embedding" + if not torch.is_tensor(data_type): + data_type = torch.tensor([data_type], dtype=torch.int, device=vae_hidden_states.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + data_type = data_type * torch.ones(batch_size, dtype=data_type.dtype, device=data_type.device) + + data_type_token = self.data_type_token_embedding(data_type).unsqueeze(dim=1) + hidden_states = torch.cat( + [ + timestep_img_token, + timestep_text_token, + data_type_token, + text_hidden_states, + clip_hidden_states, + vae_hidden_states, + ], + dim=1, + ) + else: + hidden_states = torch.cat( + [timestep_img_token, timestep_text_token, text_hidden_states, clip_hidden_states, vae_hidden_states], + dim=1, + ) + + # 1.5. Prepare the positional embeddings and add to hidden states + # Note: I think img_vae should always have the proper shape, so there's no need to interpolate + # the position embeddings. + if self.use_data_type_embedding: + pos_embed = torch.cat( + [self.pos_embed[:, : 1 + 1, :], self.data_type_pos_embed_token, self.pos_embed[:, 1 + 1 :, :]], dim=1 + ) + else: + pos_embed = self.pos_embed + hidden_states = hidden_states + pos_embed + hidden_states = self.pos_embed_drop(hidden_states) + + # 2. Blocks + hidden_states = self.transformer( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + timestep=None, + class_labels=None, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + hidden_states_is_embedding=True, + unpatchify=False, + )[0] + + # 3. Output + # Split out the predicted noise representation. + if self.use_data_type_embedding: + ( + t_img_token_out, + t_text_token_out, + data_type_token_out, + text_out, + img_clip_out, + img_vae_out, + ) = hidden_states.split((1, 1, 1, num_text_tokens, 1, num_img_tokens), dim=1) + else: + t_img_token_out, t_text_token_out, text_out, img_clip_out, img_vae_out = hidden_states.split( + (1, 1, num_text_tokens, 1, num_img_tokens), dim=1 + ) + + img_vae_out = self.vae_img_out(img_vae_out) + + # unpatchify + height = width = int(img_vae_out.shape[1] ** 0.5) + img_vae_out = img_vae_out.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + img_vae_out = torch.einsum("nhwpqc->nchpwq", img_vae_out) + img_vae_out = img_vae_out.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + + img_clip_out = self.clip_img_out(img_clip_out) + + text_out = self.text_out(text_out) + + return img_vae_out, img_clip_out, text_out diff --git a/diffusers/pipelines/deprecated/unidiffuser/pipeline_unidiffuser.py b/diffusers/pipelines/deprecated/unidiffuser/pipeline_unidiffuser.py new file mode 100644 index 0000000000000000000000000000000000000000..3b3db863744b3d63bdb57d3105d6919d46daba5d --- /dev/null +++ b/diffusers/pipelines/deprecated/unidiffuser/pipeline_unidiffuser.py @@ -0,0 +1,1405 @@ +import inspect +from dataclasses import dataclass +from typing import Callable + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + GPT2Tokenizer, +) + +from ....image_processor import VaeImageProcessor +from ....loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ....models import AutoencoderKL +from ....models.lora import adjust_lora_scale_text_encoder +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ....utils.outputs import BaseOutput +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline +from .modeling_text_decoder import UniDiffuserTextDecoder +from .modeling_uvit import UniDiffuserModel + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# New BaseOutput child class for joint image-text output +@dataclass +class ImageTextPipelineOutput(BaseOutput): + """ + Output class for joint image-text pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + text (`list[str]` or `list[list[str]]`) + list of generated text strings of length `batch_size` or a list of list of strings whose outer list has + length `batch_size`. + """ + + images: list[PIL.Image.Image] | np.ndarray | None + text: list[str] | list[list[str]] | None + + +class UniDiffuserPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + r""" + Pipeline for a bimodal image-text model which supports unconditional text and image generation, text-conditioned + image generation, image-conditioned text generation, and joint image-text generation. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. This + is part of the UniDiffuser image representation along with the CLIP vision encoding. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + image_encoder ([`CLIPVisionModel`]): + A [`~transformers.CLIPVisionModel`] to encode images as part of its image representation along with the VAE + latent representation. + image_processor ([`CLIPImageProcessor`]): + [`~transformers.CLIPImageProcessor`] to preprocess an image before CLIP encoding it with `image_encoder`. + clip_tokenizer ([`CLIPTokenizer`]): + A [`~transformers.CLIPTokenizer`] to tokenize the prompt before encoding it with `text_encoder`. + text_decoder ([`UniDiffuserTextDecoder`]): + Frozen text decoder. This is a GPT-style model which is used to generate text from the UniDiffuser + embedding. + text_tokenizer ([`GPT2Tokenizer`]): + A [`~transformers.GPT2Tokenizer`] to decode text for text generation; used along with the `text_decoder`. + unet ([`UniDiffuserModel`]): + A [U-ViT](https://github.com/baofff/U-ViT) model with UNNet-style skip connections between transformer + layers to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image and/or text latents. The + original UniDiffuser paper uses the [`DPMSolverMultistepScheduler`] scheduler. + """ + + _last_supported_version = "0.33.1" + # TODO: support for moving submodules for components with enable_model_cpu_offload + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae->text_decoder" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + image_encoder: CLIPVisionModelWithProjection, + clip_image_processor: CLIPImageProcessor, + clip_tokenizer: CLIPTokenizer, + text_decoder: UniDiffuserTextDecoder, + text_tokenizer: GPT2Tokenizer, + unet: UniDiffuserModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + if text_encoder.config.hidden_size != text_decoder.prefix_inner_dim: + raise ValueError( + f"The text encoder hidden size and text decoder prefix inner dim must be the same, but" + f" `text_encoder.config.hidden_size`: {text_encoder.config.hidden_size} and `text_decoder.prefix_inner_dim`: {text_decoder.prefix_inner_dim}" + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + image_encoder=image_encoder, + clip_image_processor=clip_image_processor, + clip_tokenizer=clip_tokenizer, + text_decoder=text_decoder, + text_tokenizer=text_tokenizer, + unet=unet, + scheduler=scheduler, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.num_channels_latents = vae.config.latent_channels + self.text_encoder_seq_len = text_encoder.config.max_position_embeddings + self.text_encoder_hidden_size = text_encoder.config.hidden_size + self.image_encoder_projection_dim = image_encoder.config.projection_dim + self.unet_resolution = unet.config.sample_size + + self.text_intermediate_dim = self.text_encoder_hidden_size + if self.text_decoder.prefix_hidden_dim is not None: + self.text_intermediate_dim = self.text_decoder.prefix_hidden_dim + + self.mode = None + + # TODO: handle safety checking? + self.safety_checker = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def _infer_mode(self, prompt, prompt_embeds, image, latents, prompt_latents, vae_latents, clip_latents): + r""" + Infer the generation task ('mode') from the inputs to `__call__`. If the mode has been manually set, the set + mode will be used. + """ + prompt_available = (prompt is not None) or (prompt_embeds is not None) + image_available = image is not None + input_available = prompt_available or image_available + + prompt_latents_available = prompt_latents is not None + vae_latents_available = vae_latents is not None + clip_latents_available = clip_latents is not None + full_latents_available = latents is not None + image_latents_available = vae_latents_available and clip_latents_available + all_indv_latents_available = prompt_latents_available and image_latents_available + + if self.mode is not None: + # Preferentially use the mode set by the user + mode = self.mode + elif prompt_available: + mode = "text2img" + elif image_available: + mode = "img2text" + else: + # Neither prompt nor image supplied, infer based on availability of latents + if full_latents_available or all_indv_latents_available: + mode = "joint" + elif prompt_latents_available: + mode = "text" + elif image_latents_available: + mode = "img" + else: + # No inputs or latents available + mode = "joint" + + # Give warnings for ambiguous cases + if self.mode is None and prompt_available and image_available: + logger.warning( + f"You have supplied both a text prompt and image to the pipeline and mode has not been set manually," + f" defaulting to mode '{mode}'." + ) + + if self.mode is None and not input_available: + if vae_latents_available != clip_latents_available: + # Exactly one of vae_latents and clip_latents is supplied + logger.warning( + f"You have supplied exactly one of `vae_latents` and `clip_latents`, whereas either both or none" + f" are expected to be supplied. Defaulting to mode '{mode}'." + ) + elif not prompt_latents_available and not vae_latents_available and not clip_latents_available: + # No inputs or latents supplied + logger.warning( + f"No inputs or latents have been supplied, and mode has not been manually set," + f" defaulting to mode '{mode}'." + ) + + return mode + + # Functions to manually set the mode + def set_text_mode(self): + r"""Manually set the generation mode to unconditional ("marginal") text generation.""" + self.mode = "text" + + def set_image_mode(self): + r"""Manually set the generation mode to unconditional ("marginal") image generation.""" + self.mode = "img" + + def set_text_to_image_mode(self): + r"""Manually set the generation mode to text-conditioned image generation.""" + self.mode = "text2img" + + def set_image_to_text_mode(self): + r"""Manually set the generation mode to image-conditioned text generation.""" + self.mode = "img2text" + + def set_joint_mode(self): + r"""Manually set the generation mode to unconditional joint image-text generation.""" + self.mode = "joint" + + def reset_mode(self): + r"""Removes a manually set mode; after calling this, the pipeline will infer the mode from inputs.""" + self.mode = None + + def _infer_batch_size( + self, + mode, + prompt, + prompt_embeds, + image, + num_images_per_prompt, + num_prompts_per_image, + latents, + prompt_latents, + vae_latents, + clip_latents, + ): + r"""Infers the batch size and multiplier depending on mode and supplied arguments to `__call__`.""" + if num_images_per_prompt is None: + num_images_per_prompt = 1 + if num_prompts_per_image is None: + num_prompts_per_image = 1 + + assert num_images_per_prompt > 0, "num_images_per_prompt must be a positive integer" + assert num_prompts_per_image > 0, "num_prompts_per_image must be a positive integer" + + if mode in ["text2img"]: + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + # Either prompt or prompt_embeds must be present for text2img. + batch_size = prompt_embeds.shape[0] + multiplier = num_images_per_prompt + elif mode in ["img2text"]: + if isinstance(image, PIL.Image.Image): + batch_size = 1 + else: + # Image must be available and type either PIL.Image.Image or torch.Tensor. + # Not currently supporting something like image_embeds. + batch_size = image.shape[0] + multiplier = num_prompts_per_image + elif mode in ["img"]: + if vae_latents is not None: + batch_size = vae_latents.shape[0] + elif clip_latents is not None: + batch_size = clip_latents.shape[0] + else: + batch_size = 1 + multiplier = num_images_per_prompt + elif mode in ["text"]: + if prompt_latents is not None: + batch_size = prompt_latents.shape[0] + else: + batch_size = 1 + multiplier = num_prompts_per_image + elif mode in ["joint"]: + if latents is not None: + batch_size = latents.shape[0] + elif prompt_latents is not None: + batch_size = prompt_latents.shape[0] + elif vae_latents is not None: + batch_size = vae_latents.shape[0] + elif clip_latents is not None: + batch_size = clip_latents.shape[0] + else: + batch_size = 1 + + if num_images_per_prompt == num_prompts_per_image: + multiplier = num_images_per_prompt + else: + multiplier = min(num_images_per_prompt, num_prompts_per_image) + logger.warning( + f"You are using mode `{mode}` and `num_images_per_prompt`: {num_images_per_prompt} and" + f" num_prompts_per_image: {num_prompts_per_image} are not equal. Using batch size equal to" + f" `min(num_images_per_prompt, num_prompts_per_image) = {batch_size}." + ) + return batch_size, multiplier + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with self.tokenizer->self.clip_tokenizer + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.clip_tokenizer) + + text_inputs = self.clip_tokenizer( + prompt, + padding="max_length", + max_length=self.clip_tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.clip_tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.clip_tokenizer.batch_decode( + untruncated_ids[:, self.clip_tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.clip_tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.clip_tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.clip_tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_instruct_pix2pix.StableDiffusionInstructPix2PixPipeline.prepare_image_latents + # Add num_prompts_per_image argument, sample from autoencoder moment distribution + def encode_image_vae_latents( + self, + image, + batch_size, + num_prompts_per_image, + dtype, + device, + do_classifier_free_guidance, + generator=None, + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_prompts_per_image + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if isinstance(generator, list): + image_latents = [ + self.vae.encode(image[i : i + 1]).latent_dist.sample(generator=generator[i]) + * self.vae.config.scaling_factor + for i in range(batch_size) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = self.vae.encode(image).latent_dist.sample(generator=generator) + # Scale image_latents by the VAE's scaling factor + image_latents = image_latents * self.vae.config.scaling_factor + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if do_classifier_free_guidance: + uncond_image_latents = torch.zeros_like(image_latents) + image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0) + + return image_latents + + def encode_image_clip_latents( + self, + image, + batch_size, + num_prompts_per_image, + dtype, + device, + generator=None, + ): + # Map image to CLIP embedding. + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + preprocessed_image = self.clip_image_processor.preprocess( + image, + return_tensors="pt", + ) + preprocessed_image = preprocessed_image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_prompts_per_image + if isinstance(generator, list): + image_latents = [ + self.image_encoder(**preprocessed_image[i : i + 1]).image_embeds for i in range(batch_size) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = self.image_encoder(**preprocessed_image).image_embeds + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + return image_latents + + def prepare_text_latents( + self, batch_size, num_images_per_prompt, seq_len, hidden_size, dtype, device, generator, latents=None + ): + # Prepare latents for the CLIP embedded prompt. + shape = (batch_size * num_images_per_prompt, seq_len, hidden_size) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + # latents is assumed to have shace (B, L, D) + latents = latents.repeat(num_images_per_prompt, 1, 1) + latents = latents.to(device=device, dtype=dtype) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + # Rename prepare_latents -> prepare_image_vae_latents and add num_prompts_per_image argument. + def prepare_image_vae_latents( + self, + batch_size, + num_prompts_per_image, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + shape = ( + batch_size * num_prompts_per_image, + num_channels_latents, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + # latents is assumed to have shape (B, C, H, W) + latents = latents.repeat(num_prompts_per_image, 1, 1, 1) + latents = latents.to(device=device, dtype=dtype) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_clip_latents( + self, batch_size, num_prompts_per_image, clip_img_dim, dtype, device, generator, latents=None + ): + # Prepare latents for the CLIP embedded image. + shape = (batch_size * num_prompts_per_image, 1, clip_img_dim) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + # latents is assumed to have shape (B, L, D) + latents = latents.repeat(num_prompts_per_image, 1, 1) + latents = latents.to(device=device, dtype=dtype) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def decode_text_latents(self, text_latents, device): + output_token_list, seq_lengths = self.text_decoder.generate_captions( + text_latents, self.text_tokenizer.eos_token_id, device=device + ) + output_list = output_token_list.cpu().numpy() + generated_text = [ + self.text_tokenizer.decode(output[: int(length)], skip_special_tokens=True) + for output, length in zip(output_list, seq_lengths) + ] + return generated_text + + def _split(self, x, height, width): + r""" + Splits a flattened embedding x of shape (B, C * H * W + clip_img_dim) into two tensors of shape (B, C, H, W) + and (B, 1, clip_img_dim) + """ + batch_size = x.shape[0] + latent_height = height // self.vae_scale_factor + latent_width = width // self.vae_scale_factor + img_vae_dim = self.num_channels_latents * latent_height * latent_width + + img_vae, img_clip = x.split([img_vae_dim, self.image_encoder_projection_dim], dim=1) + + img_vae = torch.reshape(img_vae, (batch_size, self.num_channels_latents, latent_height, latent_width)) + img_clip = torch.reshape(img_clip, (batch_size, 1, self.image_encoder_projection_dim)) + return img_vae, img_clip + + def _combine(self, img_vae, img_clip): + r""" + Combines a latent image img_vae of shape (B, C, H, W) and a CLIP-embedded image img_clip of shape (B, 1, + clip_img_dim) into a single tensor of shape (B, C * H * W + clip_img_dim). + """ + img_vae = torch.reshape(img_vae, (img_vae.shape[0], -1)) + img_clip = torch.reshape(img_clip, (img_clip.shape[0], -1)) + return torch.concat([img_vae, img_clip], dim=-1) + + def _split_joint(self, x, height, width): + r""" + Splits a flattened embedding x of shape (B, C * H * W + clip_img_dim + text_seq_len * text_dim] into (img_vae, + img_clip, text) where img_vae is of shape (B, C, H, W), img_clip is of shape (B, 1, clip_img_dim), and text is + of shape (B, text_seq_len, text_dim). + """ + batch_size = x.shape[0] + latent_height = height // self.vae_scale_factor + latent_width = width // self.vae_scale_factor + img_vae_dim = self.num_channels_latents * latent_height * latent_width + text_dim = self.text_encoder_seq_len * self.text_intermediate_dim + + img_vae, img_clip, text = x.split([img_vae_dim, self.image_encoder_projection_dim, text_dim], dim=1) + + img_vae = torch.reshape(img_vae, (batch_size, self.num_channels_latents, latent_height, latent_width)) + img_clip = torch.reshape(img_clip, (batch_size, 1, self.image_encoder_projection_dim)) + text = torch.reshape(text, (batch_size, self.text_encoder_seq_len, self.text_intermediate_dim)) + return img_vae, img_clip, text + + def _combine_joint(self, img_vae, img_clip, text): + r""" + Combines a latent image img_vae of shape (B, C, H, W), a CLIP-embedded image img_clip of shape (B, L_img, + clip_img_dim), and a text embedding text of shape (B, L_text, text_dim) into a single embedding x of shape (B, + C * H * W + L_img * clip_img_dim + L_text * text_dim). + """ + img_vae = torch.reshape(img_vae, (img_vae.shape[0], -1)) + img_clip = torch.reshape(img_clip, (img_clip.shape[0], -1)) + text = torch.reshape(text, (text.shape[0], -1)) + return torch.concat([img_vae, img_clip, text], dim=-1) + + def _get_noise_pred( + self, + mode, + latents, + t, + prompt_embeds, + img_vae, + img_clip, + max_timestep, + data_type, + guidance_scale, + generator, + device, + height, + width, + ): + r""" + Gets the noise prediction using the `unet` and performs classifier-free guidance, if necessary. + """ + if mode == "joint": + # Joint text-image generation + img_vae_latents, img_clip_latents, text_latents = self._split_joint(latents, height, width) + + img_vae_out, img_clip_out, text_out = self.unet( + img_vae_latents, img_clip_latents, text_latents, timestep_img=t, timestep_text=t, data_type=data_type + ) + + x_out = self._combine_joint(img_vae_out, img_clip_out, text_out) + + if guidance_scale <= 1.0: + return x_out + + # Classifier-free guidance + img_vae_T = randn_tensor(img_vae.shape, generator=generator, device=device, dtype=img_vae.dtype) + img_clip_T = randn_tensor(img_clip.shape, generator=generator, device=device, dtype=img_clip.dtype) + text_T = randn_tensor(prompt_embeds.shape, generator=generator, device=device, dtype=prompt_embeds.dtype) + + _, _, text_out_uncond = self.unet( + img_vae_T, img_clip_T, text_latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type + ) + + img_vae_out_uncond, img_clip_out_uncond, _ = self.unet( + img_vae_latents, + img_clip_latents, + text_T, + timestep_img=t, + timestep_text=max_timestep, + data_type=data_type, + ) + + x_out_uncond = self._combine_joint(img_vae_out_uncond, img_clip_out_uncond, text_out_uncond) + + return guidance_scale * x_out + (1.0 - guidance_scale) * x_out_uncond + elif mode == "text2img": + # Text-conditioned image generation + img_vae_latents, img_clip_latents = self._split(latents, height, width) + + img_vae_out, img_clip_out, text_out = self.unet( + img_vae_latents, img_clip_latents, prompt_embeds, timestep_img=t, timestep_text=0, data_type=data_type + ) + + img_out = self._combine(img_vae_out, img_clip_out) + + if guidance_scale <= 1.0: + return img_out + + # Classifier-free guidance + text_T = randn_tensor(prompt_embeds.shape, generator=generator, device=device, dtype=prompt_embeds.dtype) + + img_vae_out_uncond, img_clip_out_uncond, text_out_uncond = self.unet( + img_vae_latents, + img_clip_latents, + text_T, + timestep_img=t, + timestep_text=max_timestep, + data_type=data_type, + ) + + img_out_uncond = self._combine(img_vae_out_uncond, img_clip_out_uncond) + + return guidance_scale * img_out + (1.0 - guidance_scale) * img_out_uncond + elif mode == "img2text": + # Image-conditioned text generation + img_vae_out, img_clip_out, text_out = self.unet( + img_vae, img_clip, latents, timestep_img=0, timestep_text=t, data_type=data_type + ) + + if guidance_scale <= 1.0: + return text_out + + # Classifier-free guidance + img_vae_T = randn_tensor(img_vae.shape, generator=generator, device=device, dtype=img_vae.dtype) + img_clip_T = randn_tensor(img_clip.shape, generator=generator, device=device, dtype=img_clip.dtype) + + img_vae_out_uncond, img_clip_out_uncond, text_out_uncond = self.unet( + img_vae_T, img_clip_T, latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type + ) + + return guidance_scale * text_out + (1.0 - guidance_scale) * text_out_uncond + elif mode == "text": + # Unconditional ("marginal") text generation (no CFG) + img_vae_out, img_clip_out, text_out = self.unet( + img_vae, img_clip, latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type + ) + + return text_out + elif mode == "img": + # Unconditional ("marginal") image generation (no CFG) + img_vae_latents, img_clip_latents = self._split(latents, height, width) + + img_vae_out, img_clip_out, text_out = self.unet( + img_vae_latents, + img_clip_latents, + prompt_embeds, + timestep_img=t, + timestep_text=max_timestep, + data_type=data_type, + ) + + img_out = self._combine(img_vae_out, img_clip_out) + return img_out + + def check_latents_shape(self, latents_name, latents, expected_shape): + latents_shape = latents.shape + expected_num_dims = len(expected_shape) + 1 # expected dimensions plus the batch dimension + expected_shape_str = ", ".join(str(dim) for dim in expected_shape) + if len(latents_shape) != expected_num_dims: + raise ValueError( + f"`{latents_name}` should have shape (batch_size, {expected_shape_str}), but the current shape" + f" {latents_shape} has {len(latents_shape)} dimensions." + ) + for i in range(1, expected_num_dims): + if latents_shape[i] != expected_shape[i - 1]: + raise ValueError( + f"`{latents_name}` should have shape (batch_size, {expected_shape_str}), but the current shape" + f" {latents_shape} has {latents_shape[i]} != {expected_shape[i - 1]} at dimension {i}." + ) + + def check_inputs( + self, + mode, + prompt, + image, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + latents=None, + prompt_latents=None, + vae_latents=None, + clip_latents=None, + ): + # Check inputs before running the generative process. + if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0: + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor} but are {height} and {width}." + ) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if mode == "text2img": + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if mode == "img2text": + if image is None: + raise ValueError("`img2text` mode requires an image to be provided.") + + # Check provided latents + latent_height = height // self.vae_scale_factor + latent_width = width // self.vae_scale_factor + full_latents_available = latents is not None + prompt_latents_available = prompt_latents is not None + vae_latents_available = vae_latents is not None + clip_latents_available = clip_latents is not None + + if full_latents_available: + individual_latents_available = ( + prompt_latents is not None or vae_latents is not None or clip_latents is not None + ) + if individual_latents_available: + logger.warning( + "You have supplied both `latents` and at least one of `prompt_latents`, `vae_latents`, and" + " `clip_latents`. The value of `latents` will override the value of any individually supplied latents." + ) + # Check shape of full latents + img_vae_dim = self.num_channels_latents * latent_height * latent_width + text_dim = self.text_encoder_seq_len * self.text_encoder_hidden_size + latents_dim = img_vae_dim + self.image_encoder_projection_dim + text_dim + latents_expected_shape = (latents_dim,) + self.check_latents_shape("latents", latents, latents_expected_shape) + + # Check individual latent shapes, if present + if prompt_latents_available: + prompt_latents_expected_shape = (self.text_encoder_seq_len, self.text_encoder_hidden_size) + self.check_latents_shape("prompt_latents", prompt_latents, prompt_latents_expected_shape) + + if vae_latents_available: + vae_latents_expected_shape = (self.num_channels_latents, latent_height, latent_width) + self.check_latents_shape("vae_latents", vae_latents, vae_latents_expected_shape) + + if clip_latents_available: + clip_latents_expected_shape = (1, self.image_encoder_projection_dim) + self.check_latents_shape("clip_latents", clip_latents, clip_latents_expected_shape) + + if mode in ["text2img", "img"] and vae_latents_available and clip_latents_available: + if vae_latents.shape[0] != clip_latents.shape[0]: + raise ValueError( + f"Both `vae_latents` and `clip_latents` are supplied, but their batch dimensions are not equal:" + f" {vae_latents.shape[0]} != {clip_latents.shape[0]}." + ) + + if mode == "joint" and prompt_latents_available and vae_latents_available and clip_latents_available: + if prompt_latents.shape[0] != vae_latents.shape[0] or prompt_latents.shape[0] != clip_latents.shape[0]: + raise ValueError( + f"All of `prompt_latents`, `vae_latents`, and `clip_latents` are supplied, but their batch" + f" dimensions are not equal: {prompt_latents.shape[0]} != {vae_latents.shape[0]}" + f" != {clip_latents.shape[0]}." + ) + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] | None = None, + image: torch.Tensor | PIL.Image.Image | None = None, + height: int | None = None, + width: int | None = None, + data_type: int | None = 1, + num_inference_steps: int = 50, + guidance_scale: float = 8.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + num_prompts_per_image: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_latents: torch.Tensor | None = None, + vae_latents: torch.Tensor | None = None, + clip_latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + Required for text-conditioned image generation (`text2img`) mode. + image (`torch.Tensor` or `PIL.Image.Image`, *optional*): + `Image` or tensor representing an image batch. Required for image-conditioned text generation + (`img2text`) mode. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + data_type (`int`, *optional*, defaults to 1): + The data type (either 0 or 1). Only used if you are loading a checkpoint which supports a data type + embedding; this is added for compatibility with the + [UniDiffuser-v1](https://huggingface.co/thu-ml/unidiffuser-v1) checkpoint. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 8.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). Used in + text-conditioned image generation (`text2img`) mode. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. Used in `text2img` (text-conditioned image generation) and + `img` mode. If the mode is joint and both `num_images_per_prompt` and `num_prompts_per_image` are + supplied, `min(num_images_per_prompt, num_prompts_per_image)` samples are generated. + num_prompts_per_image (`int`, *optional*, defaults to 1): + The number of prompts to generate per image. Used in `img2text` (image-conditioned text generation) and + `text` mode. If the mode is joint and both `num_images_per_prompt` and `num_prompts_per_image` are + supplied, `min(num_images_per_prompt, num_prompts_per_image)` samples are generated. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for joint + image-text generation. Can be used to tweak the same generation with different prompts. If not + provided, a latents tensor is generated by sampling using the supplied random `generator`. This assumes + a full set of VAE, CLIP, and text latents, if supplied, overrides the value of `prompt_latents`, + `vae_latents`, and `clip_latents`. + prompt_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for text + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + vae_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + clip_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. Used in text-conditioned + image generation (`text2img`) mode. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are be generated from the `negative_prompt` input argument. Used + in text-conditioned image generation (`text2img`) mode. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImageTextPipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Returns: + [`~pipelines.unidiffuser.ImageTextPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.unidiffuser.ImageTextPipelineOutput`] is returned, otherwise a + `tuple` is returned where the first element is a list with the generated images and the second element + is a list of generated texts. + """ + + # 0. Default height and width to unet + height = height or self.unet_resolution * self.vae_scale_factor + width = width or self.unet_resolution * self.vae_scale_factor + + # 1. Check inputs + # Recalculate mode for each call to the pipeline. + mode = self._infer_mode(prompt, prompt_embeds, image, latents, prompt_latents, vae_latents, clip_latents) + self.check_inputs( + mode, + prompt, + image, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + latents, + prompt_latents, + vae_latents, + clip_latents, + ) + + # 2. Define call parameters + batch_size, multiplier = self._infer_batch_size( + mode, + prompt, + prompt_embeds, + image, + num_images_per_prompt, + num_prompts_per_image, + latents, + prompt_latents, + vae_latents, + clip_latents, + ) + device = self._execution_device + reduce_text_emb_dim = self.text_intermediate_dim < self.text_encoder_hidden_size or self.mode != "text2img" + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + # Note that this differs from the formulation in the unidiffusers paper! + do_classifier_free_guidance = guidance_scale > 1.0 + + # check if scheduler is in sigmas space + # scheduler_is_in_sigma_space = hasattr(self.scheduler, "sigmas") + + # 3. Encode input prompt, if available; otherwise prepare text latents + if latents is not None: + # Overwrite individual latents + vae_latents, clip_latents, prompt_latents = self._split_joint(latents, height, width) + + if mode in ["text2img"]: + # 3.1. Encode input prompt, if available + assert prompt is not None or prompt_embeds is not None + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=multiplier, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # if do_classifier_free_guidance: + # prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + else: + # 3.2. Prepare text latent variables, if input not available + prompt_embeds = self.prepare_text_latents( + batch_size=batch_size, + num_images_per_prompt=multiplier, + seq_len=self.text_encoder_seq_len, + hidden_size=self.text_encoder_hidden_size, + dtype=self.text_encoder.dtype, # Should work with both full precision and mixed precision + device=device, + generator=generator, + latents=prompt_latents, + ) + + if reduce_text_emb_dim: + prompt_embeds = self.text_decoder.encode(prompt_embeds) + + # 4. Encode image, if available; otherwise prepare image latents + if mode in ["img2text"]: + # 4.1. Encode images, if available + assert image is not None, "`img2text` requires a conditioning image" + # Encode image using VAE + image_vae = self.image_processor.preprocess(image) + height, width = image_vae.shape[-2:] + image_vae_latents = self.encode_image_vae_latents( + image=image_vae, + batch_size=batch_size, + num_prompts_per_image=multiplier, + dtype=prompt_embeds.dtype, + device=device, + do_classifier_free_guidance=False, # Copied from InstructPix2Pix, don't use their version of CFG + generator=generator, + ) + + # Encode image using CLIP + image_clip_latents = self.encode_image_clip_latents( + image=image, + batch_size=batch_size, + num_prompts_per_image=multiplier, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + ) + # (batch_size, clip_hidden_size) => (batch_size, 1, clip_hidden_size) + image_clip_latents = image_clip_latents.unsqueeze(1) + else: + # 4.2. Prepare image latent variables, if input not available + # Prepare image VAE latents in latent space + image_vae_latents = self.prepare_image_vae_latents( + batch_size=batch_size, + num_prompts_per_image=multiplier, + num_channels_latents=self.num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=vae_latents, + ) + + # Prepare image CLIP latents + image_clip_latents = self.prepare_image_clip_latents( + batch_size=batch_size, + num_prompts_per_image=multiplier, + clip_img_dim=self.image_encoder_projection_dim, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=clip_latents, + ) + + # 5. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + # max_timestep = timesteps[0] + max_timestep = self.scheduler.config.num_train_timesteps + + # 6. Prepare latent variables + if mode == "joint": + latents = self._combine_joint(image_vae_latents, image_clip_latents, prompt_embeds) + elif mode in ["text2img", "img"]: + latents = self._combine(image_vae_latents, image_clip_latents) + elif mode in ["img2text", "text"]: + latents = prompt_embeds + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + logger.debug(f"Scheduler extra step kwargs: {extra_step_kwargs}") + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # predict the noise residual + # Also applies classifier-free guidance as described in the UniDiffuser paper + noise_pred = self._get_noise_pred( + mode, + latents, + t, + prompt_embeds, + image_vae_latents, + image_clip_latents, + max_timestep, + data_type, + guidance_scale, + generator, + device, + height, + width, + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing + image = None + text = None + if mode == "joint": + image_vae_latents, image_clip_latents, text_latents = self._split_joint(latents, height, width) + + if not output_type == "latent": + # Map latent VAE image back to pixel space + image = self.vae.decode(image_vae_latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = image_vae_latents + + text = self.decode_text_latents(text_latents, device) + elif mode in ["text2img", "img"]: + image_vae_latents, image_clip_latents = self._split(latents, height, width) + + if not output_type == "latent": + # Map latent VAE image back to pixel space + image = self.vae.decode(image_vae_latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = image_vae_latents + elif mode in ["img2text", "text"]: + text_latents = latents + text = self.decode_text_latents(text_latents, device) + + self.maybe_free_model_hooks() + + # 10. Postprocess the image, if necessary + if image is not None: + do_denormalize = [True] * image.shape[0] + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload last model to CPU + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.final_offload_hook.offload() + + if not return_dict: + return (image, text) + + return ImageTextPipelineOutput(images=image, text=text) diff --git a/diffusers/pipelines/deprecated/versatile_diffusion/__init__.py b/diffusers/pipelines/deprecated/versatile_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8ea6ef6e2f65b96aebebdf72cb80135003e4f08d --- /dev/null +++ b/diffusers/pipelines/deprecated/versatile_diffusion/__init__.py @@ -0,0 +1,71 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, + is_transformers_available, + is_transformers_version, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import ( + VersatileDiffusionDualGuidedPipeline, + VersatileDiffusionImageVariationPipeline, + VersatileDiffusionPipeline, + VersatileDiffusionTextToImagePipeline, + ) + + _dummy_objects.update( + { + "VersatileDiffusionDualGuidedPipeline": VersatileDiffusionDualGuidedPipeline, + "VersatileDiffusionImageVariationPipeline": VersatileDiffusionImageVariationPipeline, + "VersatileDiffusionPipeline": VersatileDiffusionPipeline, + "VersatileDiffusionTextToImagePipeline": VersatileDiffusionTextToImagePipeline, + } + ) +else: + _import_structure["modeling_text_unet"] = ["UNetFlatConditionModel"] + _import_structure["pipeline_versatile_diffusion"] = ["VersatileDiffusionPipeline"] + _import_structure["pipeline_versatile_diffusion_dual_guided"] = ["VersatileDiffusionDualGuidedPipeline"] + _import_structure["pipeline_versatile_diffusion_image_variation"] = ["VersatileDiffusionImageVariationPipeline"] + _import_structure["pipeline_versatile_diffusion_text_to_image"] = ["VersatileDiffusionTextToImagePipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import ( + VersatileDiffusionDualGuidedPipeline, + VersatileDiffusionImageVariationPipeline, + VersatileDiffusionPipeline, + VersatileDiffusionTextToImagePipeline, + ) + else: + from .pipeline_versatile_diffusion import VersatileDiffusionPipeline + from .pipeline_versatile_diffusion_dual_guided import VersatileDiffusionDualGuidedPipeline + from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline + from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py b/diffusers/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py new file mode 100644 index 0000000000000000000000000000000000000000..7be159d77af5ddf3392fbae4c1b808d953bdb801 --- /dev/null +++ b/diffusers/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py @@ -0,0 +1,2438 @@ +from typing import Any + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F + +from diffusers.utils import deprecate + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models import ModelMixin +from ....models.activations import get_activation +from ....models.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, + AttnProcessor, +) +from ....models.embeddings import ( + GaussianFourierProjection, + ImageHintTimeEmbedding, + ImageProjection, + ImageTimeEmbedding, + TextImageProjection, + TextImageTimeEmbedding, + TextTimeEmbedding, + TimestepEmbedding, + Timesteps, +) +from ....models.resnet import ResnetBlockCondNorm2D +from ....models.transformers.dual_transformer_2d import DualTransformer2DModel +from ....models.transformers.transformer_2d import Transformer2DModel +from ....models.unets.unet_2d_condition import UNet2DConditionOutput +from ....utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers +from ....utils.torch_utils import apply_freeu + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_down_block( + down_block_type, + num_layers, + in_channels, + out_channels, + temb_channels, + add_downsample, + resnet_eps, + resnet_act_fn, + num_attention_heads, + transformer_layers_per_block, + attention_type, + attention_head_dim, + resnet_groups=None, + cross_attention_dim=None, + downsample_padding=None, + dual_cross_attention=False, + use_linear_projection=False, + only_cross_attention=False, + upcast_attention=False, + resnet_time_scale_shift="default", + resnet_skip_time_act=False, + resnet_out_scale_factor=1.0, + cross_attention_norm=None, + dropout=0.0, +): + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlockFlat": + return DownBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "CrossAttnDownBlockFlat": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockFlat") + return CrossAttnDownBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + raise ValueError(f"{down_block_type} is not supported.") + + +def get_up_block( + up_block_type, + num_layers, + in_channels, + out_channels, + prev_output_channel, + temb_channels, + add_upsample, + resnet_eps, + resnet_act_fn, + num_attention_heads, + transformer_layers_per_block, + resolution_idx, + attention_type, + attention_head_dim, + resnet_groups=None, + cross_attention_dim=None, + dual_cross_attention=False, + use_linear_projection=False, + only_cross_attention=False, + upcast_attention=False, + resnet_time_scale_shift="default", + resnet_skip_time_act=False, + resnet_out_scale_factor=1.0, + cross_attention_norm=None, + dropout=0.0, +): + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlockFlat": + return UpBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "CrossAttnUpBlockFlat": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockFlat") + return CrossAttnUpBlockFlat( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + raise ValueError(f"{up_block_type} is not supported.") + + +class FourierEmbedder(nn.Module): + def __init__(self, num_freqs=64, temperature=100): + super().__init__() + + self.num_freqs = num_freqs + self.temperature = temperature + + freq_bands = temperature ** (torch.arange(num_freqs) / num_freqs) + freq_bands = freq_bands[None, None, None] + self.register_buffer("freq_bands", freq_bands, persistent=False) + + def __call__(self, x): + x = self.freq_bands * x.unsqueeze(-1) + return torch.stack((x.sin(), x.cos()), dim=-1).permute(0, 1, 3, 4, 2).reshape(*x.shape[:2], -1) + + +class GLIGENTextBoundingboxProjection(nn.Module): + def __init__(self, positive_len, out_dim, feature_type, fourier_freqs=8): + super().__init__() + self.positive_len = positive_len + self.out_dim = out_dim + + self.fourier_embedder = FourierEmbedder(num_freqs=fourier_freqs) + self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy + + if isinstance(out_dim, tuple): + out_dim = out_dim[0] + + if feature_type == "text-only": + self.linears = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + elif feature_type == "text-image": + self.linears_text = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.linears_image = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim])) + + def forward( + self, + boxes, + masks, + positive_embeddings=None, + phrases_masks=None, + image_masks=None, + phrases_embeddings=None, + image_embeddings=None, + ): + masks = masks.unsqueeze(-1) + + xyxy_embedding = self.fourier_embedder(boxes) + xyxy_null = self.null_position_feature.view(1, 1, -1) + xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null + + if positive_embeddings: + positive_null = self.null_positive_feature.view(1, 1, -1) + positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null + + objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1)) + else: + phrases_masks = phrases_masks.unsqueeze(-1) + image_masks = image_masks.unsqueeze(-1) + + text_null = self.null_text_feature.view(1, 1, -1) + image_null = self.null_image_feature.view(1, 1, -1) + + phrases_embeddings = phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null + image_embeddings = image_embeddings * image_masks + (1 - image_masks) * image_null + + objs_text = self.linears_text(torch.cat([phrases_embeddings, xyxy_embedding], dim=-1)) + objs_image = self.linears_image(torch.cat([image_embeddings, xyxy_embedding], dim=-1)) + objs = torch.cat([objs_text, objs_image], dim=1) + + return objs + + +class UNetFlatConditionModel(ModelMixin, ConfigMixin): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "DownBlockFlat")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlockFlatCrossAttn"`): + Block type for middle of UNet, it can be one of `UNetMidBlockFlatCrossAttn`, `UNetMidBlockFlat`, or + `UNetMidBlockFlatSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`tuple[str]`, *optional*, defaults to `("UpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int`, `tuple[int]`, or `tuple[tuple]` , *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlockFlat`], [`~models.unet_2d_blocks.CrossAttnUpBlockFlat`], + [`~models.unet_2d_blocks.UNetMidBlockFlatCrossAttn`]. + reverse_transformer_layers_per_block : (`tuple[tuple]`, *optional*, defaults to None): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling + blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `tuple[tuple]` and for + [`~models.unet_2d_blocks.CrossAttnDownBlockFlat`], [`~models.unet_2d_blocks.CrossAttnUpBlockFlat`], + [`~models.unet_2d_blocks.UNetMidBlockFlatCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlockFlat`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`, + *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`, + *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlockFlatSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + _no_split_modules = ["BasicTransformerBlock", "ResnetBlockFlat", "CrossAttnUpBlockFlat"] + + @register_to_config + def __init__( + self, + sample_size: int | None = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: tuple[str] = ( + "CrossAttnDownBlockFlat", + "CrossAttnDownBlockFlat", + "CrossAttnDownBlockFlat", + "DownBlockFlat", + ), + mid_block_type: str = "UNetMidBlockFlatCrossAttn", + up_block_types: tuple[str] = ( + "UpBlockFlat", + "CrossAttnUpBlockFlat", + "CrossAttnUpBlockFlat", + "CrossAttnUpBlockFlat", + ), + only_cross_attention: bool | tuple[bool] = False, + block_out_channels: tuple[int] = (320, 640, 1280, 1280), + layers_per_block: int | tuple[int] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + dropout: float = 0.0, + act_fn: str = "silu", + norm_num_groups: int | None = 32, + norm_eps: float = 1e-5, + cross_attention_dim: int | tuple[int] = 1280, + transformer_layers_per_block: int | tuple[int] | tuple[tuple] = 1, + reverse_transformer_layers_per_block: tuple[tuple[int]] | None = None, + encoder_hid_dim: int | None = None, + encoder_hid_dim_type: str | None = None, + attention_head_dim: int | tuple[int] = 8, + num_attention_heads: int | tuple[int] | None = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: str | None = None, + addition_embed_type: str | None = None, + addition_time_embed_dim: int | None = None, + num_class_embeds: int | None = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: int = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: int | None = None, + time_embedding_act_fn: str | None = None, + timestep_post_act: str | None = None, + time_cond_proj_dim: int | None = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: int | None = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: bool | None = None, + cross_attention_norm: str | None = None, + addition_embed_type_num_heads=64, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None: + for layer_number_per_block in transformer_layers_per_block: + if isinstance(layer_number_per_block, list): + raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.") + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = LinearMultiDim( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim_type`: {encoder_hid_dim_type} must be None, 'text_proj', 'text_image_proj' or 'image_proj'." + ) + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.down_blocks.append(down_block) + + # mid + if mid_block_type == "UNetMidBlockFlatCrossAttn": + self.mid_block = UNetMidBlockFlatCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim[-1], + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + elif mid_block_type == "UNetMidBlockFlatSimpleCrossAttn": + self.mid_block = UNetMidBlockFlatSimpleCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + cross_attention_dim=cross_attention_dim[-1], + attention_head_dim=attention_head_dim[-1], + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + only_cross_attention=mid_block_only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif mid_block_type == "UNetMidBlockFlat": + self.mid_block = UNetMidBlockFlat( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + num_layers=0, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + add_attention=False, + ) + elif mid_block_type is None: + self.mid_block = None + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = ( + list(reversed(transformer_layers_per_block)) + if reverse_transformer_layers_per_block is None + else reverse_transformer_layers_per_block + ) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resolution_idx=i, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = LinearMultiDim( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + if attention_type in ["gated", "gated-text-image"]: + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, (list, tuple)): + positive_len = cross_attention_dim[0] + + feature_type = "text-only" if attention_type == "gated" else "text-image" + self.position_net = GLIGENTextBoundingboxProjection( + positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type + ) + + @property + def attn_processors(self) -> dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor() + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor(self, processor: AttentionProcessor | dict[str, AttentionProcessor]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + def set_attention_slice(self, slice_size): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: list[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def enable_freeu(self, s1, s2, b1, b2): + r"""Enables the FreeU mechanism from https://huggingface.co/papers/2309.11497. + + The suffixes after the scaling factors represent the stage blocks where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that + are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + for i, upsample_block in enumerate(self.up_blocks): + setattr(upsample_block, "s1", s1) + setattr(upsample_block, "s2", s2) + setattr(upsample_block, "b1", b1) + setattr(upsample_block, "b2", b2) + + def disable_freeu(self): + """Disables the FreeU mechanism.""" + freeu_keys = {"s1", "s2", "b1", "b2"} + for i, upsample_block in enumerate(self.up_blocks): + for k in freeu_keys: + if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None: + setattr(upsample_block, k, None) + + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value) + are fused. For cross-attention modules, key and value projection matrices are fused. + + > [!WARNING] > This API is 🧪 experimental. + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + > [!WARNING] > This API is 🧪 experimental. + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) + + def unload_lora(self): + """Unloads LoRA weights.""" + deprecate( + "unload_lora", + "0.28.0", + "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().", + ) + for module in self.modules(): + if hasattr(module, "set_lora_layer"): + module.set_lora_layer(None) + + def forward( + self, + sample: torch.Tensor, + timestep: torch.Tensor | float | int, + encoder_hidden_states: torch.Tensor, + class_labels: torch.Tensor | None = None, + timestep_cond: torch.Tensor | None = None, + attention_mask: torch.Tensor | None = None, + cross_attention_kwargs: dict[str, Any] | None = None, + added_cond_kwargs: dict[str, torch.Tensor] | None = None, + down_block_additional_residuals: tuple[torch.Tensor] | None = None, + mid_block_additional_residual: torch.Tensor | None = None, + down_intrablock_additional_residuals: tuple[torch.Tensor] | None = None, + encoder_attention_mask: torch.Tensor | None = None, + return_dict: bool = True, + ) -> UNet2DConditionOutput | tuple: + r""" + The [`UNetFlatConditionModel`] forward method. + + Args: + sample (`torch.Tensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.Tensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.Tensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`): + Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed + through the `self.time_embedding` layer to obtain the timestep embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*): + A tuple of tensors that if specified are added to the residuals of down unet blocks. + mid_block_additional_residual: (`torch.Tensor`, *optional*): + A tensor that if specified is added to the residual of the middle unet block. + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttnProcessor`]. + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added to UNet long skip connections from down blocks to up blocks for + example from ControlNet side model(s) + mid_block_additional_residual (`torch.Tensor`, *optional*): + additional residual to be added to UNet mid block output, for example from ControlNet side model + down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) + + Returns: + [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, + otherwise a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + for dim in sample.shape[-2:]: + if dim % default_overall_up_factor != 0: + # Forward upsample size to force interpolation output size. + forward_upsample_size = True + break + + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + is_npu = sample.device.type == "npu" + if isinstance(timestep, float): + dtype = torch.float32 if (is_mps or is_npu) else torch.float64 + else: + dtype = torch.int32 if (is_mps or is_npu) else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb, hint = self.add_embedding(image_embs, hint) + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj": + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + image_embeds = self.encoder_hid_proj(image_embeds) + encoder_hidden_states = (encoder_hidden_states, image_embeds) + + # 2. pre-process + sample = self.conv_in(sample) + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + # 3. down + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets + is_adapter = down_intrablock_additional_residuals is not None + # maintain backward compatibility for legacy usage, where + # T2I-Adapter and ControlNet both use down_block_additional_residuals arg + # but can only use one or the other + if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None: + deprecate( + "T2I should not use down_block_additional_residuals", + "1.3.0", + "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \ + and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \ + for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ", + standard_warn=False, + ) + down_intrablock_additional_residuals = down_block_additional_residuals + is_adapter = True + + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlockFlat + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + scale=lora_scale, + ) + + # 6. post-process + if self.conv_norm_out: + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample) + + +class LinearMultiDim(nn.Linear): + def __init__(self, in_features, out_features=None, second_dim=4, *args, **kwargs): + in_features = [in_features, second_dim, 1] if isinstance(in_features, int) else list(in_features) + if out_features is None: + out_features = in_features + out_features = [out_features, second_dim, 1] if isinstance(out_features, int) else list(out_features) + self.in_features_multidim = in_features + self.out_features_multidim = out_features + super().__init__(np.array(in_features).prod(), np.array(out_features).prod()) + + def forward(self, input_tensor, *args, **kwargs): + shape = input_tensor.shape + n_dim = len(self.in_features_multidim) + input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_features) + output_tensor = super().forward(input_tensor) + output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_features_multidim) + return output_tensor + + +class ResnetBlockFlat(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + dropout=0.0, + temb_channels=512, + groups=32, + groups_out=None, + pre_norm=True, + eps=1e-6, + time_embedding_norm="default", + use_in_shortcut=None, + second_dim=4, + **kwargs, + ): + super().__init__() + self.pre_norm = pre_norm + self.pre_norm = True + + in_channels = [in_channels, second_dim, 1] if isinstance(in_channels, int) else list(in_channels) + self.in_channels_prod = np.array(in_channels).prod() + self.channels_multidim = in_channels + + if out_channels is not None: + out_channels = [out_channels, second_dim, 1] if isinstance(out_channels, int) else list(out_channels) + out_channels_prod = np.array(out_channels).prod() + self.out_channels_multidim = out_channels + else: + out_channels_prod = self.in_channels_prod + self.out_channels_multidim = self.channels_multidim + self.time_embedding_norm = time_embedding_norm + + if groups_out is None: + groups_out = groups + + self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=self.in_channels_prod, eps=eps, affine=True) + self.conv1 = torch.nn.Conv2d(self.in_channels_prod, out_channels_prod, kernel_size=1, padding=0) + + if temb_channels is not None: + self.time_emb_proj = torch.nn.Linear(temb_channels, out_channels_prod) + else: + self.time_emb_proj = None + + self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels_prod, eps=eps, affine=True) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels_prod, out_channels_prod, kernel_size=1, padding=0) + + self.nonlinearity = nn.SiLU() + + self.use_in_shortcut = ( + self.in_channels_prod != out_channels_prod if use_in_shortcut is None else use_in_shortcut + ) + + self.conv_shortcut = None + if self.use_in_shortcut: + self.conv_shortcut = torch.nn.Conv2d( + self.in_channels_prod, out_channels_prod, kernel_size=1, stride=1, padding=0 + ) + + def forward(self, input_tensor, temb): + shape = input_tensor.shape + n_dim = len(self.channels_multidim) + input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_channels_prod, 1, 1) + input_tensor = input_tensor.view(-1, self.in_channels_prod, 1, 1) + + hidden_states = input_tensor + + hidden_states = self.norm1(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + hidden_states = self.conv1(hidden_states) + + if temb is not None: + temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None] + hidden_states = hidden_states + temb + + hidden_states = self.norm2(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + + hidden_states = self.dropout(hidden_states) + hidden_states = self.conv2(hidden_states) + + if self.conv_shortcut is not None: + input_tensor = self.conv_shortcut(input_tensor) + + output_tensor = input_tensor + hidden_states + + output_tensor = output_tensor.view(*shape[0:-n_dim], -1) + output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_channels_multidim) + + return output_tensor + + +class DownBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlockFlat( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + LinearMultiDim( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.Tensor, temb: torch.Tensor | None = None + ) -> tuple[torch.Tensor, tuple[torch.Tensor, ...]]: + output_states = () + + for resnet in self.resnets: + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func(resnet, hidden_states, temb) + else: + hidden_states = resnet(hidden_states, temb) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class CrossAttnDownBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: int | tuple[int] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + add_downsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlockFlat( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + LinearMultiDim( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.Tensor, + temb: torch.Tensor | None = None, + encoder_hidden_states: torch.Tensor | None = None, + attention_mask: torch.Tensor | None = None, + cross_attention_kwargs: dict[str, Any] | None = None, + encoder_attention_mask: torch.Tensor | None = None, + additional_residuals: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, tuple[torch.Tensor, ...]]: + output_states = () + + blocks = list(zip(self.resnets, self.attentions)) + + for i, (resnet, attn) in enumerate(blocks): + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func(resnet, hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + # apply additional residuals to the output of the last pair of resnet and attention blocks + if i == len(blocks) - 1 and additional_residuals is not None: + hidden_states = hidden_states + additional_residuals + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +# Copied from diffusers.models.unets.unet_2d_blocks.UpBlock2D with UpBlock2D->UpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim +class UpBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: int | None = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlockFlat( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.Tensor, + res_hidden_states_tuple: tuple[torch.Tensor, ...], + temb: torch.Tensor | None = None, + upsample_size: int | None = None, + *args, + **kwargs, + ) -> torch.Tensor: + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func(resnet, hidden_states, temb) + else: + hidden_states = resnet(hidden_states, temb) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + +# Copied from diffusers.models.unets.unet_2d_blocks.CrossAttnUpBlock2D with CrossAttnUpBlock2D->CrossAttnUpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim +class CrossAttnUpBlockFlat(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: int | None = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: int | tuple[int] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlockFlat( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.Tensor, + res_hidden_states_tuple: tuple[torch.Tensor, ...], + temb: torch.Tensor | None = None, + encoder_hidden_states: torch.Tensor | None = None, + cross_attention_kwargs: dict[str, Any] | None = None, + upsample_size: int | None = None, + attention_mask: torch.Tensor | None = None, + encoder_attention_mask: torch.Tensor | None = None, + ) -> torch.Tensor: + if cross_attention_kwargs is not None: + if cross_attention_kwargs.get("scale", None) is not None: + logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") + + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func(resnet, hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + +# Copied from diffusers.models.unets.unet_2d_blocks.UNetMidBlock2D with UNetMidBlock2D->UNetMidBlockFlat, ResnetBlock2D->ResnetBlockFlat +class UNetMidBlockFlat(nn.Module): + """ + A 2D UNet mid-block [`UNetMidBlockFlat`] with multiple residual blocks and optional attention blocks. + + Args: + in_channels (`int`): The number of input channels. + temb_channels (`int`): The number of temporal embedding channels. + dropout (`float`, *optional*, defaults to 0.0): The dropout rate. + num_layers (`int`, *optional*, defaults to 1): The number of residual blocks. + resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks. + resnet_time_scale_shift (`str`, *optional*, defaults to `default`): + The type of normalization to apply to the time embeddings. This can help to improve the performance of the + model on tasks with long-range temporal dependencies. + resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks. + resnet_groups (`int`, *optional*, defaults to 32): + The number of groups to use in the group normalization layers of the resnet blocks. + attn_groups (`int | None`, *optional*, defaults to None): The number of groups for the attention blocks. + resnet_pre_norm (`bool`, *optional*, defaults to `True`): + Whether to use pre-normalization for the resnet blocks. + add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks. + attention_head_dim (`int`, *optional*, defaults to 1): + Dimension of a single attention head. The number of attention heads is determined based on this value and + the number of input channels. + output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor. + + Returns: + `torch.Tensor`: The output of the last residual block, which is a tensor of shape `(batch_size, in_channels, + height, width)`. + + """ + + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + attn_groups: int | None = None, + resnet_pre_norm: bool = True, + add_attention: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + ): + super().__init__() + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + self.add_attention = add_attention + + if attn_groups is None: + attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None + + # there is always at least one resnet + if resnet_time_scale_shift == "spatial": + resnets = [ + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ] + else: + resnets = [ + ResnetBlockFlat( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + if attention_head_dim is None: + logger.warning( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}." + ) + attention_head_dim = in_channels + + for _ in range(num_layers): + if self.add_attention: + attentions.append( + Attention( + in_channels, + heads=in_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=attn_groups, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + else: + attentions.append(None) + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlockFlat( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def forward(self, hidden_states: torch.Tensor, temb: torch.Tensor | None = None) -> torch.Tensor: + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if torch.is_grad_enabled() and self.gradient_checkpointing: + if attn is not None: + hidden_states = attn(hidden_states, temb=temb) + hidden_states = self._gradient_checkpointing_func(resnet, hidden_states, temb) + else: + if attn is not None: + hidden_states = attn(hidden_states, temb=temb) + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +# Copied from diffusers.models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn with UNetMidBlock2DCrossAttn->UNetMidBlockFlatCrossAttn, ResnetBlock2D->ResnetBlockFlat +class UNetMidBlockFlatCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + out_channels: int | None = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: int | tuple[int] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_groups_out: int | None = None, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + + out_channels = out_channels or in_channels + self.in_channels = in_channels + self.out_channels = out_channels + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # support for variable transformer layers per block + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + resnet_groups_out = resnet_groups_out or resnet_groups + + # there is always at least one resnet + resnets = [ + ResnetBlockFlat( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + groups_out=resnet_groups_out, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for i in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups_out, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + resnets.append( + ResnetBlockFlat( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups_out, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.Tensor, + temb: torch.Tensor | None = None, + encoder_hidden_states: torch.Tensor | None = None, + attention_mask: torch.Tensor | None = None, + cross_attention_kwargs: dict[str, Any] | None = None, + encoder_attention_mask: torch.Tensor | None = None, + ) -> torch.Tensor: + if cross_attention_kwargs is not None: + if cross_attention_kwargs.get("scale", None) is not None: + logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") + + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = self._gradient_checkpointing_func(resnet, hidden_states, temb) + else: + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +# Copied from diffusers.models.unets.unet_2d_blocks.UNetMidBlock2DSimpleCrossAttn with UNetMidBlock2DSimpleCrossAttn->UNetMidBlockFlatSimpleCrossAttn, ResnetBlock2D->ResnetBlockFlat +class UNetMidBlockFlatSimpleCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: str | None = None, + ): + super().__init__() + + self.has_cross_attention = True + + self.attention_head_dim = attention_head_dim + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + self.num_heads = in_channels // self.attention_head_dim + + # there is always at least one resnet + resnets = [ + ResnetBlockFlat( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ] + attentions = [] + + for _ in range(num_layers): + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=in_channels, + cross_attention_dim=in_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + resnets.append( + ResnetBlockFlat( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward( + self, + hidden_states: torch.Tensor, + temb: torch.Tensor | None = None, + encoder_hidden_states: torch.Tensor | None = None, + attention_mask: torch.Tensor | None = None, + cross_attention_kwargs: dict[str, Any] | None = None, + encoder_attention_mask: torch.Tensor | None = None, + ) -> torch.Tensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + if cross_attention_kwargs.get("scale", None) is not None: + logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + # attn + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + # resnet + hidden_states = resnet(hidden_states, temb) + + return hidden_states diff --git a/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion.py b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..101a1b72e7f95e4d28b68964f89d3e7566ccfe57 --- /dev/null +++ b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion.py @@ -0,0 +1,421 @@ +import inspect +from typing import Callable + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModel + +from ....models import AutoencoderKL, UNet2DConditionModel +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import logging +from ...pipeline_utils import DiffusionPipeline +from .pipeline_versatile_diffusion_dual_guided import VersatileDiffusionDualGuidedPipeline +from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline +from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + tokenizer: CLIPTokenizer + image_feature_extractor: CLIPImageProcessor + text_encoder: CLIPTextModel + image_encoder: CLIPVisionModel + image_unet: UNet2DConditionModel + text_unet: UNet2DConditionModel + vae: AutoencoderKL + scheduler: KarrasDiffusionSchedulers + + def __init__( + self, + tokenizer: CLIPTokenizer, + image_feature_extractor: CLIPImageProcessor, + text_encoder: CLIPTextModel, + image_encoder: CLIPVisionModel, + image_unet: UNet2DConditionModel, + text_unet: UNet2DConditionModel, + vae: AutoencoderKL, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, + image_feature_extractor=image_feature_extractor, + text_encoder=text_encoder, + image_encoder=image_encoder, + image_unet=image_unet, + text_unet=text_unet, + vae=vae, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + + @torch.no_grad() + def image_variation( + self, + image: torch.Tensor | PIL.Image.Image, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image`, `list[PIL.Image.Image]` or `torch.Tensor`): + The image prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + + >>> pipe = VersatileDiffusionPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe.image_variation(image, generator=generator).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + expected_components = inspect.signature(VersatileDiffusionImageVariationPipeline.__init__).parameters.keys() + components = {name: component for name, component in self.components.items() if name in expected_components} + return VersatileDiffusionImageVariationPipeline(**components)( + image=image, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + ) + + @torch.no_grad() + def text_to_image( + self, + prompt: str | list[str], + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. + height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionPipeline + >>> import torch + + >>> pipe = VersatileDiffusionPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe.text_to_image("an astronaut riding on a horse on mars", generator=generator).images[0] + >>> image.save("./astronaut.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + expected_components = inspect.signature(VersatileDiffusionTextToImagePipeline.__init__).parameters.keys() + components = {name: component for name, component in self.components.items() if name in expected_components} + temp_pipeline = VersatileDiffusionTextToImagePipeline(**components) + output = temp_pipeline( + prompt=prompt, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + ) + # swap the attention blocks back to the original state + temp_pipeline._swap_unet_attention_blocks() + + return output + + @torch.no_grad() + def dual_guided( + self, + prompt: PIL.Image.Image | list[PIL.Image.Image], + image: str | list[str], + text_to_image_strength: float = 0.5, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. + height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + >>> text = "a red car in the sun" + + >>> pipe = VersatileDiffusionPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> text_to_image_strength = 0.75 + + >>> image = pipe.dual_guided( + ... prompt=text, image=image, text_to_image_strength=text_to_image_strength, generator=generator + ... ).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + expected_components = inspect.signature(VersatileDiffusionDualGuidedPipeline.__init__).parameters.keys() + components = {name: component for name, component in self.components.items() if name in expected_components} + temp_pipeline = VersatileDiffusionDualGuidedPipeline(**components) + output = temp_pipeline( + prompt=prompt, + image=image, + text_to_image_strength=text_to_image_strength, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + ) + temp_pipeline._revert_dual_attention() + + return output diff --git a/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py new file mode 100644 index 0000000000000000000000000000000000000000..038169770d8dcb6c62a1335e7002460c7063e267 --- /dev/null +++ b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py @@ -0,0 +1,563 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import PIL.Image +import torch +import torch.utils.checkpoint +from transformers import ( + CLIPImageProcessor, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ....image_processor import VaeImageProcessor +from ....models import AutoencoderKL, DualTransformer2DModel, Transformer2DModel, UNet2DConditionModel +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import deprecate, logging +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from .modeling_text_unet import UNetFlatConditionModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline): + r""" + Pipeline for image-text dual-guided generation using Versatile Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) model to encode and decode images to and from latent representations. + bert ([`LDMBertModel`]): + Text-encoder model based on [`~transformers.BERT`]. + tokenizer ([`~transformers.BertTokenizer`]): + A `BertTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "bert->unet->vqvae" + + tokenizer: CLIPTokenizer + image_feature_extractor: CLIPImageProcessor + text_encoder: CLIPTextModelWithProjection + image_encoder: CLIPVisionModelWithProjection + image_unet: UNet2DConditionModel + text_unet: UNetFlatConditionModel + vae: AutoencoderKL + scheduler: KarrasDiffusionSchedulers + + _optional_components = ["text_unet"] + + def __init__( + self, + tokenizer: CLIPTokenizer, + image_feature_extractor: CLIPImageProcessor, + text_encoder: CLIPTextModelWithProjection, + image_encoder: CLIPVisionModelWithProjection, + image_unet: UNet2DConditionModel, + text_unet: UNetFlatConditionModel, + vae: AutoencoderKL, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + self.register_modules( + tokenizer=tokenizer, + image_feature_extractor=image_feature_extractor, + text_encoder=text_encoder, + image_encoder=image_encoder, + image_unet=image_unet, + text_unet=text_unet, + vae=vae, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + if self.text_unet is not None and ( + "dual_cross_attention" not in self.image_unet.config or not self.image_unet.config.dual_cross_attention + ): + # if loading from a universal checkpoint rather than a saved dual-guided pipeline + self._convert_to_dual_attention() + + def remove_unused_weights(self): + self.register_modules(text_unet=None) + + def _convert_to_dual_attention(self): + """ + Replace image_unet's `Transformer2DModel` blocks with `DualTransformer2DModel` that contains transformer blocks + from both `image_unet` and `text_unet` + """ + for name, module in self.image_unet.named_modules(): + if isinstance(module, Transformer2DModel): + parent_name, index = name.rsplit(".", 1) + index = int(index) + + image_transformer = self.image_unet.get_submodule(parent_name)[index] + text_transformer = self.text_unet.get_submodule(parent_name)[index] + + config = image_transformer.config + dual_transformer = DualTransformer2DModel( + num_attention_heads=config.num_attention_heads, + attention_head_dim=config.attention_head_dim, + in_channels=config.in_channels, + num_layers=config.num_layers, + dropout=config.dropout, + norm_num_groups=config.norm_num_groups, + cross_attention_dim=config.cross_attention_dim, + attention_bias=config.attention_bias, + sample_size=config.sample_size, + num_vector_embeds=config.num_vector_embeds, + activation_fn=config.activation_fn, + num_embeds_ada_norm=config.num_embeds_ada_norm, + ) + dual_transformer.transformers[0] = image_transformer + dual_transformer.transformers[1] = text_transformer + + self.image_unet.get_submodule(parent_name)[index] = dual_transformer + self.image_unet.register_to_config(dual_cross_attention=True) + + def _revert_dual_attention(self): + """ + Revert the image_unet `DualTransformer2DModel` blocks back to `Transformer2DModel` with image_unet weights Call + this function if you reuse `image_unet` in another pipeline, e.g. `VersatileDiffusionPipeline` + """ + for name, module in self.image_unet.named_modules(): + if isinstance(module, DualTransformer2DModel): + parent_name, index = name.rsplit(".", 1) + index = int(index) + self.image_unet.get_submodule(parent_name)[index] = module.transformers[0] + + self.image_unet.register_to_config(dual_cross_attention=False) + + def _encode_text_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + """ + + def normalize_embeddings(encoder_output): + embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state) + embeds_pooled = encoder_output.text_embeds + embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True) + return embeds + + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + prompt_embeds = normalize_embeddings(prompt_embeds) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def _encode_image_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + """ + + def normalize_embeddings(encoder_output): + embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state) + embeds = self.image_encoder.visual_projection(embeds) + embeds_pooled = embeds[:, 0:1] + embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True) + return embeds + + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + image_input = self.image_feature_extractor(images=prompt, return_tensors="pt") + pixel_values = image_input.pixel_values.to(device).to(self.image_encoder.dtype) + image_embeddings = self.image_encoder(pixel_values) + image_embeddings = normalize_embeddings(image_embeddings) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size + uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt") + pixel_values = uncond_images.pixel_values.to(device).to(self.image_encoder.dtype) + negative_prompt_embeds = self.image_encoder(pixel_values) + negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and conditional embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings]) + + return image_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs(self, prompt, image, height, width, callback_steps): + if not isinstance(prompt, str) and not isinstance(prompt, PIL.Image.Image) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` `PIL.Image` or `list` but is {type(prompt)}") + if not isinstance(image, str) and not isinstance(image, PIL.Image.Image) and not isinstance(image, list): + raise ValueError(f"`image` has to be of type `str` `PIL.Image` or `list` but is {type(image)}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def set_transformer_params(self, mix_ratio: float = 0.5, condition_types: tuple = ("text", "image")): + for name, module in self.image_unet.named_modules(): + if isinstance(module, DualTransformer2DModel): + module.mix_ratio = mix_ratio + + for i, type in enumerate(condition_types): + if type == "text": + module.condition_lengths[i] = self.text_encoder.config.max_position_embeddings + module.transformer_index_for_condition[i] = 1 # use the second (text) transformer + else: + module.condition_lengths[i] = 257 + module.transformer_index_for_condition[i] = 0 # use the first (image) transformer + + @torch.no_grad() + def __call__( + self, + prompt: PIL.Image.Image | list[PIL.Image.Image], + image: str | list[str], + text_to_image_strength: float = 0.5, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. + image (`PIL.Image.Image` or `list[PIL.Image.Image]`): + The image or images to condition the generation on alongside `prompt`. + text_to_image_strength (`float`, *optional*, defaults to 0.5): + Mixing ratio between the text and image conditioning. A value of 1.0 corresponds to pure text-to-image, + while 0.0 corresponds to pure image variation. + height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionDualGuidedPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + >>> text = "a red car in the sun" + + >>> pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe.remove_unused_weights() + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> text_to_image_strength = 0.75 + + >>> image = pipe( + ... prompt=text, image=image, text_to_image_strength=text_to_image_strength, generator=generator + ... ).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.image_unet.config.sample_size * self.vae_scale_factor + width = width or self.image_unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, image, height, width, callback_steps) + + # 2. Define call parameters + prompt = [prompt] if not isinstance(prompt, list) else prompt + image = [image] if not isinstance(image, list) else image + batch_size = len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompts + prompt_embeds = self._encode_text_prompt(prompt, device, num_images_per_prompt, do_classifier_free_guidance) + image_embeddings = self._encode_image_prompt(image, device, num_images_per_prompt, do_classifier_free_guidance) + dual_prompt_embeddings = torch.cat([prompt_embeds, image_embeddings], dim=1) + prompt_types = ("text", "image") + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.image_unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + dual_prompt_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Combine the attention blocks of the image and text UNets + self.set_transformer_params(text_to_image_strength, prompt_types) + + # 8. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=dual_prompt_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py new file mode 100644 index 0000000000000000000000000000000000000000..7481ab1c8cba53d7e233244f4b0d5e71a30fdd9a --- /dev/null +++ b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py @@ -0,0 +1,401 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ....image_processor import VaeImageProcessor +from ....models import AutoencoderKL, UNet2DConditionModel +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import deprecate, logging +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionImageVariationPipeline(DiffusionPipeline): + r""" + Pipeline for image variation using Versatile Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) model to encode and decode images to and from latent representations. + bert ([`LDMBertModel`]): + Text-encoder model based on [`~transformers.BERT`]. + tokenizer ([`~transformers.BertTokenizer`]): + A `BertTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "bert->unet->vqvae" + + image_feature_extractor: CLIPImageProcessor + image_encoder: CLIPVisionModelWithProjection + image_unet: UNet2DConditionModel + vae: AutoencoderKL + scheduler: KarrasDiffusionSchedulers + + def __init__( + self, + image_feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection, + image_unet: UNet2DConditionModel, + vae: AutoencoderKL, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + self.register_modules( + image_feature_extractor=image_feature_extractor, + image_encoder=image_encoder, + image_unet=image_unet, + vae=vae, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + + def normalize_embeddings(encoder_output): + embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state) + embeds = self.image_encoder.visual_projection(embeds) + embeds_pooled = embeds[:, 0:1] + embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True) + return embeds + + if isinstance(prompt, torch.Tensor) and len(prompt.shape) == 4: + prompt = list(prompt) + + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + image_input = self.image_feature_extractor(images=prompt, return_tensors="pt") + pixel_values = image_input.pixel_values.to(device).to(self.image_encoder.dtype) + image_embeddings = self.image_encoder(pixel_values) + image_embeddings = normalize_embeddings(image_embeddings) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_images: list[str] + if negative_prompt is None: + uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, PIL.Image.Image): + uncond_images = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_images = negative_prompt + + uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt") + pixel_values = uncond_images.pixel_values.to(device).to(self.image_encoder.dtype) + negative_prompt_embeds = self.image_encoder(pixel_values) + negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and conditional embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings]) + + return image_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_image_variation.StableDiffusionImageVariationPipeline.check_inputs + def check_inputs(self, image, height, width, callback_steps): + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `list[PIL.Image.Image]` but is" + f" {type(image)}" + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + image: PIL.Image.Image | list[PIL.Image.Image] | torch.Tensor, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image`, `list[PIL.Image.Image]` or `torch.Tensor`): + The image prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionImageVariationPipeline + >>> import torch + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + + >>> # let's download an initial image + >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg" + + >>> response = requests.get(url) + >>> image = Image.open(BytesIO(response.content)).convert("RGB") + + >>> pipe = VersatileDiffusionImageVariationPipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe(image, generator=generator).images[0] + >>> image.save("./car_variation.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.image_unet.config.sample_size * self.vae_scale_factor + width = width or self.image_unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(image, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(image, PIL.Image.Image) else len(image) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + image_embeddings = self._encode_prompt( + image, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.image_unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + image_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py new file mode 100644 index 0000000000000000000000000000000000000000..5ee5758fac8aecd12c3a88d9eaa125f070cbd621 --- /dev/null +++ b/diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py @@ -0,0 +1,479 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import torch +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer + +from ....image_processor import VaeImageProcessor +from ....models import AutoencoderKL, Transformer2DModel, UNet2DConditionModel +from ....schedulers import KarrasDiffusionSchedulers +from ....utils import deprecate, logging +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from .modeling_text_unet import UNetFlatConditionModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class VersatileDiffusionTextToImagePipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using Versatile Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) model to encode and decode images to and from latent representations. + bert ([`LDMBertModel`]): + Text-encoder model based on [`~transformers.BERT`]. + tokenizer ([`~transformers.BertTokenizer`]): + A `BertTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "bert->unet->vqvae" + + tokenizer: CLIPTokenizer + image_feature_extractor: CLIPImageProcessor + text_encoder: CLIPTextModelWithProjection + image_unet: UNet2DConditionModel + text_unet: UNetFlatConditionModel + vae: AutoencoderKL + scheduler: KarrasDiffusionSchedulers + + _optional_components = ["text_unet"] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModelWithProjection, + image_unet: UNet2DConditionModel, + text_unet: UNetFlatConditionModel, + vae: AutoencoderKL, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + image_unet=image_unet, + text_unet=text_unet, + vae=vae, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + if self.text_unet is not None: + self._swap_unet_attention_blocks() + + def _swap_unet_attention_blocks(self): + """ + Swap the `Transformer2DModel` blocks between the image and text UNets + """ + for name, module in self.image_unet.named_modules(): + if isinstance(module, Transformer2DModel): + parent_name, index = name.rsplit(".", 1) + index = int(index) + self.image_unet.get_submodule(parent_name)[index], self.text_unet.get_submodule(parent_name)[index] = ( + self.text_unet.get_submodule(parent_name)[index], + self.image_unet.get_submodule(parent_name)[index], + ) + + def remove_unused_weights(self): + self.register_modules(text_unet=None) + + def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + + def normalize_embeddings(encoder_output): + embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state) + embeds_pooled = encoder_output.text_embeds + embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True) + return embeds + + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + prompt_embeds = normalize_embeddings(prompt_embeds) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. + height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + + ```py + >>> from diffusers import VersatileDiffusionTextToImagePipeline + >>> import torch + + >>> pipe = VersatileDiffusionTextToImagePipeline.from_pretrained( + ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16 + ... ) + >>> pipe.remove_unused_weights() + >>> pipe = pipe.to("cuda") + + >>> generator = torch.Generator(device="cuda").manual_seed(0) + >>> image = pipe("an astronaut riding on a horse on mars", generator=generator).images[0] + >>> image.save("./astronaut.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.image_unet.config.sample_size * self.vae_scale_factor + width = width or self.image_unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.image_unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=prompt_embeds).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/deprecated/vq_diffusion/__init__.py b/diffusers/pipelines/deprecated/vq_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..070903377c7188415af0417d4839d74a8a34dc01 --- /dev/null +++ b/diffusers/pipelines/deprecated/vq_diffusion/__init__.py @@ -0,0 +1,57 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import ( + LearnedClassifierFreeSamplingEmbeddings, + VQDiffusionPipeline, + ) + + _dummy_objects.update( + { + "LearnedClassifierFreeSamplingEmbeddings": LearnedClassifierFreeSamplingEmbeddings, + "VQDiffusionPipeline": VQDiffusionPipeline, + } + ) +else: + _import_structure["pipeline_vq_diffusion"] = ["LearnedClassifierFreeSamplingEmbeddings", "VQDiffusionPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import ( + LearnedClassifierFreeSamplingEmbeddings, + VQDiffusionPipeline, + ) + else: + from .pipeline_vq_diffusion import LearnedClassifierFreeSamplingEmbeddings, VQDiffusionPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/vq_diffusion/pipeline_vq_diffusion.py b/diffusers/pipelines/deprecated/vq_diffusion/pipeline_vq_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..8f3b9512e888fce4ba9b639987f74da7a63bb755 --- /dev/null +++ b/diffusers/pipelines/deprecated/vq_diffusion/pipeline_vq_diffusion.py @@ -0,0 +1,325 @@ +# Copyright 2025 Microsoft and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models import ModelMixin, Transformer2DModel, VQModel +from ....schedulers import VQDiffusionScheduler +from ....utils import logging +from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class LearnedClassifierFreeSamplingEmbeddings(ModelMixin, ConfigMixin): + """ + Utility class for storing learned text embeddings for classifier free sampling + """ + + @register_to_config + def __init__(self, learnable: bool, hidden_size: int | None = None, length: int | None = None): + super().__init__() + + self.learnable = learnable + + if self.learnable: + assert hidden_size is not None, "learnable=True requires `hidden_size` to be set" + assert length is not None, "learnable=True requires `length` to be set" + + embeddings = torch.zeros(length, hidden_size) + else: + embeddings = None + + self.embeddings = torch.nn.Parameter(embeddings) + + +class VQDiffusionPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using VQ Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vqvae ([`VQModel`]): + Vector Quantized Variational Auto-Encoder (VAE) model to encode and decode images to and from latent + representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + transformer ([`Transformer2DModel`]): + A conditional `Transformer2DModel` to denoise the encoded image latents. + scheduler ([`VQDiffusionScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + """ + + vqvae: VQModel + text_encoder: CLIPTextModel + tokenizer: CLIPTokenizer + transformer: Transformer2DModel + learned_classifier_free_sampling_embeddings: LearnedClassifierFreeSamplingEmbeddings + scheduler: VQDiffusionScheduler + + def __init__( + self, + vqvae: VQModel, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + transformer: Transformer2DModel, + scheduler: VQDiffusionScheduler, + learned_classifier_free_sampling_embeddings: LearnedClassifierFreeSamplingEmbeddings, + ): + super().__init__() + + self.register_modules( + vqvae=vqvae, + transformer=transformer, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + learned_classifier_free_sampling_embeddings=learned_classifier_free_sampling_embeddings, + ) + + def _encode_prompt(self, prompt, num_images_per_prompt, do_classifier_free_guidance): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + if text_input_ids.shape[-1] > self.tokenizer.model_max_length: + removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + prompt_embeds = self.text_encoder(text_input_ids.to(self.device))[0] + + # NOTE: This additional step of normalizing the text embeddings is from VQ-Diffusion. + # While CLIP does normalize the pooled output of the text transformer when combining + # the image and text embeddings, CLIP does not directly normalize the last hidden state. + # + # CLIP normalizing the pooled output. + # https://github.com/huggingface/transformers/blob/d92e22d1f28324f513f3080e5c47c071a3916721/src/transformers/models/clip/modeling_clip.py#L1052-L1053 + prompt_embeds = prompt_embeds / prompt_embeds.norm(dim=-1, keepdim=True) + + # duplicate text embeddings for each generation per prompt + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + if self.learned_classifier_free_sampling_embeddings.learnable: + negative_prompt_embeds = self.learned_classifier_free_sampling_embeddings.embeddings + negative_prompt_embeds = negative_prompt_embeds.unsqueeze(0).repeat(batch_size, 1, 1) + else: + uncond_tokens = [""] * batch_size + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + negative_prompt_embeds = self.text_encoder(uncond_input.input_ids.to(self.device))[0] + # See comment for normalizing text embeddings + negative_prompt_embeds = negative_prompt_embeds / negative_prompt_embeds.norm(dim=-1, keepdim=True) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + num_inference_steps: int = 100, + guidance_scale: float = 5.0, + truncation_rate: float = 1.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ) -> ImagePipelineOutput | tuple: + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image generation. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + truncation_rate (`float`, *optional*, defaults to 1.0 (equivalent to no truncation)): + Used to "truncate" the predicted classes for x_0 such that the cumulative probability for a pixel is at + most `truncation_rate`. The lowest probabilities that would increase the cumulative probability above + `truncation_rate` are set to zero. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor` of shape (batch), *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Must be valid embedding indices.If not provided, a latents tensor will be generated of + completely masked latent pixels. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt(prompt, num_images_per_prompt, do_classifier_free_guidance) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # get the initial completely masked latents unless the user supplied it + + latents_shape = (batch_size, self.transformer.num_latent_pixels) + if latents is None: + mask_class = self.transformer.num_vector_embeds - 1 + latents = torch.full(latents_shape, mask_class).to(self.device) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + if (latents < 0).any() or (latents >= self.transformer.num_vector_embeds).any(): + raise ValueError( + "Unexpected latents value(s). All latents be valid embedding indices i.e. in the range 0," + f" {self.transformer.num_vector_embeds - 1} (inclusive)." + ) + latents = latents.to(self.device) + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=self.device) + + timesteps_tensor = self.scheduler.timesteps.to(self.device) + + sample = latents + + for i, t in enumerate(self.progress_bar(timesteps_tensor)): + # expand the sample if we are doing classifier free guidance + latent_model_input = torch.cat([sample] * 2) if do_classifier_free_guidance else sample + + # predict the un-noised image + # model_output == `log_p_x_0` + model_output = self.transformer(latent_model_input, encoder_hidden_states=prompt_embeds, timestep=t).sample + + if do_classifier_free_guidance: + model_output_uncond, model_output_text = model_output.chunk(2) + model_output = model_output_uncond + guidance_scale * (model_output_text - model_output_uncond) + model_output -= torch.logsumexp(model_output, dim=1, keepdim=True) + + model_output = self.truncate(model_output, truncation_rate) + + # remove `log(0)`'s (`-inf`s) + model_output = model_output.clamp(-70) + + # compute the previous noisy sample x_t -> x_t-1 + sample = self.scheduler.step(model_output, timestep=t, sample=sample, generator=generator).prev_sample + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + callback(i, t, sample) + + embedding_channels = self.vqvae.config.vq_embed_dim + embeddings_shape = (batch_size, self.transformer.height, self.transformer.width, embedding_channels) + embeddings = self.vqvae.quantize.get_codebook_entry(sample, shape=embeddings_shape) + image = self.vqvae.decode(embeddings, force_not_quantize=True).sample + + image = (image / 2 + 0.5).clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) + + def truncate(self, log_p_x_0: torch.Tensor, truncation_rate: float) -> torch.Tensor: + """ + Truncates `log_p_x_0` such that for each column vector, the total cumulative probability is `truncation_rate` + The lowest probabilities that would increase the cumulative probability above `truncation_rate` are set to + zero. + """ + sorted_log_p_x_0, indices = torch.sort(log_p_x_0, 1, descending=True) + sorted_p_x_0 = torch.exp(sorted_log_p_x_0) + keep_mask = sorted_p_x_0.cumsum(dim=1) < truncation_rate + + # Ensure that at least the largest probability is not zeroed out + all_true = torch.full_like(keep_mask[:, 0:1, :], True) + keep_mask = torch.cat((all_true, keep_mask), dim=1) + keep_mask = keep_mask[:, :-1, :] + + keep_mask = keep_mask.gather(1, indices.argsort(1)) + + rv = log_p_x_0.clone() + + rv[~keep_mask] = -torch.inf # -inf = log(0) + + return rv diff --git a/diffusers/pipelines/deprecated/wuerstchen/__init__.py b/diffusers/pipelines/deprecated/wuerstchen/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..26f259512cbd146dbb27419d5df08db109a9499c --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/__init__.py @@ -0,0 +1,56 @@ +from typing import TYPE_CHECKING + +from ....utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ....utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["modeling_paella_vq_model"] = ["PaellaVQModel"] + _import_structure["modeling_wuerstchen_diffnext"] = ["WuerstchenDiffNeXt"] + _import_structure["modeling_wuerstchen_prior"] = ["WuerstchenPrior"] + _import_structure["pipeline_wuerstchen"] = ["WuerstchenDecoderPipeline"] + _import_structure["pipeline_wuerstchen_combined"] = ["WuerstchenCombinedPipeline"] + _import_structure["pipeline_wuerstchen_prior"] = ["DEFAULT_STAGE_C_TIMESTEPS", "WuerstchenPriorPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ....utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .modeling_paella_vq_model import PaellaVQModel + from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt + from .modeling_wuerstchen_prior import WuerstchenPrior + from .pipeline_wuerstchen import WuerstchenDecoderPipeline + from .pipeline_wuerstchen_combined import WuerstchenCombinedPipeline + from .pipeline_wuerstchen_prior import DEFAULT_STAGE_C_TIMESTEPS, WuerstchenPriorPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/deprecated/wuerstchen/modeling_paella_vq_model.py b/diffusers/pipelines/deprecated/wuerstchen/modeling_paella_vq_model.py new file mode 100644 index 0000000000000000000000000000000000000000..a18ab8eed6eca86cc0466690d9a71997ee926585 --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/modeling_paella_vq_model.py @@ -0,0 +1,171 @@ +# Copyright (c) 2022 Dominic Rampas MIT License +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import torch +import torch.nn as nn + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models.autoencoders.vae import DecoderOutput, VectorQuantizer +from ....models.modeling_utils import ModelMixin +from ....models.vq_model import VQEncoderOutput +from ....utils.accelerate_utils import apply_forward_hook + + +class MixingResidualBlock(nn.Module): + """ + Residual block with mixing used by Paella's VQ-VAE. + """ + + def __init__(self, inp_channels, embed_dim): + super().__init__() + # depthwise + self.norm1 = nn.LayerNorm(inp_channels, elementwise_affine=False, eps=1e-6) + self.depthwise = nn.Sequential( + nn.ReplicationPad2d(1), nn.Conv2d(inp_channels, inp_channels, kernel_size=3, groups=inp_channels) + ) + + # channelwise + self.norm2 = nn.LayerNorm(inp_channels, elementwise_affine=False, eps=1e-6) + self.channelwise = nn.Sequential( + nn.Linear(inp_channels, embed_dim), nn.GELU(), nn.Linear(embed_dim, inp_channels) + ) + + self.gammas = nn.Parameter(torch.zeros(6), requires_grad=True) + + def forward(self, x): + mods = self.gammas + x_temp = self.norm1(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * (1 + mods[0]) + mods[1] + x = x + self.depthwise(x_temp) * mods[2] + x_temp = self.norm2(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * (1 + mods[3]) + mods[4] + x = x + self.channelwise(x_temp.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * mods[5] + return x + + +class PaellaVQModel(ModelMixin, ConfigMixin): + r"""VQ-VAE model from Paella model. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library + implements for all the model (such as downloading or saving, etc.) + + Parameters: + in_channels (int, *optional*, defaults to 3): Number of channels in the input image. + out_channels (int, *optional*, defaults to 3): Number of channels in the output. + up_down_scale_factor (int, *optional*, defaults to 2): Up and Downscale factor of the input image. + levels (int, *optional*, defaults to 2): Number of levels in the model. + bottleneck_blocks (int, *optional*, defaults to 12): Number of bottleneck blocks in the model. + embed_dim (int, *optional*, defaults to 384): Number of hidden channels in the model. + latent_channels (int, *optional*, defaults to 4): Number of latent channels in the VQ-VAE model. + num_vq_embeddings (int, *optional*, defaults to 8192): Number of codebook vectors in the VQ-VAE. + scale_factor (float, *optional*, defaults to 0.3764): Scaling factor of the latent space. + """ + + @register_to_config + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + up_down_scale_factor: int = 2, + levels: int = 2, + bottleneck_blocks: int = 12, + embed_dim: int = 384, + latent_channels: int = 4, + num_vq_embeddings: int = 8192, + scale_factor: float = 0.3764, + ): + super().__init__() + + c_levels = [embed_dim // (2**i) for i in reversed(range(levels))] + # Encoder blocks + self.in_block = nn.Sequential( + nn.PixelUnshuffle(up_down_scale_factor), + nn.Conv2d(in_channels * up_down_scale_factor**2, c_levels[0], kernel_size=1), + ) + down_blocks = [] + for i in range(levels): + if i > 0: + down_blocks.append(nn.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1)) + block = MixingResidualBlock(c_levels[i], c_levels[i] * 4) + down_blocks.append(block) + down_blocks.append( + nn.Sequential( + nn.Conv2d(c_levels[-1], latent_channels, kernel_size=1, bias=False), + nn.BatchNorm2d(latent_channels), # then normalize them to have mean 0 and std 1 + ) + ) + self.down_blocks = nn.Sequential(*down_blocks) + + # Vector Quantizer + self.vquantizer = VectorQuantizer(num_vq_embeddings, vq_embed_dim=latent_channels, legacy=False, beta=0.25) + + # Decoder blocks + up_blocks = [nn.Sequential(nn.Conv2d(latent_channels, c_levels[-1], kernel_size=1))] + for i in range(levels): + for j in range(bottleneck_blocks if i == 0 else 1): + block = MixingResidualBlock(c_levels[levels - 1 - i], c_levels[levels - 1 - i] * 4) + up_blocks.append(block) + if i < levels - 1: + up_blocks.append( + nn.ConvTranspose2d( + c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2, padding=1 + ) + ) + self.up_blocks = nn.Sequential(*up_blocks) + self.out_block = nn.Sequential( + nn.Conv2d(c_levels[0], out_channels * up_down_scale_factor**2, kernel_size=1), + nn.PixelShuffle(up_down_scale_factor), + ) + + @apply_forward_hook + def encode(self, x: torch.Tensor, return_dict: bool = True) -> VQEncoderOutput: + h = self.in_block(x) + h = self.down_blocks(h) + + if not return_dict: + return (h,) + + return VQEncoderOutput(latents=h) + + @apply_forward_hook + def decode( + self, h: torch.Tensor, force_not_quantize: bool = True, return_dict: bool = True + ) -> DecoderOutput | torch.Tensor: + if not force_not_quantize: + quant, _, _ = self.vquantizer(h) + else: + quant = h + + x = self.up_blocks(quant) + dec = self.out_block(x) + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + def forward(self, sample: torch.Tensor, return_dict: bool = True) -> DecoderOutput | torch.Tensor: + r""" + Args: + sample (`torch.Tensor`): Input sample. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`DecoderOutput`] instead of a plain tuple. + """ + x = sample + h = self.encode(x).latents + dec = self.decode(h).sample + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) diff --git a/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_common.py b/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_common.py new file mode 100644 index 0000000000000000000000000000000000000000..7645a5579c5861cef203df6816da8534a879231d --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_common.py @@ -0,0 +1,76 @@ +import torch +import torch.nn as nn + +from ....models.attention_processor import Attention + + +class WuerstchenLayerNorm(nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, x): + x = x.permute(0, 2, 3, 1) + x = super().forward(x) + return x.permute(0, 3, 1, 2) + + +class TimestepBlock(nn.Module): + def __init__(self, c, c_timestep): + super().__init__() + + self.mapper = nn.Linear(c_timestep, c * 2) + + def forward(self, x, t): + a, b = self.mapper(t)[:, :, None, None].chunk(2, dim=1) + return x * (1 + a) + b + + +class ResBlock(nn.Module): + def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0): + super().__init__() + + self.depthwise = nn.Conv2d(c + c_skip, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c) + self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6) + self.channelwise = nn.Sequential( + nn.Linear(c, c * 4), nn.GELU(), GlobalResponseNorm(c * 4), nn.Dropout(dropout), nn.Linear(c * 4, c) + ) + + def forward(self, x, x_skip=None): + x_res = x + if x_skip is not None: + x = torch.cat([x, x_skip], dim=1) + x = self.norm(self.depthwise(x)).permute(0, 2, 3, 1) + x = self.channelwise(x).permute(0, 3, 1, 2) + return x + x_res + + +# from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105 +class GlobalResponseNorm(nn.Module): + def __init__(self, dim): + super().__init__() + self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim)) + self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim)) + + def forward(self, x): + agg_norm = torch.norm(x, p=2, dim=(1, 2), keepdim=True) + stand_div_norm = agg_norm / (agg_norm.mean(dim=-1, keepdim=True) + 1e-6) + return self.gamma * (x * stand_div_norm) + self.beta + x + + +class AttnBlock(nn.Module): + def __init__(self, c, c_cond, nhead, self_attn=True, dropout=0.0): + super().__init__() + + self.self_attn = self_attn + self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6) + self.attention = Attention(query_dim=c, heads=nhead, dim_head=c // nhead, dropout=dropout, bias=True) + self.kv_mapper = nn.Sequential(nn.SiLU(), nn.Linear(c_cond, c)) + + def forward(self, x, kv): + kv = self.kv_mapper(kv) + norm_x = self.norm(x) + if self.self_attn: + batch_size, channel, _, _ = x.shape + kv = torch.cat([norm_x.view(batch_size, channel, -1).transpose(1, 2), kv], dim=1) + x = x + self.attention(norm_x, encoder_hidden_states=kv) + return x diff --git a/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_diffnext.py b/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_diffnext.py new file mode 100644 index 0000000000000000000000000000000000000000..5c6c12264da9b2cf146766c5e92a342491b26477 --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_diffnext.py @@ -0,0 +1,254 @@ +# Copyright (c) 2023 Dominic Rampas MIT License +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math + +import numpy as np +import torch +import torch.nn as nn + +from ....configuration_utils import ConfigMixin, register_to_config +from ....models.modeling_utils import ModelMixin +from .modeling_wuerstchen_common import AttnBlock, GlobalResponseNorm, TimestepBlock, WuerstchenLayerNorm + + +class WuerstchenDiffNeXt(ModelMixin, ConfigMixin): + @register_to_config + def __init__( + self, + c_in=4, + c_out=4, + c_r=64, + patch_size=2, + c_cond=1024, + c_hidden=[320, 640, 1280, 1280], + nhead=[-1, 10, 20, 20], + blocks=[4, 4, 14, 4], + level_config=["CT", "CTA", "CTA", "CTA"], + inject_effnet=[False, True, True, True], + effnet_embd=16, + clip_embd=1024, + kernel_size=3, + dropout=0.1, + ): + super().__init__() + self.c_r = c_r + self.c_cond = c_cond + if not isinstance(dropout, list): + dropout = [dropout] * len(c_hidden) + + # CONDITIONING + self.clip_mapper = nn.Linear(clip_embd, c_cond) + self.effnet_mappers = nn.ModuleList( + [ + nn.Conv2d(effnet_embd, c_cond, kernel_size=1) if inject else None + for inject in inject_effnet + list(reversed(inject_effnet)) + ] + ) + self.seq_norm = nn.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6) + + self.embedding = nn.Sequential( + nn.PixelUnshuffle(patch_size), + nn.Conv2d(c_in * (patch_size**2), c_hidden[0], kernel_size=1), + WuerstchenLayerNorm(c_hidden[0], elementwise_affine=False, eps=1e-6), + ) + + def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0): + if block_type == "C": + return ResBlockStageB(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout) + elif block_type == "A": + return AttnBlock(c_hidden, c_cond, nhead, self_attn=True, dropout=dropout) + elif block_type == "T": + return TimestepBlock(c_hidden, c_r) + else: + raise ValueError(f"Block type {block_type} not supported") + + # BLOCKS + # -- down blocks + self.down_blocks = nn.ModuleList() + for i in range(len(c_hidden)): + down_block = nn.ModuleList() + if i > 0: + down_block.append( + nn.Sequential( + WuerstchenLayerNorm(c_hidden[i - 1], elementwise_affine=False, eps=1e-6), + nn.Conv2d(c_hidden[i - 1], c_hidden[i], kernel_size=2, stride=2), + ) + ) + for _ in range(blocks[i]): + for block_type in level_config[i]: + c_skip = c_cond if inject_effnet[i] else 0 + down_block.append(get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i])) + self.down_blocks.append(down_block) + + # -- up blocks + self.up_blocks = nn.ModuleList() + for i in reversed(range(len(c_hidden))): + up_block = nn.ModuleList() + for j in range(blocks[i]): + for k, block_type in enumerate(level_config[i]): + c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0 + c_skip += c_cond if inject_effnet[i] else 0 + up_block.append(get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i])) + if i > 0: + up_block.append( + nn.Sequential( + WuerstchenLayerNorm(c_hidden[i], elementwise_affine=False, eps=1e-6), + nn.ConvTranspose2d(c_hidden[i], c_hidden[i - 1], kernel_size=2, stride=2), + ) + ) + self.up_blocks.append(up_block) + + # OUTPUT + self.clf = nn.Sequential( + WuerstchenLayerNorm(c_hidden[0], elementwise_affine=False, eps=1e-6), + nn.Conv2d(c_hidden[0], 2 * c_out * (patch_size**2), kernel_size=1), + nn.PixelShuffle(patch_size), + ) + + # --- WEIGHT INIT --- + self.apply(self._init_weights) + + def _init_weights(self, m): + # General init + if isinstance(m, (nn.Conv2d, nn.Linear)): + nn.init.xavier_uniform_(m.weight) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + for mapper in self.effnet_mappers: + if mapper is not None: + nn.init.normal_(mapper.weight, std=0.02) # conditionings + nn.init.normal_(self.clip_mapper.weight, std=0.02) # conditionings + nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs + nn.init.constant_(self.clf[1].weight, 0) # outputs + + # blocks + for level_block in self.down_blocks + self.up_blocks: + for block in level_block: + if isinstance(block, ResBlockStageB): + block.channelwise[-1].weight.data *= np.sqrt(1 / sum(self.config.blocks)) + elif isinstance(block, TimestepBlock): + nn.init.constant_(block.mapper.weight, 0) + + def gen_r_embedding(self, r, max_positions=10000): + r = r * max_positions + half_dim = self.c_r // 2 + emb = math.log(max_positions) / (half_dim - 1) + emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp() + emb = r[:, None] * emb[None, :] + emb = torch.cat([emb.sin(), emb.cos()], dim=1) + if self.c_r % 2 == 1: # zero pad + emb = nn.functional.pad(emb, (0, 1), mode="constant") + return emb.to(dtype=r.dtype) + + def gen_c_embeddings(self, clip): + clip = self.clip_mapper(clip) + clip = self.seq_norm(clip) + return clip + + def _down_encode(self, x, r_embed, effnet, clip=None): + level_outputs = [] + for i, down_block in enumerate(self.down_blocks): + effnet_c = None + for block in down_block: + if isinstance(block, ResBlockStageB): + if effnet_c is None and self.effnet_mappers[i] is not None: + dtype = effnet.dtype + effnet_c = self.effnet_mappers[i]( + nn.functional.interpolate( + effnet.float(), size=x.shape[-2:], mode="bicubic", antialias=True, align_corners=True + ).to(dtype) + ) + skip = effnet_c if self.effnet_mappers[i] is not None else None + x = block(x, skip) + elif isinstance(block, AttnBlock): + x = block(x, clip) + elif isinstance(block, TimestepBlock): + x = block(x, r_embed) + else: + x = block(x) + level_outputs.insert(0, x) + return level_outputs + + def _up_decode(self, level_outputs, r_embed, effnet, clip=None): + x = level_outputs[0] + for i, up_block in enumerate(self.up_blocks): + effnet_c = None + for j, block in enumerate(up_block): + if isinstance(block, ResBlockStageB): + if effnet_c is None and self.effnet_mappers[len(self.down_blocks) + i] is not None: + dtype = effnet.dtype + effnet_c = self.effnet_mappers[len(self.down_blocks) + i]( + nn.functional.interpolate( + effnet.float(), size=x.shape[-2:], mode="bicubic", antialias=True, align_corners=True + ).to(dtype) + ) + skip = level_outputs[i] if j == 0 and i > 0 else None + if effnet_c is not None: + if skip is not None: + skip = torch.cat([skip, effnet_c], dim=1) + else: + skip = effnet_c + x = block(x, skip) + elif isinstance(block, AttnBlock): + x = block(x, clip) + elif isinstance(block, TimestepBlock): + x = block(x, r_embed) + else: + x = block(x) + return x + + def forward(self, x, r, effnet, clip=None, x_cat=None, eps=1e-3, return_noise=True): + if x_cat is not None: + x = torch.cat([x, x_cat], dim=1) + # Process the conditioning embeddings + r_embed = self.gen_r_embedding(r) + if clip is not None: + clip = self.gen_c_embeddings(clip) + + # Model Blocks + x_in = x + x = self.embedding(x) + level_outputs = self._down_encode(x, r_embed, effnet, clip) + x = self._up_decode(level_outputs, r_embed, effnet, clip) + a, b = self.clf(x).chunk(2, dim=1) + b = b.sigmoid() * (1 - eps * 2) + eps + if return_noise: + return (x_in - a) / b + else: + return a, b + + +class ResBlockStageB(nn.Module): + def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0): + super().__init__() + self.depthwise = nn.Conv2d(c, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c) + self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6) + self.channelwise = nn.Sequential( + nn.Linear(c + c_skip, c * 4), + nn.GELU(), + GlobalResponseNorm(c * 4), + nn.Dropout(dropout), + nn.Linear(c * 4, c), + ) + + def forward(self, x, x_skip=None): + x_res = x + x = self.norm(self.depthwise(x)) + if x_skip is not None: + x = torch.cat([x, x_skip], dim=1) + x = self.channelwise(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + return x + x_res diff --git a/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_prior.py b/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_prior.py new file mode 100644 index 0000000000000000000000000000000000000000..f19eec46ec3be9e4ca1133f4f45477bb46e86b61 --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/modeling_wuerstchen_prior.py @@ -0,0 +1,113 @@ +# Copyright (c) 2023 Dominic Rampas MIT License +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math + +import torch +import torch.nn as nn + +from ....configuration_utils import ConfigMixin, register_to_config +from ....loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin +from ....models.attention import AttentionMixin +from ....models.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + AttnAddedKVProcessor, + AttnProcessor, +) +from ....models.modeling_utils import ModelMixin +from .modeling_wuerstchen_common import AttnBlock, ResBlock, TimestepBlock, WuerstchenLayerNorm + + +class WuerstchenPrior(ModelMixin, AttentionMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin): + unet_name = "prior" + _supports_gradient_checkpointing = True + + @register_to_config + def __init__(self, c_in=16, c=1280, c_cond=1024, c_r=64, depth=16, nhead=16, dropout=0.1): + super().__init__() + + self.c_r = c_r + self.projection = nn.Conv2d(c_in, c, kernel_size=1) + self.cond_mapper = nn.Sequential( + nn.Linear(c_cond, c), + nn.LeakyReLU(0.2), + nn.Linear(c, c), + ) + + self.blocks = nn.ModuleList() + for _ in range(depth): + self.blocks.append(ResBlock(c, dropout=dropout)) + self.blocks.append(TimestepBlock(c, c_r)) + self.blocks.append(AttnBlock(c, c, nhead, self_attn=True, dropout=dropout)) + self.out = nn.Sequential( + WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6), + nn.Conv2d(c, c_in * 2, kernel_size=1), + ) + + self.gradient_checkpointing = False + self.set_default_attn_processor() + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + def gen_r_embedding(self, r, max_positions=10000): + r = r * max_positions + half_dim = self.c_r // 2 + emb = math.log(max_positions) / (half_dim - 1) + emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp() + emb = r[:, None] * emb[None, :] + emb = torch.cat([emb.sin(), emb.cos()], dim=1) + if self.c_r % 2 == 1: # zero pad + emb = nn.functional.pad(emb, (0, 1), mode="constant") + return emb.to(dtype=r.dtype) + + def forward(self, x, r, c): + x_in = x + x = self.projection(x) + c_embed = self.cond_mapper(c) + r_embed = self.gen_r_embedding(r) + + if torch.is_grad_enabled() and self.gradient_checkpointing: + for block in self.blocks: + if isinstance(block, AttnBlock): + x = self._gradient_checkpointing_func(block, x, c_embed) + elif isinstance(block, TimestepBlock): + x = self._gradient_checkpointing_func(block, x, r_embed) + else: + x = self._gradient_checkpointing_func(block, x) + else: + for block in self.blocks: + if isinstance(block, AttnBlock): + x = block(x, c_embed) + elif isinstance(block, TimestepBlock): + x = block(x, r_embed) + else: + x = block(x) + a, b = self.out(x).chunk(2, dim=1) + return (x_in - a) / ((1 - b).abs() + 1e-5) diff --git a/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen.py b/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen.py new file mode 100644 index 0000000000000000000000000000000000000000..aeef16bc45e0d3cebb0a42dde3783cc72d235cf0 --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen.py @@ -0,0 +1,449 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ....schedulers import DDPMWuerstchenScheduler +from ....utils import deprecate, is_torch_xla_available, logging, replace_example_docstring +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput +from .modeling_paella_vq_model import PaellaVQModel +from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import WuerstchenPriorPipeline, WuerstchenDecoderPipeline + + >>> prior_pipe = WuerstchenPriorPipeline.from_pretrained( + ... "warp-ai/wuerstchen-prior", torch_dtype=torch.float16 + ... ).to("cuda") + >>> gen_pipe = WuerstchenDecoderPipeline.from_pretrain("warp-ai/wuerstchen", torch_dtype=torch.float16).to( + ... "cuda" + ... ) + + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> prior_output = pipe(prompt) + >>> images = gen_pipe(prior_output.image_embeddings, prompt=prompt) + ``` +""" + + +class WuerstchenDecoderPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline for generating images from the Wuerstchen model. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + tokenizer (`CLIPTokenizer`): + The CLIP tokenizer. + text_encoder (`CLIPTextModel`): + The CLIP text encoder. + decoder ([`WuerstchenDiffNeXt`]): + The WuerstchenDiffNeXt unet decoder. + vqgan ([`PaellaVQModel`]): + The VQGAN model. + scheduler ([`DDPMWuerstchenScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + latent_dim_scale (float, `optional`, defaults to 10.67): + Multiplier to determine the VQ latent space size from the image embeddings. If the image embeddings are + height=24 and width=24, the VQ latent shape needs to be height=int(24*10.67)=256 and + width=int(24*10.67)=256 in order to match the training conditions. + """ + + model_cpu_offload_seq = "text_encoder->decoder->vqgan" + _callback_tensor_inputs = [ + "latents", + "text_encoder_hidden_states", + "negative_prompt_embeds", + "image_embeddings", + ] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + decoder: WuerstchenDiffNeXt, + scheduler: DDPMWuerstchenScheduler, + vqgan: PaellaVQModel, + latent_dim_scale: float = 10.67, + ) -> None: + super().__init__() + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + ) + self.register_to_config(latent_dim_scale=latent_dim_scale) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + attention_mask = attention_mask[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask.to(device)) + text_encoder_hidden_states = text_encoder_output.last_hidden_state + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + + uncond_text_encoder_hidden_states = None + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + negative_prompt_embeds_text_encoder_output = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=uncond_input.attention_mask.to(device) + ) + + uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + return text_encoder_hidden_states, uncond_text_encoder_hidden_states + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image_embeddings: torch.Tensor | list[torch.Tensor], + prompt: str | list[str] = None, + num_inference_steps: int = 12, + timesteps: list[float] | None = None, + guidance_scale: float = 0.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + image_embeddings (`torch.Tensor` or `list[torch.Tensor]`): + Image Embeddings either extracted from an image or generated by a Prior Model. + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + num_inference_steps (`int`, *optional*, defaults to 12): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` + timesteps are used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 0.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `decoder_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `decoder_guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is a list with the generated image + embeddings. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # 0. Define commonly used variables + device = self._execution_device + dtype = self.decoder.dtype + self._guidance_scale = guidance_scale + + # 1. Check inputs. Raise error if not correct + if not isinstance(prompt, list): + if isinstance(prompt, str): + prompt = [prompt] + else: + raise TypeError(f"'prompt' must be of type 'list' or 'str', but got {type(prompt)}.") + + if self.do_classifier_free_guidance: + if negative_prompt is not None and not isinstance(negative_prompt, list): + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + else: + raise TypeError( + f"'negative_prompt' must be of type 'list' or 'str', but got {type(negative_prompt)}." + ) + + if isinstance(image_embeddings, list): + image_embeddings = torch.cat(image_embeddings, dim=0) + if isinstance(image_embeddings, np.ndarray): + image_embeddings = torch.Tensor(image_embeddings, device=device).to(dtype=dtype) + if not isinstance(image_embeddings, torch.Tensor): + raise TypeError( + f"'image_embeddings' must be of type 'torch.Tensor' or 'np.array', but got {type(image_embeddings)}." + ) + + if not isinstance(num_inference_steps, int): + raise TypeError( + f"'num_inference_steps' must be of type 'int', but got {type(num_inference_steps)}\ + In Case you want to provide explicit timesteps, please use the 'timesteps' argument." + ) + + # 2. Encode caption + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + image_embeddings.size(0) * num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + ) + text_encoder_hidden_states = ( + torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds + ) + effnet = ( + torch.cat([image_embeddings, torch.zeros_like(image_embeddings)]) + if self.do_classifier_free_guidance + else image_embeddings + ) + + # 3. Determine latent shape of latents + latent_height = int(image_embeddings.size(2) * self.config.latent_dim_scale) + latent_width = int(image_embeddings.size(3) * self.config.latent_dim_scale) + latent_features_shape = (image_embeddings.size(0) * num_images_per_prompt, 4, latent_height, latent_width) + + # 4. Prepare and set timesteps + if timesteps is not None: + self.scheduler.set_timesteps(timesteps=timesteps, device=device) + timesteps = self.scheduler.timesteps + num_inference_steps = len(timesteps) + else: + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latents + latents = self.prepare_latents(latent_features_shape, dtype, device, generator, latents, self.scheduler) + + # 6. Run denoising loop + self._num_timesteps = len(timesteps[:-1]) + for i, t in enumerate(self.progress_bar(timesteps[:-1])): + ratio = t.expand(latents.size(0)).to(dtype) + # 7. Denoise latents + predicted_latents = self.decoder( + torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents, + r=torch.cat([ratio] * 2) if self.do_classifier_free_guidance else ratio, + effnet=effnet, + clip=text_encoder_hidden_states, + ) + + # 8. Check for classifier free guidance and apply it + if self.do_classifier_free_guidance: + predicted_latents_text, predicted_latents_uncond = predicted_latents.chunk(2) + predicted_latents = torch.lerp(predicted_latents_uncond, predicted_latents_text, self.guidance_scale) + + # 9. Renoise latents to next timestep + latents = self.scheduler.step( + model_output=predicted_latents, + timestep=ratio, + sample=latents, + generator=generator, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + image_embeddings = callback_outputs.pop("image_embeddings", image_embeddings) + text_encoder_hidden_states = callback_outputs.pop( + "text_encoder_hidden_states", text_encoder_hidden_states + ) + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type not in ["pt", "np", "pil", "latent"]: + raise ValueError( + f"Only the output types `pt`, `np`, `pil` and `latent` are supported not output_type={output_type}" + ) + + if not output_type == "latent": + # 10. Scale and decode the image latents with vq-vae + latents = self.vqgan.config.scale_factor * latents + images = self.vqgan.decode(latents).sample.clamp(0, 1) + if output_type == "np": + images = images.permute(0, 2, 3, 1).cpu().float().numpy() + elif output_type == "pil": + images = images.permute(0, 2, 3, 1).cpu().float().numpy() + images = self.numpy_to_pil(images) + else: + images = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return images + return ImagePipelineOutput(images) diff --git a/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen_combined.py b/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen_combined.py new file mode 100644 index 0000000000000000000000000000000000000000..3bffdcae251dfe31f3427b77474ee19c6e1bfe1a --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen_combined.py @@ -0,0 +1,307 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Callable + +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ....schedulers import DDPMWuerstchenScheduler +from ....utils import deprecate, replace_example_docstring +from ...pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline +from .modeling_paella_vq_model import PaellaVQModel +from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt +from .modeling_wuerstchen_prior import WuerstchenPrior +from .pipeline_wuerstchen import WuerstchenDecoderPipeline +from .pipeline_wuerstchen_prior import WuerstchenPriorPipeline + + +TEXT2IMAGE_EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusions import WuerstchenCombinedPipeline + + >>> pipe = WuerstchenCombinedPipeline.from_pretrained("warp-ai/Wuerstchen", torch_dtype=torch.float16).to( + ... "cuda" + ... ) + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> images = pipe(prompt=prompt) + ``` +""" + + +class WuerstchenCombinedPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Combined Pipeline for text-to-image generation using Wuerstchen + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + tokenizer (`CLIPTokenizer`): + The decoder tokenizer to be used for text inputs. + text_encoder (`CLIPTextModel`): + The decoder text encoder to be used for text inputs. + decoder (`WuerstchenDiffNeXt`): + The decoder model to be used for decoder image generation pipeline. + scheduler (`DDPMWuerstchenScheduler`): + The scheduler to be used for decoder image generation pipeline. + vqgan (`PaellaVQModel`): + The VQGAN model to be used for decoder image generation pipeline. + prior_tokenizer (`CLIPTokenizer`): + The prior tokenizer to be used for text inputs. + prior_text_encoder (`CLIPTextModel`): + The prior text encoder to be used for text inputs. + prior_prior (`WuerstchenPrior`): + The prior model to be used for prior pipeline. + prior_scheduler (`DDPMWuerstchenScheduler`): + The scheduler to be used for prior pipeline. + """ + + _last_supported_version = "0.33.1" + _load_connected_pipes = True + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + decoder: WuerstchenDiffNeXt, + scheduler: DDPMWuerstchenScheduler, + vqgan: PaellaVQModel, + prior_tokenizer: CLIPTokenizer, + prior_text_encoder: CLIPTextModel, + prior_prior: WuerstchenPrior, + prior_scheduler: DDPMWuerstchenScheduler, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + prior_prior=prior_prior, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + ) + self.prior_pipe = WuerstchenPriorPipeline( + prior=prior_prior, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + ) + self.decoder_pipe = WuerstchenDecoderPipeline( + text_encoder=text_encoder, + tokenizer=tokenizer, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_model_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + """ + self.prior_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗 + Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a + GPU only when their specific submodule's `forward` method is called. Offloading happens on a submodule basis. + Memory savings are higher than using `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + height: int = 512, + width: int = 512, + prior_num_inference_steps: int = 60, + prior_timesteps: list[float] | None = None, + prior_guidance_scale: float = 4.0, + num_inference_steps: int = 12, + decoder_timesteps: list[float] | None = None, + decoder_guidance_scale: float = 0.0, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + prior_callback_on_step_end: Callable[[int, int], None] | None = None, + prior_callback_on_step_end_tensor_inputs: list[str] = ["latents"], + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation for the prior and decoder. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* + prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` + input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `prior_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `prior_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int | dict[float, int]`, *optional*, defaults to 60): + The number of prior denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. For more specific timestep spacing, you can pass customized + `prior_timesteps` + num_inference_steps (`int`, *optional*, defaults to 12): + The number of decoder denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. For more specific timestep spacing, you can pass customized + `timesteps` + prior_timesteps (`list[float]`, *optional*): + Custom timesteps to use for the denoising process for the prior. If not defined, equal spaced + `prior_num_inference_steps` timesteps are used. Must be in descending order. + decoder_timesteps (`list[float]`, *optional*): + Custom timesteps to use for the denoising process for the decoder. If not defined, equal spaced + `num_inference_steps` timesteps are used. Must be in descending order. + decoder_guidance_scale (`float`, *optional*, defaults to 0.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + prior_callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `prior_callback_on_step_end(self: DiffusionPipeline, step: int, timestep: + int, callback_kwargs: Dict)`. + prior_callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the + list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in + the `._callback_tensor_inputs` attribute of your pipeline class. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + prior_kwargs = {} + if kwargs.get("prior_callback", None) is not None: + prior_kwargs["callback"] = kwargs.pop("prior_callback") + deprecate( + "prior_callback", + "1.0.0", + "Passing `prior_callback` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`", + ) + if kwargs.get("prior_callback_steps", None) is not None: + deprecate( + "prior_callback_steps", + "1.0.0", + "Passing `prior_callback_steps` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`", + ) + prior_kwargs["callback_steps"] = kwargs.pop("prior_callback_steps") + + prior_outputs = self.prior_pipe( + prompt=prompt if prompt_embeds is None else None, + height=height, + width=width, + num_inference_steps=prior_num_inference_steps, + timesteps=prior_timesteps, + guidance_scale=prior_guidance_scale, + negative_prompt=negative_prompt if negative_prompt_embeds is None else None, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + output_type="pt", + return_dict=False, + callback_on_step_end=prior_callback_on_step_end, + callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs, + **prior_kwargs, + ) + image_embeddings = prior_outputs[0] + + outputs = self.decoder_pipe( + image_embeddings=image_embeddings, + prompt=prompt if prompt is not None else "", + num_inference_steps=num_inference_steps, + timesteps=decoder_timesteps, + guidance_scale=decoder_guidance_scale, + negative_prompt=negative_prompt, + generator=generator, + output_type=output_type, + return_dict=return_dict, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + **kwargs, + ) + + return outputs diff --git a/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen_prior.py b/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen_prior.py new file mode 100644 index 0000000000000000000000000000000000000000..4a95796b64819ced413f8226a4a8f59b773471fb --- /dev/null +++ b/diffusers/pipelines/deprecated/wuerstchen/pipeline_wuerstchen_prior.py @@ -0,0 +1,528 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from math import ceil +from typing import Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ....loaders import StableDiffusionLoraLoaderMixin +from ....schedulers import DDPMWuerstchenScheduler +from ....utils import BaseOutput, deprecate, is_torch_xla_available, logging, replace_example_docstring +from ....utils.torch_utils import randn_tensor +from ...pipeline_utils import DiffusionPipeline +from .modeling_wuerstchen_prior import WuerstchenPrior + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +DEFAULT_STAGE_C_TIMESTEPS = list(np.linspace(1.0, 2 / 3, 20)) + list(np.linspace(2 / 3, 0.0, 11))[1:] + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import WuerstchenPriorPipeline + + >>> prior_pipe = WuerstchenPriorPipeline.from_pretrained( + ... "warp-ai/wuerstchen-prior", torch_dtype=torch.float16 + ... ).to("cuda") + + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> prior_output = pipe(prompt) + ``` +""" + + +@dataclass +class WuerstchenPriorPipelineOutput(BaseOutput): + """ + Output class for WuerstchenPriorPipeline. + + Args: + image_embeddings (`torch.Tensor` or `np.ndarray`) + Prior image embeddings for text prompt + + """ + + image_embeddings: torch.Tensor | np.ndarray + + +class WuerstchenPriorPipeline(DiffusionPipeline, StableDiffusionLoraLoaderMixin): + """ + Pipeline for generating image prior for Wuerstchen. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + prior ([`Prior`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + scheduler ([`DDPMWuerstchenScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + latent_mean ('float', *optional*, defaults to 42.0): + Mean value for latent diffusers. + latent_std ('float', *optional*, defaults to 1.0): + Standard value for latent diffusers. + resolution_multiple ('float', *optional*, defaults to 42.67): + Default resolution for multiple images generated. + """ + + unet_name = "prior" + text_encoder_name = "text_encoder" + model_cpu_offload_seq = "text_encoder->prior" + _callback_tensor_inputs = ["latents", "text_encoder_hidden_states", "negative_prompt_embeds"] + _lora_loadable_modules = ["prior", "text_encoder"] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + prior: WuerstchenPrior, + scheduler: DDPMWuerstchenScheduler, + latent_mean: float = 42.0, + latent_std: float = 1.0, + resolution_multiple: float = 42.67, + ) -> None: + super().__init__() + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + prior=prior, + scheduler=scheduler, + ) + self.register_to_config( + latent_mean=latent_mean, latent_std=latent_std, resolution_multiple=resolution_multiple + ) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def encode_prompt( + self, + device, + num_images_per_prompt, + do_classifier_free_guidance, + prompt=None, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ): + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + attention_mask = attention_mask[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask.to(device) + ) + prompt_embeds = text_encoder_output.last_hidden_state + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + if negative_prompt_embeds is None and do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + negative_prompt_embeds_text_encoder_output = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=uncond_input.attention_mask.to(device) + ) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.last_hidden_state + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + # done duplicates + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + num_inference_steps, + do_classifier_free_guidance, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if not isinstance(num_inference_steps, int): + raise TypeError( + f"'num_inference_steps' must be of type 'int', but got {type(num_inference_steps)}\ + In Case you want to provide explicit timesteps, please use the 'timesteps' argument." + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + height: int = 1024, + width: int = 1024, + num_inference_steps: int = 60, + timesteps: list[float] = None, + guidance_scale: float = 8.0, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pt", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 60): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` + timesteps are used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 8.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `decoder_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `decoder_guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.WuerstchenPriorPipelineOutput`] or `tuple` [`~pipelines.WuerstchenPriorPipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated image embeddings. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # 0. Define commonly used variables + device = self._execution_device + self._guidance_scale = guidance_scale + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 1. Check inputs. Raise error if not correct + if prompt is not None and not isinstance(prompt, list): + if isinstance(prompt, str): + prompt = [prompt] + else: + raise TypeError(f"'prompt' must be of type 'list' or 'str', but got {type(prompt)}.") + + if self.do_classifier_free_guidance: + if negative_prompt is not None and not isinstance(negative_prompt, list): + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + else: + raise TypeError( + f"'negative_prompt' must be of type 'list' or 'str', but got {type(negative_prompt)}." + ) + + self.check_inputs( + prompt, + negative_prompt, + num_inference_steps, + self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # 2. Encode caption + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_encoder_hidden_states = ( + torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds + ) + + # 3. Determine latent shape of image embeddings + dtype = text_encoder_hidden_states.dtype + latent_height = ceil(height / self.config.resolution_multiple) + latent_width = ceil(width / self.config.resolution_multiple) + num_channels = self.prior.config.c_in + effnet_features_shape = (num_images_per_prompt * batch_size, num_channels, latent_height, latent_width) + + # 4. Prepare and set timesteps + if timesteps is not None: + self.scheduler.set_timesteps(timesteps=timesteps, device=device) + timesteps = self.scheduler.timesteps + num_inference_steps = len(timesteps) + else: + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latents + latents = self.prepare_latents(effnet_features_shape, dtype, device, generator, latents, self.scheduler) + + # 6. Run denoising loop + self._num_timesteps = len(timesteps[:-1]) + for i, t in enumerate(self.progress_bar(timesteps[:-1])): + ratio = t.expand(latents.size(0)).to(dtype) + + # 7. Denoise image embeddings + predicted_image_embedding = self.prior( + torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents, + r=torch.cat([ratio] * 2) if self.do_classifier_free_guidance else ratio, + c=text_encoder_hidden_states, + ) + + # 8. Check for classifier free guidance and apply it + if self.do_classifier_free_guidance: + predicted_image_embedding_text, predicted_image_embedding_uncond = predicted_image_embedding.chunk(2) + predicted_image_embedding = torch.lerp( + predicted_image_embedding_uncond, predicted_image_embedding_text, self.guidance_scale + ) + + # 9. Renoise latents to next timestep + latents = self.scheduler.step( + model_output=predicted_image_embedding, + timestep=ratio, + sample=latents, + generator=generator, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + text_encoder_hidden_states = callback_outputs.pop( + "text_encoder_hidden_states", text_encoder_hidden_states + ) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 10. Denormalize the latents + latents = latents * self.config.latent_mean - self.config.latent_std + + # Offload all models + self.maybe_free_model_hooks() + + if output_type == "np": + latents = latents.cpu().float().numpy() + + if not return_dict: + return (latents,) + + return WuerstchenPriorPipelineOutput(latents) diff --git a/diffusers/pipelines/diffusion_gemma/__init__.py b/diffusers/pipelines/diffusion_gemma/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5825d94e95949cabf40ca7b3a5c58284d69975a1 --- /dev/null +++ b/diffusers/pipelines/diffusion_gemma/__init__.py @@ -0,0 +1,49 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_diffusion_gemma"] = ["DiffusionGemmaPipeline"] + _import_structure["pipeline_output"] = ["DiffusionGemmaPipelineOutput"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_diffusion_gemma import DiffusionGemmaPipeline + from .pipeline_output import DiffusionGemmaPipelineOutput +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/diffusion_gemma/pipeline_diffusion_gemma.py b/diffusers/pipelines/diffusion_gemma/pipeline_diffusion_gemma.py new file mode 100644 index 0000000000000000000000000000000000000000..5222ead8813b856d227894bab4fd0a76266834d6 --- /dev/null +++ b/diffusers/pipelines/diffusion_gemma/pipeline_diffusion_gemma.py @@ -0,0 +1,460 @@ +# Copyright 2025 The Google and HuggingFace Teams. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import inspect +from typing import Any, Callable + +import torch +import torch.nn.functional as F +from transformers import DynamicCache, StaticCache + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...schedulers import BlockRefinementScheduler, DiscreteDDIMScheduler, EntropyBoundScheduler +from ...utils import logging, replace_example_docstring +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import DiffusionGemmaPipelineOutput + + +logger = logging.get_logger(__name__) + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from transformers import AutoProcessor, DiffusionGemmaForBlockDiffusion + >>> from diffusers import BlockRefinementScheduler, DiffusionGemmaPipeline + + >>> model_id = "google/diffusiongemma-26B-A4B-it" + >>> model = DiffusionGemmaForBlockDiffusion.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto") + >>> processor = AutoProcessor.from_pretrained(model_id) + >>> scheduler = BlockRefinementScheduler() + + >>> pipe = DiffusionGemmaPipeline(model=model, scheduler=scheduler, processor=processor) + >>> output = pipe(prompt="Why is the sky blue?", gen_length=256) + >>> print(output.texts[0]) + ``` +""" + + +class DiffusionGemmaPipeline(DiffusionPipeline): + r""" + Pipeline for DiffusionGemma block-diffusion text generation. + + DiffusionGemma is a block-diffusion encoder-decoder model: a causal encoder reads the clean prompt (and any + previously generated blocks) into a KV cache, and a bidirectional decoder denoises a fixed-size "canvas" of + `canvas_length` tokens by cross-attending to that cache. Generation alternates an outer autoregressive loop over + canvases with an inner denoising loop, where each step samples candidate tokens, commits the most confident ones + via [`BlockRefinementScheduler`] (uniform corruption mode, `mask_token_id=None`), and renoises the rest. + + The model is expected to be a `DiffusionGemmaForBlockDiffusion` instance exposing `forward(input_ids, + decoder_input_ids=..., self_conditioning_logits=..., ...)` and returning logits of shape `[batch, canvas_length, + vocab_size]` over the canvas. See the model card at https://huggingface.co/google/diffusiongemma-26B-A4B-it. + + Args: + model ([`~transformers.DiffusionGemmaForBlockDiffusion`]): + The block-diffusion denoiser (causal encoder + bidirectional decoder with tied weights). + scheduler ([`BlockRefinementScheduler`], [`DiscreteDDIMScheduler`] or [`EntropyBoundScheduler`]): + The sampler that commits and renoises canvas tokens each denoising step. + processor ([`~transformers.ProcessorMixin`]): + The processor used to apply the chat template and decode the generated tokens. + """ + + _callback_tensor_inputs = ["canvas", "logits"] + + def __init__( + self, + model: Any, + scheduler: BlockRefinementScheduler | DiscreteDDIMScheduler | EntropyBoundScheduler, + processor: Any, + ): + super().__init__() + self.register_modules(model=model, scheduler=scheduler, processor=processor) + tokenizer = getattr(processor, "tokenizer", processor) + self.eos_token_id = getattr(tokenizer, "eos_token_id", None) if tokenizer is not None else None + + @property + def num_timesteps(self): + return self._num_timesteps + + # --- Prompt encoding --- + + def _prepare_inputs( + self, + *, + prompt: str | list[str] | None, + messages: list[dict] | None, + image: Any | list[Any] | None, + add_generation_prompt: bool, + ) -> tuple[torch.LongTensor, torch.LongTensor, dict[str, torch.Tensor]]: + """Tokenize a raw `prompt` (optionally with an `image`) or a raw `messages` conversation into + `(input_ids, attention_mask, multimodal_inputs)`, where `multimodal_inputs` holds the image tensors the + processor produced for the encoder prefill.""" + + def build_content(text, img): + if img is None: + return text + return [{"type": "image", "image": img}, {"type": "text", "text": text}] + + if messages is None: + if isinstance(prompt, list): + images = image if isinstance(image, list) else [image] * len(prompt) + messages = [[{"role": "user", "content": build_content(p, im)}] for p, im in zip(prompt, images)] + else: + messages = [{"role": "user", "content": build_content(prompt, image)}] + + encoded = self.processor.apply_chat_template( + messages, + add_generation_prompt=add_generation_prompt, + tokenize=True, + return_tensors="pt", + return_dict=True, + ) + ids = encoded["input_ids"] + mask = encoded.get("attention_mask") + if mask is None: + mask = torch.ones_like(ids, dtype=torch.long) + multimodal_keys = ("pixel_values", "image_position_ids", "mm_token_type_ids") + multimodal_inputs = {k: encoded[k] for k in multimodal_keys if k in encoded} + return ids, mask.to(dtype=torch.long), multimodal_inputs + + def check_inputs( + self, + prompt: str | list[str] | None, + messages: list[dict] | None, + gen_length: int, + num_inference_steps: int, + output_type: str, + callback_on_step_end_tensor_inputs: list[str] | None, + ): + if output_type not in {"seq", "text"}: + raise ValueError(f"`output_type` must be 'seq' or 'text', got {output_type!r}.") + if gen_length <= 0: + raise ValueError(f"`gen_length` must be > 0, got {gen_length}.") + if num_inference_steps <= 0: + raise ValueError(f"`num_inference_steps` must be > 0, got {num_inference_steps}.") + if prompt is None and messages is None: + raise ValueError("Provide either `prompt` or `messages`.") + if prompt is not None and messages is not None: + raise ValueError("Provide either `prompt` or `messages`, not both.") + if self.processor is None: + raise ValueError("`processor` is required to encode the prompt.") + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found " + f"{[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + messages: list[dict] | None = None, + image: Any | list[Any] | None = None, + add_generation_prompt: bool = True, + gen_length: int = 256, + num_inference_steps: int = 48, + temperature: float = 0.0, + cache_implementation: str | None = None, + eos_early_stop: bool = True, + eos_token_id: int | None = None, + stability_threshold: int = 1, + confidence_threshold: float | None = 0.005, + generator: torch.Generator | None = None, + output_type: str = "text", + return_dict: bool = True, + callback_on_step_end: Callable[[Any, int, int, dict], dict] + | PipelineCallback + | MultiPipelineCallbacks + | None = None, + callback_on_step_end_tensor_inputs: list[str] | None = None, + ) -> DiffusionGemmaPipelineOutput | tuple[torch.LongTensor, list[str] | None]: + """ + Generate text with block diffusion. + + Args: + prompt (`str` or `List[str]`, *optional*): + Prompt text, wrapped in a chat template and tokenized by the processor. Provide either this or + `messages`. + messages (`List[Dict]`, *optional*): + A raw chat conversation to encode, e.g. `[{"role": "user", "content": "Hello"}]` or a multi-turn / + multimodal conversation. Use this instead of `prompt` for anything beyond a single user turn. + image (`PIL.Image.Image` or `List`, *optional*): + Image(s) to pair with `prompt` for multimodal generation; the processor turns them into the model's + image inputs. For richer layouts, put the image content directly in `messages`. + add_generation_prompt (`bool`, defaults to `True`): + Whether to add the generation prompt when applying the chat template. + gen_length (`int`, defaults to `256`): + Number of tokens to generate, rounded up to a multiple of the model's `canvas_length`. + num_inference_steps (`int`, defaults to `48`): + Number of denoising steps per canvas. + temperature (`float`, defaults to `0.0`): + Sampling temperature for `DiscreteDDIMScheduler`/`BlockRefinementScheduler` (`0.0` is greedy); + `EntropyBoundScheduler` ignores it and anneals its own temperature. Other sampling knobs (e.g. `top_k`, + `threshold`, `t_min`/`t_max`) are scheduler config; set them on the scheduler, e.g. `pipe.scheduler = + BlockRefinementScheduler.from_config(pipe.scheduler.config, top_k=...)`. + cache_implementation (`str`, *optional*): + Set to `"static"` to prefill the encoder once per block into a persistent `StaticCache` and run the + decoder against it with fixed shapes, instead of re-encoding the full sequence on every step. The fixed + shapes also let you compile the decoder, e.g. `pipe.model.model.decoder = + torch.compile(pipe.model.model.decoder, fullgraph=True)`. + eos_early_stop (`bool`, defaults to `True`): + Whether to stop generating further canvases once every sequence has emitted EOS. + eos_token_id (`int`, *optional*): + EOS token ID for early stopping. Falls back to the processor's tokenizer. + stability_threshold (`int`, defaults to `1`): + Number of consecutive steps the argmax prediction must be unchanged for a block to count as stable. + Only used when `confidence_threshold` is set. + confidence_threshold (`float`, *optional*, defaults to `0.005`): + Leave a block's denoising loop early once every example is stable (see `stability_threshold`) and the + mean per-token entropy of the prediction is below this value. Speeds up generation at matched quality; + the default matches the released checkpoint. Set to `None` to always run all `num_inference_steps`. + generator (`torch.Generator`, *optional*): + RNG for sampling. + output_type (`str`, defaults to `"text"`): + `"text"` decodes sequences into strings (requires a processor); `"seq"` returns token IDs only. + return_dict (`bool`, defaults to `True`): + Whether to return a [`DiffusionGemmaPipelineOutput`] instead of a tuple. + callback_on_step_end (`Callable` or `PipelineCallback`, *optional*): + Callback run after each denoising step with signature `callback_on_step_end(self, step, timestep, + callback_kwargs)`. Allowed tensor keys: `canvas`, `logits`. + callback_on_step_end_tensor_inputs (`List[str]`, *optional*): + Tensor keys to pass to the callback. + + Examples: + + Returns: + [`~pipelines.diffusion_gemma.pipeline_diffusion_gemma.DiffusionGemmaPipelineOutput`] or `tuple`: + The generated token IDs (`sequences`) and, for `output_type="text"`, the decoded `texts`. + """ + if callback_on_step_end is not None and isinstance( + callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks) + ): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + if callback_on_step_end_tensor_inputs is None: + callback_on_step_end_tensor_inputs = ["canvas"] + + self.check_inputs( + prompt=prompt, + messages=messages, + gen_length=gen_length, + num_inference_steps=num_inference_steps, + output_type=output_type, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + prompt_ids, prompt_attention_mask, multimodal_inputs = self._prepare_inputs( + prompt=prompt, + messages=messages, + image=image, + add_generation_prompt=add_generation_prompt, + ) + + device = self._execution_device + prompt_ids = prompt_ids.to(device=device) + prompt_attention_mask = prompt_attention_mask.to(device=device) + multimodal_inputs = {k: v.to(device=device) for k, v in multimodal_inputs.items()} + batch_size, prompt_length = prompt_ids.shape + + if eos_token_id is None: + eos_token_id = self.eos_token_id + + canvas_length = self.model.config.canvas_length + num_canvases = (gen_length + canvas_length - 1) // canvas_length + # `num_inference_steps` is the per-block budget of model forwards. With a corrector, fold its sweeps into that + # budget (as in https://huggingface.co/papers/2605.22765) instead of adding them on top: the first + # `corrected_steps` predictor steps each run `corrector_steps` extra forwards, so the total stays + # `num_inference_steps` and the predictor-corrector costs the same as plain ancestral sampling. + corrector_steps = getattr(self.scheduler.config, "corrector_steps", 0) + if corrector_steps > 0: + corrected_steps = (num_inference_steps - 1) // (1 + corrector_steps) + predictor_steps = num_inference_steps - corrected_steps * corrector_steps + else: + corrected_steps = 0 + predictor_steps = num_inference_steps + + # Only `BlockRefinementScheduler` takes a per-call `block_length`; the DiscreteDDIM/EntropyBound schedulers do + # not, so we pass scheduler-specific kwargs by signature. + set_timesteps_kwargs = {"device": device} + if "block_length" in inspect.signature(self.scheduler.set_timesteps).parameters: + set_timesteps_kwargs["block_length"] = canvas_length + self.scheduler.set_timesteps(predictor_steps, **set_timesteps_kwargs) + step_param_names = set(inspect.signature(self.scheduler.step).parameters) + self._num_timesteps = predictor_steps * num_canvases + + cur_input_ids = prompt_ids + cur_attention_mask = prompt_attention_mask + finished = torch.zeros(batch_size, dtype=torch.bool, device=device) + global_step = 0 + + # Encode each block of context once into a reusable KV cache and run the decoder against it, rather than + # re-encoding the whole sequence on every denoising step. The default `DynamicCache` grows with the context; + # `cache_implementation="static"` uses a fixed-shape `StaticCache` so the decoder can be `torch.compile`-d. + use_static_cache = cache_implementation == "static" + text_config = self.model.config.get_text_config(decoder=True) + max_cache_len = prompt_length + num_canvases * canvas_length + if use_static_cache: + past_key_values = StaticCache(config=text_config, max_cache_len=max_cache_len) + else: + past_key_values = DynamicCache(config=text_config) + + progress_bar = self.progress_bar(total=self._num_timesteps) + for _ in range(num_canvases): + cur_len = cur_input_ids.shape[1] + decoder_position_ids = torch.arange(cur_len, cur_len + canvas_length, device=device).unsqueeze(0) + + # Encode the tokens not yet in the cache (the whole prompt on the first block, the last committed canvas + # afterwards), so the decoder reuses the encoder KV cache instead of re-encoding the full sequence. + cached_len = past_key_values.get_seq_length() + torch.compiler.cudagraph_mark_step_begin() + self.model.model.encoder( + input_ids=cur_input_ids[:, cached_len:], + attention_mask=cur_attention_mask, + past_key_values=past_key_values, + position_ids=torch.arange(cached_len, cur_len, device=device).unsqueeze(0), + # Image tensors are consumed by the prompt prefill only; later blocks encode text-only canvases. + **(multimodal_inputs if cached_len == 0 else {}), + ) + + # Decoder attends bidirectionally over the populated cache (the live padding mask) plus the always-visible + # canvas; the mask builder sizes this to the cache internally, including the static buffer for a StaticCache. + decoder_attention_mask = torch.nn.functional.pad(cur_attention_mask.bool(), (0, canvas_length), value=True) + mask_mapping = self.model.model.decoder.create_diffusion_decoder_attention_mask( + config=self.model.config, + inputs_embeds=torch.empty((batch_size, canvas_length, 0), device=device), + past_key_values=past_key_values, + decoder_attention_mask=decoder_attention_mask, + ) + + # Start from a fully random canvas and denoise it; the scheduler resets its committed state at step 0. + canvas = torch.randint( + 0, text_config.vocab_size, (batch_size, canvas_length), device=device, generator=generator + ) + self_conditioning_logits = None + # Adaptive stopping history: the last `stability_threshold` argmax predictions of this block's canvas. + argmax_history = torch.full( + (max(stability_threshold, 1), batch_size, canvas_length), -1, dtype=torch.long, device=device + ) + + # Denoise the predictor steps of this canvas; the first `corrected_steps` also run corrector sweeps. + for step_idx in range(predictor_steps): + if corrected_steps: + progress_bar.set_description( + "denoising (corrector)" if step_idx < corrected_steps else "denoising" + ) + # Mark a fresh step and clone the logits so a cudagraph-compiled decoder (`mode="reduce-overhead"`) + # does not overwrite the tensors that self-conditioning and the scheduler read next. Both are no-ops + # when the decoder is not cudagraph-compiled. + torch.compiler.cudagraph_mark_step_begin() + logits = self.model( + decoder_input_ids=canvas, + past_key_values=past_key_values, + self_conditioning_logits=self_conditioning_logits, + decoder_attention_mask=mask_mapping, + decoder_position_ids=decoder_position_ids, + ).logits.clone() + + # Pass only the kwargs the chosen scheduler accepts, so any of the schedulers can drive the pipeline. + # Sampling knobs (temperature annealing, thresholds, top-k, ...) live on the scheduler config, not here. + step_kwargs = {"mask_token_id": None, "temperature": temperature, "generator": generator} + step_kwargs = {k: v for k, v in step_kwargs.items() if k in step_param_names} + scheduler_output = self.scheduler.step( + model_output=logits, timestep=step_idx, sample=canvas, return_dict=True, **step_kwargs + ) + canvas = scheduler_output.prev_sample + # Self-condition on the logits the scheduler sampled from: temperature-shaped for the reference + # EntropyBound sampler, the raw denoiser logits for the others. + self_conditioning_logits = scheduler_output.pred_logits + + # Predictor-corrector (https://huggingface.co/papers/2605.22765): a scheduler exposing `corrector_steps` + # + `step_correct` refines the canvas with extra Gibbs sweeps on the first `corrected_steps` predictor + # steps (the budget split computed above). Each sweep needs fresh logits on the updated canvas. + if step_idx < corrected_steps: + for _ in range(corrector_steps): + torch.compiler.cudagraph_mark_step_begin() + corrector_logits = self.model( + decoder_input_ids=canvas, + past_key_values=past_key_values, + self_conditioning_logits=self_conditioning_logits, + decoder_attention_mask=mask_mapping, + decoder_position_ids=decoder_position_ids, + ).logits.clone() + canvas = self.scheduler.step_correct( + model_output=corrector_logits, timestep=step_idx, sample=canvas, generator=generator + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, global_step, step_idx, callback_kwargs) + canvas = callback_outputs.pop("canvas", canvas) + global_step += 1 + progress_bar.update() + + # Adaptive stopping: leave this block early once every example's argmax prediction is stable across + # `stability_threshold` steps and confident (mean per-token entropy below `confidence_threshold`). + if confidence_threshold is not None: + argmax_canvas = logits.argmax(dim=-1) + stable = (argmax_history == argmax_canvas[None]).all(dim=-1).all(dim=0) + argmax_history = torch.roll(argmax_history, shifts=-1, dims=0) + argmax_history[-1] = argmax_canvas + confident = torch.distributions.Categorical(logits=logits.float()).entropy().mean(-1) < ( + confidence_threshold + ) + if bool((stable & confident).all()): + # Commit the converged prediction. Ancestral schedulers (e.g. DiscreteDDIM) only clean the + # canvas on their final step, so the in-progress canvas may still hold noise tokens; the + # denoiser argmax is the converged answer (and equals the canvas for commit-style schedulers). + canvas = argmax_canvas + break + + # Append the denoised canvas and extend the context for the next block. + cur_input_ids = torch.cat([cur_input_ids, canvas], dim=-1) + cur_attention_mask = F.pad(cur_attention_mask, (0, canvas_length), value=1) + + if eos_early_stop and eos_token_id is not None: + finished = finished | (canvas == eos_token_id).any(dim=-1) + if finished.all(): + break + + progress_bar.close() + + sequences = cur_input_ids[:, prompt_length:] + + # Trim each row at its first EOS so post-EOS canvas tokens don't leak into the decoded text. + decode_sequences: list[torch.LongTensor] | torch.LongTensor = sequences + if eos_token_id is not None: + decode_sequences = [ + seq[: int((seq == eos_token_id).nonzero(as_tuple=True)[0][0]) + 1] + if (seq == eos_token_id).any() + else seq + for seq in sequences + ] + + texts = None + if output_type == "text" and self.processor is not None: + texts = self.processor.batch_decode(decode_sequences, skip_special_tokens=True) + + if not return_dict: + return sequences, texts + return DiffusionGemmaPipelineOutput(sequences=sequences, texts=texts) + + +__all__ = ["DiffusionGemmaPipeline", "DiffusionGemmaPipelineOutput"] diff --git a/diffusers/pipelines/diffusion_gemma/pipeline_output.py b/diffusers/pipelines/diffusion_gemma/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..845f71ebdb0d7d558427270faea707393b57627a --- /dev/null +++ b/diffusers/pipelines/diffusion_gemma/pipeline_output.py @@ -0,0 +1,37 @@ +# Copyright 2025 The Google and HuggingFace Teams. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from dataclasses import dataclass + +import torch + +from ...utils import BaseOutput + + +@dataclass +class DiffusionGemmaPipelineOutput(BaseOutput): + """ + Output class for DiffusionGemma block-diffusion generation. + + Args: + sequences (`torch.LongTensor` of shape `(batch_size, gen_length)`): + The generated token IDs (the prompt is stripped off). + texts (`list[str]`, *optional*): + The decoded text, one string per sequence. Only set for `output_type="text"`. + """ + + sequences: torch.LongTensor + texts: list[str] | None = None diff --git a/diffusers/pipelines/dit/__init__.py b/diffusers/pipelines/dit/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..fe2a94f3cba77d867f97111a41895918842adc27 --- /dev/null +++ b/diffusers/pipelines/dit/__init__.py @@ -0,0 +1,19 @@ +from typing import TYPE_CHECKING + +from ...utils import DIFFUSERS_SLOW_IMPORT, _LazyModule + + +_import_structure = {"pipeline_dit": ["DiTPipeline"]} + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + from .pipeline_dit import DiTPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) diff --git a/diffusers/pipelines/dit/pipeline_dit.py b/diffusers/pipelines/dit/pipeline_dit.py new file mode 100644 index 0000000000000000000000000000000000000000..dd73d364067ee2d8fdb29b7e9f0932a42d51e9a0 --- /dev/null +++ b/diffusers/pipelines/dit/pipeline_dit.py @@ -0,0 +1,249 @@ +# Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) +# William Peebles and Saining Xie +# +# Copyright (c) 2021 OpenAI +# MIT License +# +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch + +from ...models import AutoencoderKL, DiTTransformer2DModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import is_torch_xla_available +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +class DiTPipeline(DiffusionPipeline): + r""" + Pipeline for image generation based on a Transformer backbone instead of a UNet. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Parameters: + transformer ([`DiTTransformer2DModel`]): + A class conditioned `DiTTransformer2DModel` to denoise the encoded image latents. Initially published as + [`Transformer2DModel`](https://huggingface.co/facebook/DiT-XL-2-256/blob/main/transformer/config.json#L2) + in the config, but the mismatch can be ignored. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + scheduler ([`DDIMScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "transformer->vae" + + def __init__( + self, + transformer: DiTTransformer2DModel, + vae: AutoencoderKL, + scheduler: KarrasDiffusionSchedulers, + id2label: dict[int, str] | None = None, + ): + super().__init__() + self.register_modules(transformer=transformer, vae=vae, scheduler=scheduler) + + # create a imagenet -> id dictionary for easier use + self.labels = {} + if id2label is not None: + for key, value in id2label.items(): + for label in value.split(","): + self.labels[label.lstrip().rstrip()] = int(key) + self.labels = dict(sorted(self.labels.items())) + + def get_label_ids(self, label: str | list[str]) -> list[int]: + r""" + + Map label strings from ImageNet to corresponding class ids. + + Parameters: + label (`str` or `dict` of `str`): + Label strings to be mapped to class ids. + + Returns: + `list` of `int`: + Class ids to be processed by pipeline. + """ + + if not isinstance(label, list): + label = list(label) + + for l in label: + if l not in self.labels: + raise ValueError( + f"{l} does not exist. Please make sure to select one of the following labels: \n {self.labels}." + ) + + return [self.labels[l] for l in label] + + @torch.no_grad() + def __call__( + self, + class_labels: list[int], + guidance_scale: float = 4.0, + generator: torch.Generator | list[torch.Generator] | None = None, + num_inference_steps: int = 50, + output_type: str | None = "pil", + return_dict: bool = True, + ) -> ImagePipelineOutput | tuple: + r""" + The call function to the pipeline for generation. + + Args: + class_labels (list[int]): + list of ImageNet class labels for the images to be generated. + guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + num_inference_steps (`int`, *optional*, defaults to 250): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + ```py + >>> from diffusers import DiTPipeline, DPMSolverMultistepScheduler + >>> import torch + + >>> pipe = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256", torch_dtype=torch.float16) + >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config) + >>> pipe = pipe.to("cuda") + + >>> # pick words from Imagenet class labels + >>> pipe.labels # to print all available words + + >>> # pick words that exist in ImageNet + >>> words = ["white shark", "umbrella"] + + >>> class_ids = pipe.get_label_ids(words) + + >>> generator = torch.manual_seed(33) + >>> output = pipe(class_labels=class_ids, num_inference_steps=25, generator=generator) + + >>> image = output.images[0] # label 'white shark' + ``` + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images + """ + + batch_size = len(class_labels) + latent_size = self.transformer.config.sample_size + latent_channels = self.transformer.config.in_channels + + latents = randn_tensor( + shape=(batch_size, latent_channels, latent_size, latent_size), + generator=generator, + device=self._execution_device, + dtype=self.transformer.dtype, + ) + latent_model_input = torch.cat([latents] * 2) if guidance_scale > 1 else latents + + class_labels = torch.tensor(class_labels, device=self._execution_device).reshape(-1) + class_null = torch.tensor([1000] * batch_size, device=self._execution_device) + class_labels_input = torch.cat([class_labels, class_null], 0) if guidance_scale > 1 else class_labels + + # set step values + self.scheduler.set_timesteps(num_inference_steps) + for t in self.progress_bar(self.scheduler.timesteps): + if guidance_scale > 1: + half = latent_model_input[: len(latent_model_input) // 2] + latent_model_input = torch.cat([half, half], dim=0) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + timesteps = t + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = latent_model_input.device.type == "mps" + is_npu = latent_model_input.device.type == "npu" + if isinstance(timesteps, float): + dtype = torch.float32 if (is_mps or is_npu) else torch.float64 + else: + dtype = torch.int32 if (is_mps or is_npu) else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=latent_model_input.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(latent_model_input.device) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(latent_model_input.shape[0]) + # predict noise model_output + noise_pred = self.transformer( + latent_model_input, timestep=timesteps, class_labels=class_labels_input + ).sample + + # perform guidance + if guidance_scale > 1: + eps, rest = noise_pred[:, :latent_channels], noise_pred[:, latent_channels:] + cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0) + + half_eps = uncond_eps + guidance_scale * (cond_eps - uncond_eps) + eps = torch.cat([half_eps, half_eps], dim=0) + + noise_pred = torch.cat([eps, rest], dim=1) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + model_output, _ = torch.split(noise_pred, latent_channels, dim=1) + else: + model_output = noise_pred + + # compute previous image: x_t -> x_t-1 + latent_model_input = self.scheduler.step(model_output, t, latent_model_input).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + if guidance_scale > 1: + latents, _ = latent_model_input.chunk(2, dim=0) + else: + latents = latent_model_input + + latents = 1 / self.vae.config.scaling_factor * latents + samples = self.vae.decode(latents).sample + + samples = (samples / 2 + 0.5).clamp(0, 1) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + samples = samples.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + samples = self.numpy_to_pil(samples) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (samples,) + + return ImagePipelineOutput(images=samples) diff --git a/diffusers/pipelines/dreamlite/__init__.py b/diffusers/pipelines/dreamlite/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..01a0609265c1f07991939a904a79490178bfd699 --- /dev/null +++ b/diffusers/pipelines/dreamlite/__init__.py @@ -0,0 +1,42 @@ +# Copyright (c) 2026 ByteDance Ltd. and/or its affiliates. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + _LazyModule, +) + + +_import_structure = { + "pipeline_dreamlite": ["DreamLitePipeline"], + "pipeline_dreamlite_mobile": ["DreamLiteMobilePipeline"], + "pipeline_output": ["DreamLitePipelineOutput"], +} + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + from .pipeline_dreamlite import DreamLitePipeline + from .pipeline_dreamlite_mobile import DreamLiteMobilePipeline + from .pipeline_output import DreamLitePipelineOutput +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) diff --git a/diffusers/pipelines/dreamlite/pipeline_dreamlite.py b/diffusers/pipelines/dreamlite/pipeline_dreamlite.py new file mode 100644 index 0000000000000000000000000000000000000000..e9a0e3c2a767f22a4750b23cc32525832bf8fe5c --- /dev/null +++ b/diffusers/pipelines/dreamlite/pipeline_dreamlite.py @@ -0,0 +1,592 @@ +# Copyright (c) 2026 ByteDance Ltd. and/or its affiliates. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import List, Optional, Union + +import numpy as np +import torch +from PIL import Image +from torch.nn.utils.rnn import pad_sequence +from transformers import AutoTokenizer, Qwen3VLForConditionalGeneration, Qwen3VLProcessor + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin, TextualInversionLoaderMixin +from ...models import AutoencoderTiny +from ...models.unets.unet_dreamlite import DreamLiteUNetModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import DreamLitePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from PIL import Image + >>> from diffusers import DreamLitePipeline + + >>> pipe = DreamLitePipeline.from_pretrained( + ... "carlofkl/DreamLite-base", revision="diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> # Text-to-image + >>> image = pipe(prompt="A serene mountain lake at sunrise").images[0] + + >>> # Image-to-image (instruction-based edit) + >>> init_image = Image.open("input.png").convert("RGB") + >>> edited = pipe(prompt="make it snowy", image=init_image).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def calculate_shift( + image_seq_len: int, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.16, +) -> float: + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.flux.pipeline_flux.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class DreamLitePipeline(DiffusionPipeline, FromSingleFileMixin, TextualInversionLoaderMixin): + r"""DreamLite pipeline for text-to-image and instruction-based image editing. + + The same pipeline supports both modes; the operating mode is auto-detected from the inputs: + + * ``image is None`` -> text-to-image (single CFG on text). + * ``image is not None`` -> image-to-image / instruction edit (dual CFG: text + image). + + Components: + text_encoder ([`~transformers.Qwen3VLForConditionalGeneration`]): + Multimodal text/vision encoder used to produce conditioning embeddings. + tokenizer ([`~transformers.AutoTokenizer`]): + Tokenizer for text-only (generate) mode. + processor ([`~transformers.Qwen3VLProcessor`]): + Multimodal processor for edit mode (text + image template). + vae ([`~diffusers.AutoencoderTiny`]): + Mobile-friendly tiny VAE for latent encode/decode. + unet ([`~diffusers.DreamLiteUNetModel`]): + DreamLite UNet (GQA + qk_norm + depthwise-separable convs). + scheduler ([`~diffusers.FlowMatchEulerDiscreteScheduler`]): + Flow-matching Euler scheduler with dynamic shift. + + Note: + ``batch_size`` is currently forced to ``1``; ``num_images_per_prompt`` is supported. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + text_encoder: Qwen3VLForConditionalGeneration, + tokenizer: AutoTokenizer, + processor: Qwen3VLProcessor, + vae: AutoencoderTiny, + unet: DreamLiteUNetModel, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + processor=processor, + vae=vae, + unet=unet, + scheduler=scheduler, + ) + + # Safe VAE scale factor: AutoencoderTiny exposes `encoder_block_out_channels`; fall back to 8. + if self.vae is not None and hasattr(self.vae.config, "encoder_block_out_channels"): + self.vae_scale_factor = 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + else: + self.vae_scale_factor = 8 + + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.default_sample_size = 128 + + # ----- Prompt encoding templates ----- + # ``prompt_template_encode_*`` is the chat template wrapped around the user prompt before tokenisation. + # ``prompt_template_encode_*_start_idx`` is the number of tokens occupied by the template prefix + # (system + chat-template scaffolding) that must be dropped from the encoder hidden states so the cross- + # attention only attends to the **user prompt** content. The values come from running each template (with + # an empty prompt) through the matching tokenizer / processor and recording the resulting prefix length; + # they are pinned here for reproducibility, mirroring the pattern used by Qwen-Image pipelines. + self.prompt_template_encode_generate = ( + "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, " + "quantity, text, spatial relationships of the objects and background:<|im_end|>\n" + "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + ) + self.prompt_template_encode_generate_start_idx = 34 + self.prompt_template_encode_edit = ( + "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, " + "texture, objects, background), then explain how the user's text instruction should alter " + "or modify the image. Generate a new image that meets the user's requirements while maintaining " + "consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n" + "<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n" + ) + self.prompt_template_encode_edit_start_idx = 64 + + # --------------------------------------------------------------------- + # Helpers + # --------------------------------------------------------------------- + @staticmethod + def _extract_masked_hidden(hidden_states: torch.Tensor, mask: torch.Tensor) -> List[torch.Tensor]: + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1).tolist() + selected = hidden_states[bool_mask] + return torch.split(selected, valid_lengths, dim=0) + + def encode_prompt( + self, + mode: str, + prompts: List[str], + device: torch.device, + dtype: torch.dtype, + image: Optional[Image.Image] = None, + max_sequence_length: int = 500, + text_pad_embedding: Optional[torch.Tensor] = None, + ): + if mode == "edit": + template = self.prompt_template_encode_edit + drop_idx = self.prompt_template_encode_edit_start_idx + + txts = [template.format(p) for p in prompts] + # ``VaeImageProcessor.resize`` defaults to LANCZOS resampling, matching the reference preprocessing + # exactly while avoiding a bespoke ``Image.resize`` call. + cond_image = self.image_processor.resize(image, height=512, width=512) + images = [cond_image] * len(prompts) + + tk_out = self.processor(text=txts, images=images, padding=True, return_tensors="pt").to(device) + + # Pass all processor outputs (input_ids, attention_mask, pixel_values, + # image_grid_thw, mm_token_type_ids, …) to the text encoder so that + # newly-added fields (e.g. mm_token_type_ids for M-RoPE) are forwarded + # automatically. + outputs = self.text_encoder(**tk_out, output_hidden_states=True) + + elif mode == "generate": + template = self.prompt_template_encode_generate + drop_idx = self.prompt_template_encode_generate_start_idx + + txts = [template.format(p) for p in prompts] + tk_out = self.tokenizer( + text=txts, + max_length=max_sequence_length + drop_idx, + padding=True, + truncation=True, + return_tensors="pt", + ).to(device) + + outputs = self.text_encoder(**tk_out, output_hidden_states=True) + else: + raise ValueError(f"Unknown mode: {mode!r}; expected 'generate' or 'edit'.") + + hidden_states = outputs.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, tk_out.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + + prompt_embeds = pad_sequence(split_hidden_states, batch_first=True, padding_value=0).to( + dtype=dtype, device=device + ) + + B, L, _ = prompt_embeds.shape + prompt_embeds_mask = torch.zeros((B, L), dtype=torch.long, device=device) + for i, seq in enumerate(split_hidden_states): + prompt_embeds_mask[i, : seq.shape[0]] = 1 + + if text_pad_embedding is not None: + pad_emb = text_pad_embedding.to(dtype=dtype, device=device) + if pad_emb.ndim == 1: + pad_emb = pad_emb.unsqueeze(0).unsqueeze(0) + elif pad_emb.ndim == 2: + pad_emb = pad_emb.unsqueeze(0) + + mask_expanded = prompt_embeds_mask.unsqueeze(-1).to(dtype=dtype) + prompt_embeds = prompt_embeds * mask_expanded + pad_emb * (1 - mask_expanded) + + return prompt_embeds, prompt_embeds_mask + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[torch.Generator], + latents: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + height = int(height) // self.vae_scale_factor + width = int(width) // self.vae_scale_factor + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError("Generator list length must match batch size.") + + return randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + def prepare_image_latents( + self, + image: Union[torch.Tensor, Image.Image, List[Image.Image]], + dtype: torch.dtype, + device: torch.device, + generator: Optional[torch.Generator] = None, + ) -> torch.Tensor: + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax") + + return image_latents + + def check_inputs( + self, + prompt: Optional[str], + image: Optional[Union[torch.Tensor, Image.Image, List[Image.Image]]], + height: Optional[int], + width: Optional[int], + ): + if prompt is not None and not isinstance(prompt, str): + raise ValueError(f"`prompt` has to be of type `str` but is {type(prompt)}") + + if image is not None and not isinstance(image, (torch.Tensor, Image.Image, list)): + raise ValueError(f"`image` must be of type `torch.Tensor`, `PIL.Image.Image` or `list`, got {type(image)}") + + if (height is not None and height % self.vae_scale_factor != 0) or ( + width is not None and width % self.vae_scale_factor != 0 + ): + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor} but are {height} and {width}. " + "Dimensions will be resized accordingly." + ) + + # --------------------------------------------------------------------- + # Properties + # --------------------------------------------------------------------- + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def image_guidance_scale(self): + return self._image_guidance_scale + + # --------------------------------------------------------------------- + # Main entry + # --------------------------------------------------------------------- + @torch.no_grad() + def __call__( + self, + prompt: Optional[str] = None, + negative_prompt: Optional[str] = None, + image: Optional[Image.Image] = None, + height: Optional[int] = None, + width: Optional[int] = None, + guidance_scale: float = 3.5, + image_guidance_scale: float = 1.5, + num_inference_steps: int = 30, + sigmas: Optional[List[float]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + max_sequence_length: int = 200, + text_pad_embedding: Optional[torch.Tensor] = None, + ): + r"""Run the DreamLite pipeline. + + Args: + prompt: Text prompt. + negative_prompt: Negative text prompt (defaults to empty string). + image: Optional input image. If provided, the pipeline runs in **edit / image-to-image** mode + with dual classifier-free guidance; otherwise it runs in **text-to-image** mode. + height: Output resolution (height). Defaults to ``default_sample_size * vae_scale_factor`` (1024). + The same default applies in both T2I and I2I; pass an explicit value to override. + width: Output resolution (width). Defaults to ``default_sample_size * vae_scale_factor`` (1024). + The same default applies in both T2I and I2I; pass an explicit value to override. + guidance_scale: CFG scale on the text branch (both modes). + image_guidance_scale: Additional CFG scale on the image branch (edit mode only). + num_inference_steps: Number of denoising steps. + sigmas: Optional explicit FlowMatch sigmas; defaults to a uniform linspace. + num_images_per_prompt: Output images per prompt (note: ``batch_size`` is forced to 1). + generator: Random generator(s). + output_type: ``"pil"``, ``"np"``, ``"pt"`` or ``"latent"``. + return_dict: If True, returns a :class:`DreamLitePipelineOutput`; else a tuple ``(images,)``. + max_sequence_length: Maximum number of user-prompt tokens kept after dropping the chat-template + prefix. Only applies to ``generate`` mode (the ``edit`` mode uses the multimodal processor's native + padding). + text_pad_embedding: Optional learned pad embedding for masked positions. + + Returns: + :class:`DreamLitePipelineOutput` or ``tuple``. + """ + # 1. Init pipeline parameters + self.check_inputs(prompt, image, height, width) + if height is None and width is None and image is not None: + w, h = image.size + width = (w // self.vae_scale_factor) * self.vae_scale_factor + height = (h // self.vae_scale_factor) * self.vae_scale_factor + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + self._guidance_scale = guidance_scale + self._image_guidance_scale = image_guidance_scale + + task = "generate" if image is None else "edit" + device = self._execution_device + dtype = self.text_encoder.dtype + batch_size = 1 # Note: pipeline currently forces batch_size = 1. + negative_prompt = negative_prompt or "" + + if sigmas is None: + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) + + # 2. Prepare Time IDs (carries original H,W as additional conditioning) + original_size = (width, height) + add_time_ids = torch.tensor([list(original_size)], device=device, dtype=dtype) + + # 3. Prepare Noise Latents (x_t) + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + dtype, + device, + generator, + ) + + # 4. Prepare Timesteps (FlowMatch with dynamic shift) + image_seq_len = latents.shape[2] * latents.shape[3] // 4 + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 5. Prepare Conditions (Text & Image) + if task == "generate": + prompt_str = f"[Generate]: {prompt}" + prompt_embeds, text_attention_mask = self.encode_prompt( + mode="generate", + prompts=[negative_prompt, prompt_str], + device=device, + dtype=dtype, + max_sequence_length=max_sequence_length, + text_pad_embedding=text_pad_embedding, + ) + if num_images_per_prompt > 1: + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_attention_mask = text_attention_mask.repeat_interleave(num_images_per_prompt, dim=0) + image_latents = torch.zeros_like(latents) + else: + prompt_str = ( + f"[Edit]: A diptych with two side-by-side images of the same scene. " + f"Compared to the right side, the left one has {prompt}" + ) + prompt_embeds, text_attention_mask = self.encode_prompt( + mode="edit", + prompts=[negative_prompt, negative_prompt, prompt_str], + image=image, + device=device, + dtype=dtype, + ) + if num_images_per_prompt > 1: + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_attention_mask = text_attention_mask.repeat_interleave(num_images_per_prompt, dim=0) + image_processed = self.image_processor.preprocess(image, height=height, width=width) + image_latents = self.prepare_image_latents( + image_processed, + dtype=dtype, + device=device, + ) + uncond_image_latents = torch.zeros_like(latents) + + # 6. Denoising Loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Expand latents for classifier-free guidance + if task == "generate": + latents_in = torch.cat([latents] * 2) + cond_img_in = torch.cat([image_latents] * 2) + model_input = torch.cat([latents_in, cond_img_in], dim=3) + time_ids_in = torch.cat([add_time_ids] * 2) + else: # edit + latents_in = torch.cat([latents] * 3) + cond_img_in = torch.cat([uncond_image_latents, image_latents, image_latents]) + model_input = torch.cat([latents_in, cond_img_in], dim=3) + time_ids_in = torch.cat([add_time_ids] * 3) + + # UNet Forward + noise_pred = self.unet( + model_input, + timestep=t.expand(model_input.shape[0]).to(latents.dtype), + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=text_attention_mask, + added_cond_kwargs={"time_ids": time_ids_in}, + return_dict=False, + )[0] + + # Classifier-Free Guidance (single for T2I, dual for I2I) + noise_pred = noise_pred[..., : latents.shape[-1]] + if task == "generate": + noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond) + else: # edit + noise_pred_uncond, noise_pred_image, noise_pred_text = noise_pred.chunk(3) + noise_pred = ( + noise_pred_uncond + + self.guidance_scale * (noise_pred_text - noise_pred_image) + + self.image_guidance_scale * (noise_pred_image - noise_pred_uncond) + ) + + # Scheduler Step + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + latents = latents.to(latents_dtype) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 7. Decode Latents + if output_type == "latent": + image_out = latents + else: + shift_factor = getattr(self.vae.config, "shift_factor", 0.0) or 0.0 + latents = (latents / self.vae.config.scaling_factor) + shift_factor + image_out = self.vae.decode(latents, return_dict=False)[0] + image_out = self.image_processor.postprocess(image_out, output_type=output_type) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image_out,) + + return DreamLitePipelineOutput(images=image_out) diff --git a/diffusers/pipelines/dreamlite/pipeline_dreamlite_mobile.py b/diffusers/pipelines/dreamlite/pipeline_dreamlite_mobile.py new file mode 100644 index 0000000000000000000000000000000000000000..ca9e6b7b4c40432c01ded7e902aacba2d24eb13f --- /dev/null +++ b/diffusers/pipelines/dreamlite/pipeline_dreamlite_mobile.py @@ -0,0 +1,569 @@ +# Copyright (c) 2026 ByteDance Ltd. and/or its affiliates. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import List, Optional, Union + +import numpy as np +import torch +from PIL import Image +from torch.nn.utils.rnn import pad_sequence +from transformers import AutoTokenizer, Qwen3VLForConditionalGeneration, Qwen3VLProcessor + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin, TextualInversionLoaderMixin +from ...models import AutoencoderTiny +from ...models.unets.unet_dreamlite import DreamLiteUNetModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import DreamLitePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.dreamlite.pipeline_dreamlite.calculate_shift +def calculate_shift( + image_seq_len: int, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.16, +) -> float: + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.flux.pipeline_flux.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from PIL import Image + >>> from diffusers import DreamLiteMobilePipeline + + >>> pipe = DreamLiteMobilePipeline.from_pretrained( + ... "carlofkl/DreamLite-mobile", revision="diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> # Text-to-image (4 steps, no CFG) + >>> image = pipe(prompt="A serene mountain lake at sunrise").images[0] + + >>> # Image-to-image (instruction-based edit, 4 steps) + >>> init_image = Image.open("input.png").convert("RGB") + >>> edited = pipe(prompt="make it snowy", image=init_image).images[0] + ``` +""" + + +class DreamLiteMobilePipeline(DiffusionPipeline, FromSingleFileMixin, TextualInversionLoaderMixin): + r"""DreamLite **Mobile** pipeline: a distilled, classifier-free-guidance-free variant of + :class:`DreamLitePipeline` for fast few-step inference (default 4 steps). + + The operating mode is auto-detected from inputs (same as the base pipeline): + + * ``image is None`` -> text-to-image. + * ``image is not None`` -> image-to-image / instruction edit. + + Because classifier-free guidance is **distilled away**, ``guidance_scale`` and ``image_guidance_scale`` are + accepted for API parity with :class:`DreamLitePipeline` but are ignored in the denoising loop. ``negative_prompt`` + is intentionally absent. + + Components (identical to the base pipeline): + text_encoder ([`~transformers.Qwen3VLForConditionalGeneration`]): + Multimodal text/vision encoder. + tokenizer ([`~transformers.AutoTokenizer`]): + Tokenizer for text-only (generate) mode. + processor ([`~transformers.Qwen3VLProcessor`]): + Multimodal processor for edit mode. + vae ([`~diffusers.AutoencoderTiny`]): + Mobile-friendly tiny VAE. + unet ([`~diffusers.DreamLiteUNetModel`]): + DreamLite UNet. + scheduler ([`~diffusers.FlowMatchEulerDiscreteScheduler`]): + Flow-matching Euler scheduler with dynamic shift. + + Note: + ``batch_size`` is currently forced to ``1``; ``num_images_per_prompt`` is supported. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + text_encoder: Qwen3VLForConditionalGeneration, + tokenizer: AutoTokenizer, + processor: Qwen3VLProcessor, + vae: AutoencoderTiny, + unet: DreamLiteUNetModel, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + processor=processor, + vae=vae, + unet=unet, + scheduler=scheduler, + ) + + # Safe VAE scale factor: AutoencoderTiny exposes `encoder_block_out_channels`; fall back to 8. + if self.vae is not None and hasattr(self.vae.config, "encoder_block_out_channels"): + self.vae_scale_factor = 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + else: + self.vae_scale_factor = 8 + + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.default_sample_size = 128 + + # ----- Prompt encoding templates ----- + # See ``DreamLitePipeline.__init__`` for the meaning of these template strings and their associated + # ``*_start_idx`` token-prefix offsets. + self.prompt_template_encode_generate = ( + "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, " + "quantity, text, spatial relationships of the objects and background:<|im_end|>\n" + "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + ) + self.prompt_template_encode_generate_start_idx = 34 + self.prompt_template_encode_edit = ( + "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, " + "texture, objects, background), then explain how the user's text instruction should alter " + "or modify the image. Generate a new image that meets the user's requirements while maintaining " + "consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n" + "<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n" + ) + self.prompt_template_encode_edit_start_idx = 64 + + # --------------------------------------------------------------------- + # Helpers (identical to DreamLitePipeline) + # --------------------------------------------------------------------- + @staticmethod + # Copied from diffusers.pipelines.dreamlite.pipeline_dreamlite.DreamLitePipeline._extract_masked_hidden + def _extract_masked_hidden(hidden_states: torch.Tensor, mask: torch.Tensor) -> List[torch.Tensor]: + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1).tolist() + selected = hidden_states[bool_mask] + return torch.split(selected, valid_lengths, dim=0) + + # Copied from diffusers.pipelines.dreamlite.pipeline_dreamlite.DreamLitePipeline.encode_prompt + def encode_prompt( + self, + mode: str, + prompts: List[str], + device: torch.device, + dtype: torch.dtype, + image: Optional[Image.Image] = None, + max_sequence_length: int = 500, + text_pad_embedding: Optional[torch.Tensor] = None, + ): + if mode == "edit": + template = self.prompt_template_encode_edit + drop_idx = self.prompt_template_encode_edit_start_idx + + txts = [template.format(p) for p in prompts] + # ``VaeImageProcessor.resize`` defaults to LANCZOS resampling, matching the reference preprocessing + # exactly while avoiding a bespoke ``Image.resize`` call. + cond_image = self.image_processor.resize(image, height=512, width=512) + images = [cond_image] * len(prompts) + + tk_out = self.processor(text=txts, images=images, padding=True, return_tensors="pt").to(device) + + # Pass all processor outputs (input_ids, attention_mask, pixel_values, + # image_grid_thw, mm_token_type_ids, …) to the text encoder so that + # newly-added fields (e.g. mm_token_type_ids for M-RoPE) are forwarded + # automatically. + outputs = self.text_encoder(**tk_out, output_hidden_states=True) + + elif mode == "generate": + template = self.prompt_template_encode_generate + drop_idx = self.prompt_template_encode_generate_start_idx + + txts = [template.format(p) for p in prompts] + tk_out = self.tokenizer( + text=txts, + max_length=max_sequence_length + drop_idx, + padding=True, + truncation=True, + return_tensors="pt", + ).to(device) + + outputs = self.text_encoder(**tk_out, output_hidden_states=True) + else: + raise ValueError(f"Unknown mode: {mode!r}; expected 'generate' or 'edit'.") + + hidden_states = outputs.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, tk_out.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + + prompt_embeds = pad_sequence(split_hidden_states, batch_first=True, padding_value=0).to( + dtype=dtype, device=device + ) + + B, L, _ = prompt_embeds.shape + prompt_embeds_mask = torch.zeros((B, L), dtype=torch.long, device=device) + for i, seq in enumerate(split_hidden_states): + prompt_embeds_mask[i, : seq.shape[0]] = 1 + + if text_pad_embedding is not None: + pad_emb = text_pad_embedding.to(dtype=dtype, device=device) + if pad_emb.ndim == 1: + pad_emb = pad_emb.unsqueeze(0).unsqueeze(0) + elif pad_emb.ndim == 2: + pad_emb = pad_emb.unsqueeze(0) + + mask_expanded = prompt_embeds_mask.unsqueeze(-1).to(dtype=dtype) + prompt_embeds = prompt_embeds * mask_expanded + pad_emb * (1 - mask_expanded) + + return prompt_embeds, prompt_embeds_mask + + # Copied from diffusers.pipelines.dreamlite.pipeline_dreamlite.DreamLitePipeline.prepare_latents + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[torch.Generator], + latents: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + height = int(height) // self.vae_scale_factor + width = int(width) // self.vae_scale_factor + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError("Generator list length must match batch size.") + + return randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # Copied from diffusers.pipelines.dreamlite.pipeline_dreamlite.DreamLitePipeline.prepare_image_latents + def prepare_image_latents( + self, + image: Union[torch.Tensor, Image.Image, List[Image.Image]], + dtype: torch.dtype, + device: torch.device, + generator: Optional[torch.Generator] = None, + ) -> torch.Tensor: + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax") + + return image_latents + + # Copied from diffusers.pipelines.dreamlite.pipeline_dreamlite.DreamLitePipeline.check_inputs + def check_inputs( + self, + prompt: Optional[str], + image: Optional[Union[torch.Tensor, Image.Image, List[Image.Image]]], + height: Optional[int], + width: Optional[int], + ): + if prompt is not None and not isinstance(prompt, str): + raise ValueError(f"`prompt` has to be of type `str` but is {type(prompt)}") + + if image is not None and not isinstance(image, (torch.Tensor, Image.Image, list)): + raise ValueError(f"`image` must be of type `torch.Tensor`, `PIL.Image.Image` or `list`, got {type(image)}") + + if (height is not None and height % self.vae_scale_factor != 0) or ( + width is not None and width % self.vae_scale_factor != 0 + ): + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor} but are {height} and {width}. " + "Dimensions will be resized accordingly." + ) + + # --------------------------------------------------------------------- + # Main entry + # --------------------------------------------------------------------- + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: Optional[Image.Image] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 4, + guidance_scale: Optional[float] = None, + image_guidance_scale: Optional[float] = None, + sigmas: Optional[List[float]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + max_sequence_length: int = 200, + text_pad_embedding: Optional[torch.Tensor] = None, + ): + r"""Run the distilled DreamLite Mobile pipeline. + + Args: + prompt: Text prompt. + image: Optional input image. If provided, runs in **edit / image-to-image** mode; + otherwise runs in **text-to-image** mode. + height: Output resolution (height). Defaults to ``default_sample_size * vae_scale_factor`` (1024). + width: Output resolution (width). Defaults to ``default_sample_size * vae_scale_factor`` (1024). + num_inference_steps: Number of denoising steps. Defaults to **4** (distilled). + guidance_scale: Accepted for API parity with :class:`DreamLitePipeline`; **ignored** + because CFG was distilled away. + image_guidance_scale: Accepted for API parity with :class:`DreamLitePipeline`; **ignored** + because CFG was distilled away. + sigmas: Optional explicit FlowMatch sigmas; defaults to a uniform linspace. + num_images_per_prompt: Output images per prompt (note: ``batch_size`` is forced to 1). + generator: Random generator(s). + output_type: ``"pil"``, ``"np"``, ``"pt"`` or ``"latent"``. + return_dict: If True, returns a :class:`DreamLitePipelineOutput`; else ``(images,)``. + max_sequence_length: Maximum number of user-prompt tokens kept after dropping the chat-template + prefix. Only applies to ``generate`` mode (the ``edit`` mode uses the multimodal processor's native + padding). + text_pad_embedding: Optional learned pad embedding for masked positions. + + Returns: + :class:`DreamLitePipelineOutput` or ``tuple``. + """ + # 1. Init pipeline parameters + self.check_inputs(prompt, image, height, width) + if height is None and width is None and image is not None: + w, h = image.size + width = (w // self.vae_scale_factor) * self.vae_scale_factor + height = (h // self.vae_scale_factor) * self.vae_scale_factor + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + task = "generate" if image is None else "edit" + device = self._execution_device + dtype = self.text_encoder.dtype + batch_size = 1 # Note: pipeline currently forces batch_size = 1. + + if guidance_scale is not None or image_guidance_scale is not None: + logger.warning( + "`guidance_scale` / `image_guidance_scale` are ignored by DreamLiteMobilePipeline " + "because classifier-free guidance was distilled away." + ) + + if sigmas is None: + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) + + # 2. Prepare Time IDs + original_size = (width, height) + add_time_ids = torch.tensor([list(original_size)], device=device, dtype=dtype) + + # 3. Prepare Noise Latents (x_t) + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + dtype, + device, + generator, + ) + + # 4. Prepare Timesteps (FlowMatch with dynamic shift) + image_seq_len = latents.shape[2] * latents.shape[3] // 4 + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 5. Prepare Conditions (Text & Image) — no negatives because CFG is distilled away + if task == "generate": + prompt_str = f"[Generate]: {prompt}" + prompt_embeds, text_attention_mask = self.encode_prompt( + mode="generate", + prompts=[prompt_str], + device=device, + dtype=dtype, + max_sequence_length=max_sequence_length, + text_pad_embedding=text_pad_embedding, + ) + if num_images_per_prompt > 1: + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_attention_mask = text_attention_mask.repeat_interleave(num_images_per_prompt, dim=0) + image_latents = torch.zeros_like(latents) + else: + prompt_str = ( + f"[Edit]: A diptych with two side-by-side images of the same scene. " + f"Compared to the right side, the left one has {prompt}" + ) + prompt_embeds, text_attention_mask = self.encode_prompt( + mode="edit", + prompts=[prompt_str], + image=image, + device=device, + dtype=dtype, + ) + if num_images_per_prompt > 1: + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_attention_mask = text_attention_mask.repeat_interleave(num_images_per_prompt, dim=0) + image_processed = self.image_processor.preprocess(image, height=height, width=width) + image_latents = self.prepare_image_latents( + image_processed, + dtype=dtype, + device=device, + ) + + # 6. Denoising Loop (no CFG: single forward per step, no cat/chunk) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + model_input = torch.cat([latents, image_latents], dim=3) + time_ids_in = add_time_ids + + # UNet Forward + noise_pred = self.unet( + model_input, + timestep=t.expand(model_input.shape[0]).to(latents.dtype), + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=text_attention_mask, + added_cond_kwargs={"time_ids": time_ids_in}, + return_dict=False, + )[0] + + # Drop extra channels (image-conditioning half of the spatial concat) + noise_pred = noise_pred[..., : latents.shape[-1]] + + # Scheduler Step + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + latents = latents.to(latents_dtype) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 7. Decode Latents + if output_type == "latent": + image_out = latents + else: + shift_factor = getattr(self.vae.config, "shift_factor", 0.0) or 0.0 + latents = (latents / self.vae.config.scaling_factor) + shift_factor + image_out = self.vae.decode(latents, return_dict=False)[0] + image_out = self.image_processor.postprocess(image_out, output_type=output_type) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image_out,) + + return DreamLitePipelineOutput(images=image_out) diff --git a/diffusers/pipelines/dreamlite/pipeline_output.py b/diffusers/pipelines/dreamlite/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..6d4cf15aed6ec705b5d0197033a4aeeab255fdeb --- /dev/null +++ b/diffusers/pipelines/dreamlite/pipeline_output.py @@ -0,0 +1,35 @@ +# Copyright (c) 2026 ByteDance Ltd. and/or its affiliates. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import List, Union + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class DreamLitePipelineOutput(BaseOutput): + """ + Output class for DreamLite pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`): + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. PIL images or NumPy array present the denoised images of the diffusion pipeline. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] diff --git a/diffusers/pipelines/easyanimate/__init__.py b/diffusers/pipelines/easyanimate/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..49923423f951c1e6124ddb9b0506f740f28a7a5a --- /dev/null +++ b/diffusers/pipelines/easyanimate/__init__.py @@ -0,0 +1,52 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_easyanimate"] = ["EasyAnimatePipeline"] + _import_structure["pipeline_easyanimate_control"] = ["EasyAnimateControlPipeline"] + _import_structure["pipeline_easyanimate_inpaint"] = ["EasyAnimateInpaintPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_easyanimate import EasyAnimatePipeline + from .pipeline_easyanimate_control import EasyAnimateControlPipeline + from .pipeline_easyanimate_inpaint import EasyAnimateInpaintPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/easyanimate/pipeline_easyanimate.py b/diffusers/pipelines/easyanimate/pipeline_easyanimate.py new file mode 100644 index 0000000000000000000000000000000000000000..72e19a8cce1fc10b82f2f7ab7fc3e54e835e5961 --- /dev/null +++ b/diffusers/pipelines/easyanimate/pipeline_easyanimate.py @@ -0,0 +1,773 @@ +# Copyright 2025 The EasyAnimate team and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import torch +from transformers import ( + BertModel, + BertTokenizer, + Qwen2Tokenizer, + Qwen2VLForConditionalGeneration, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...models import AutoencoderKLMagvit, EasyAnimateTransformer3DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from .pipeline_output import EasyAnimatePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import EasyAnimatePipeline + >>> from diffusers.utils import export_to_video + + >>> # Models: "alibaba-pai/EasyAnimateV5.1-12b-zh" + >>> pipe = EasyAnimatePipeline.from_pretrained( + ... "alibaba-pai/EasyAnimateV5.1-7b-zh-diffusers", torch_dtype=torch.float16 + ... ).to("cuda") + >>> prompt = ( + ... "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. " + ... "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other " + ... "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, " + ... "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. " + ... "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical " + ... "atmosphere of this unique musical performance." + ... ) + >>> sample_size = (512, 512) + >>> video = pipe( + ... prompt=prompt, + ... guidance_scale=6, + ... negative_prompt="bad detailed", + ... height=sample_size[0], + ... width=sample_size[1], + ... num_inference_steps=50, + ... ).frames[0] + >>> export_to_video(video, "output.mp4", fps=8) + ``` +""" + + +# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid +def get_resize_crop_region_for_grid(src, tgt_width, tgt_height): + tw = tgt_width + th = tgt_height + h, w = src + r = h / w + if r > (th / tw): + resize_height = th + resize_width = int(round(th / h * w)) + else: + resize_width = tw + resize_height = int(round(tw / w * h)) + + crop_top = int(round((th - resize_height) / 2.0)) + crop_left = int(round((tw - resize_width) / 2.0)) + + return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class EasyAnimatePipeline(DiffusionPipeline): + r""" + Pipeline for text-to-video generation using EasyAnimate. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + EasyAnimate uses one text encoder [qwen2 vl](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) in V5.1. + + Args: + vae ([`AutoencoderKLMagvit`]): + Variational Auto-Encoder (VAE) Model to encode and decode video to and from latent representations. + text_encoder (`~transformers.Qwen2VLForConditionalGeneration`, `~transformers.BertModel` | None): + EasyAnimate uses [qwen2 vl](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) in V5.1. + tokenizer (`~transformers.Qwen2Tokenizer`, `~transformers.BertTokenizer` | None): + A `Qwen2Tokenizer` or `BertTokenizer` to tokenize text. + transformer ([`EasyAnimateTransformer3DModel`]): + The EasyAnimate model designed by EasyAnimate Team. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with EasyAnimate to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKLMagvit, + text_encoder: Qwen2VLForConditionalGeneration | BertModel, + tokenizer: Qwen2Tokenizer | BertTokenizer, + transformer: EasyAnimateTransformer3DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.enable_text_attention_mask = ( + self.transformer.config.enable_text_attention_mask + if getattr(self, "transformer", None) is not None + else True + ) + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 4 + ) + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + + def encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + dtype (`torch.dtype`): + torch dtype + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds` is passed directly. + max_sequence_length (`int`, *optional*): maximum sequence length to use for the prompt. + """ + dtype = dtype or self.text_encoder.dtype + device = device or self.text_encoder.device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + if isinstance(prompt, str): + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": prompt}], + } + ] + else: + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": _prompt}], + } + for _prompt in prompt + ] + text = [ + self.tokenizer.apply_chat_template([m], tokenize=False, add_generation_prompt=True) for m in messages + ] + + text_inputs = self.tokenizer( + text=text, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_attention_mask=True, + padding_side="right", + return_tensors="pt", + ) + text_inputs = text_inputs.to(self.text_encoder.device) + + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + if self.enable_text_attention_mask: + # Inference: Generation of the output + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ).hidden_states[-2] + else: + raise ValueError("LLM needs attention_mask") + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.to(device=device) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + if negative_prompt is not None and isinstance(negative_prompt, str): + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": negative_prompt}], + } + ] + else: + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": _negative_prompt}], + } + for _negative_prompt in negative_prompt + ] + text = [ + self.tokenizer.apply_chat_template([m], tokenize=False, add_generation_prompt=True) for m in messages + ] + + text_inputs = self.tokenizer( + text=text, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_attention_mask=True, + padding_side="right", + return_tensors="pt", + ) + text_inputs = text_inputs.to(self.text_encoder.device) + + text_input_ids = text_inputs.input_ids + negative_prompt_attention_mask = text_inputs.attention_mask + if self.enable_text_attention_mask: + # Inference: Generation of the output + negative_prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=negative_prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-2] + else: + raise ValueError("LLM needs attention_mask") + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.to(device=device) + + return prompt_embeds, negative_prompt_embeds, prompt_attention_mask, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents( + self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None + ): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + (num_frames - 1) // self.vae_temporal_compression_ratio + 1, + height // self.vae_spatial_compression_ratio, + width // self.vae_spatial_compression_ratio, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # scale the initial noise by the standard deviation required by the scheduler + if hasattr(self.scheduler, "init_noise_sigma"): + latents = latents * self.scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + num_frames: int | None = 49, + height: int | None = 512, + width: int | None = 512, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + timesteps: list[int] | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + guidance_rescale: float = 0.0, + ): + r""" + Generates images or video using the EasyAnimate pipeline based on the provided prompts. + + Args: + prompt (`str` or `list[str]`, *optional*): + Text prompts to guide the image or video generation. If not provided, use `prompt_embeds` instead. + num_frames (`int`, *optional*): + Length of the generated video (in frames). + height (`int`, *optional*): + Height of the generated image in pixels. + width (`int`, *optional*): + Width of the generated image in pixels. + num_inference_steps (`int`, *optional*, defaults to 50): + Number of denoising steps during generation. More steps generally yield higher quality images but slow + down inference. + guidance_scale (`float`, *optional*, defaults to 5.0): + Encourages the model to align outputs with prompts. A higher value may decrease image quality. + negative_prompt (`str` or `list[str]`, *optional*): + Prompts indicating what to exclude in generation. If not specified, use `negative_prompt_embeds`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + Number of images to generate for each prompt. + eta (`float`, *optional*, defaults to 0.0): + Applies to DDIM scheduling. Controlled by the eta parameter from the related literature. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A generator to ensure reproducibility in image generation. + latents (`torch.Tensor`, *optional*): + Predefined latent tensors to condition generation. + prompt_embeds (`torch.Tensor`, *optional*): + Text embeddings for the prompts. Overrides prompt string inputs for more flexibility. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Embeddings for negative prompts. Overrides string inputs if defined. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the primary prompt embeddings. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for negative prompt embeddings. + output_type (`str`, *optional*, defaults to "latent"): + Format of the generated output, either as a PIL image or as a NumPy array. + return_dict (`bool`, *optional*, defaults to `True`): + If `True`, returns a structured output. Otherwise returns a simple tuple. + callback_on_step_end (`Callable`, *optional*): + Functions called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + Tensor names to be included in callback function calls. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Adjusts noise levels based on guidance scale. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, the scheduler's default schedule for + `num_inference_steps` is used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. default height and width + height = int((height // 16) * 16) + width = int((width // 16) * 16) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = self.transformer.dtype + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + device=device, + dtype=dtype, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler): + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, mu=1 + ) + else: + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps + ) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + num_frames, + height, + width, + dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask]) + + prompt_embeds = prompt_embeds.to(device=device) + prompt_attention_mask = prompt_attention_mask.to(device=device) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + if hasattr(self.scheduler, "scale_model_input"): + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # expand scalar t to 1-D tensor to match the 1st dim of latent_model_input + t_expand = torch.tensor([t] * latent_model_input.shape[0], device=device).to( + dtype=latent_model_input.dtype + ) + + # predict the noise residual + noise_pred = self.transformer( + latent_model_input, + t_expand, + encoder_hidden_states=prompt_embeds, + return_dict=False, + )[0] + + if noise_pred.size()[1] != self.vae.config.latent_channels: + noise_pred, _ = noise_pred.chunk(2, dim=1) + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + latents = 1 / self.vae.config.scaling_factor * latents + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video=video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return EasyAnimatePipelineOutput(frames=video) diff --git a/diffusers/pipelines/easyanimate/pipeline_easyanimate_control.py b/diffusers/pipelines/easyanimate/pipeline_easyanimate_control.py new file mode 100644 index 0000000000000000000000000000000000000000..4ad3a48b70ec31bf3b34456a40b48fd5b20880a0 --- /dev/null +++ b/diffusers/pipelines/easyanimate/pipeline_easyanimate_control.py @@ -0,0 +1,1009 @@ +# Copyright 2025 The EasyAnimate team and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import torch +import torch.nn.functional as F +from PIL import Image +from transformers import ( + BertModel, + BertTokenizer, + Qwen2Tokenizer, + Qwen2VLForConditionalGeneration, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKLMagvit, EasyAnimateTransformer3DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from .pipeline_output import EasyAnimatePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import EasyAnimateControlPipeline + >>> from diffusers.pipelines.easyanimate.pipeline_easyanimate_control import get_video_to_video_latent + >>> from diffusers.utils import export_to_video, load_video + + >>> pipe = EasyAnimateControlPipeline.from_pretrained( + ... "alibaba-pai/EasyAnimateV5.1-12b-zh-Control-diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> control_video = load_video( + ... "https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-Control/blob/main/asset/pose.mp4" + ... ) + >>> prompt = ( + ... "In this sunlit outdoor garden, a beautiful woman is dressed in a knee-length, sleeveless white dress. " + ... "The hem of her dress gently sways with her graceful dance, much like a butterfly fluttering in the breeze. " + ... "Sunlight filters through the leaves, casting dappled shadows that highlight her soft features and clear eyes, " + ... "making her appear exceptionally elegant. It seems as if every movement she makes speaks of youth and vitality. " + ... "As she twirls on the grass, her dress flutters, as if the entire garden is rejoicing in her dance. " + ... "The colorful flowers around her sway in the gentle breeze, with roses, chrysanthemums, and lilies each " + ... "releasing their fragrances, creating a relaxed and joyful atmosphere." + ... ) + >>> sample_size = (672, 384) + >>> num_frames = 49 + + >>> input_video, _, _ = get_video_to_video_latent(control_video, num_frames, sample_size) + >>> video = pipe( + ... prompt, + ... num_frames=num_frames, + ... negative_prompt="Twisted body, limb deformities, text subtitles, comics, stillness, ugliness, errors, garbled text.", + ... height=sample_size[0], + ... width=sample_size[1], + ... control_video=input_video, + ... ).frames[0] + >>> export_to_video(video, "output.mp4", fps=8) + ``` +""" + + +def preprocess_image(image, sample_size): + """ + Preprocess a single image (PIL.Image, numpy.ndarray, or torch.Tensor) to a resized tensor. + """ + if isinstance(image, torch.Tensor): + # If input is a tensor, assume it's in CHW format and resize using interpolation + image = torch.nn.functional.interpolate( + image.unsqueeze(0), size=sample_size, mode="bilinear", align_corners=False + ).squeeze(0) + elif isinstance(image, Image.Image): + # If input is a PIL image, resize and convert to numpy array + image = image.resize((sample_size[1], sample_size[0])) + image = np.array(image) + elif isinstance(image, np.ndarray): + # If input is a numpy array, resize using PIL + image = Image.fromarray(image).resize((sample_size[1], sample_size[0])) + image = np.array(image) + else: + raise ValueError("Unsupported input type. Expected PIL.Image, numpy.ndarray, or torch.Tensor.") + + # Convert to tensor if not already + if not isinstance(image, torch.Tensor): + image = torch.from_numpy(image).permute(2, 0, 1).float() / 255.0 # HWC -> CHW, normalize to [0, 1] + + return image + + +def get_video_to_video_latent(input_video, num_frames, sample_size, validation_video_mask=None, ref_image=None): + if input_video is not None: + # Convert each frame in the list to tensor + input_video = [preprocess_image(frame, sample_size=sample_size) for frame in input_video] + + # Stack all frames into a single tensor (F, C, H, W) + input_video = torch.stack(input_video)[:num_frames] + + # Add batch dimension (B, F, C, H, W) + input_video = input_video.permute(1, 0, 2, 3).unsqueeze(0) + + if validation_video_mask is not None: + # Handle mask input + validation_video_mask = preprocess_image(validation_video_mask, size=sample_size) + input_video_mask = torch.where(validation_video_mask < 240 / 255.0, 0.0, 255) + + # Adjust mask dimensions to match video + input_video_mask = input_video_mask.unsqueeze(0).unsqueeze(-1).permute([3, 0, 1, 2]).unsqueeze(0) + input_video_mask = torch.tile(input_video_mask, [1, 1, input_video.size()[2], 1, 1]) + input_video_mask = input_video_mask.to(input_video.device, input_video.dtype) + else: + input_video_mask = torch.zeros_like(input_video[:, :1]) + input_video_mask[:, :, :] = 255 + else: + input_video, input_video_mask = None, None + + if ref_image is not None: + # Convert reference image to tensor + ref_image = preprocess_image(ref_image, size=sample_size) + ref_image = ref_image.permute(1, 0, 2, 3).unsqueeze(0) # Add batch dimension (B, C, H, W) + else: + ref_image = None + + return input_video, input_video_mask, ref_image + + +# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid +def get_resize_crop_region_for_grid(src, tgt_width, tgt_height): + tw = tgt_width + th = tgt_height + h, w = src + r = h / w + if r > (th / tw): + resize_height = th + resize_width = int(round(th / h * w)) + else: + resize_width = tw + resize_height = int(round(tw / w * h)) + + crop_top = int(round((th - resize_height) / 2.0)) + crop_left = int(round((tw - resize_width) / 2.0)) + + return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Resize mask information in magvit +def resize_mask(mask, latent, process_first_frame_only=True): + latent_size = latent.size() + + if process_first_frame_only: + target_size = list(latent_size[2:]) + target_size[0] = 1 + first_frame_resized = F.interpolate( + mask[:, :, 0:1, :, :], size=target_size, mode="trilinear", align_corners=False + ) + + target_size = list(latent_size[2:]) + target_size[0] = target_size[0] - 1 + if target_size[0] != 0: + remaining_frames_resized = F.interpolate( + mask[:, :, 1:, :, :], size=target_size, mode="trilinear", align_corners=False + ) + resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2) + else: + resized_mask = first_frame_resized + else: + target_size = list(latent_size[2:]) + resized_mask = F.interpolate(mask, size=target_size, mode="trilinear", align_corners=False) + return resized_mask + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class EasyAnimateControlPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-video generation using EasyAnimate. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + EasyAnimate uses one text encoder [qwen2 vl](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) in V5.1. + + Args: + vae ([`AutoencoderKLMagvit`]): + Variational Auto-Encoder (VAE) Model to encode and decode video to and from latent representations. + text_encoder (`~transformers.Qwen2VLForConditionalGeneration`, `~transformers.BertModel` | None): + EasyAnimate uses [qwen2 vl](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) in V5.1. + tokenizer (`~transformers.Qwen2Tokenizer`, `~transformers.BertTokenizer` | None): + A `Qwen2Tokenizer` or `BertTokenizer` to tokenize text. + transformer ([`EasyAnimateTransformer3DModel`]): + The EasyAnimate model designed by EasyAnimate Team. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with EasyAnimate to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKLMagvit, + text_encoder: Qwen2VLForConditionalGeneration | BertModel, + tokenizer: Qwen2Tokenizer | BertTokenizer, + transformer: EasyAnimateTransformer3DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.enable_text_attention_mask = ( + self.transformer.config.enable_text_attention_mask + if getattr(self, "transformer", None) is not None + else True + ) + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 4 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_spatial_compression_ratio, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + + # Copied from diffusers.pipelines.easyanimate.pipeline_easyanimate.EasyAnimatePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + dtype (`torch.dtype`): + torch dtype + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds` is passed directly. + max_sequence_length (`int`, *optional*): maximum sequence length to use for the prompt. + """ + dtype = dtype or self.text_encoder.dtype + device = device or self.text_encoder.device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + if isinstance(prompt, str): + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": prompt}], + } + ] + else: + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": _prompt}], + } + for _prompt in prompt + ] + text = [ + self.tokenizer.apply_chat_template([m], tokenize=False, add_generation_prompt=True) for m in messages + ] + + text_inputs = self.tokenizer( + text=text, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_attention_mask=True, + padding_side="right", + return_tensors="pt", + ) + text_inputs = text_inputs.to(self.text_encoder.device) + + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + if self.enable_text_attention_mask: + # Inference: Generation of the output + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ).hidden_states[-2] + else: + raise ValueError("LLM needs attention_mask") + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.to(device=device) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + if negative_prompt is not None and isinstance(negative_prompt, str): + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": negative_prompt}], + } + ] + else: + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": _negative_prompt}], + } + for _negative_prompt in negative_prompt + ] + text = [ + self.tokenizer.apply_chat_template([m], tokenize=False, add_generation_prompt=True) for m in messages + ] + + text_inputs = self.tokenizer( + text=text, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_attention_mask=True, + padding_side="right", + return_tensors="pt", + ) + text_inputs = text_inputs.to(self.text_encoder.device) + + text_input_ids = text_inputs.input_ids + negative_prompt_attention_mask = text_inputs.attention_mask + if self.enable_text_attention_mask: + # Inference: Generation of the output + negative_prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=negative_prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-2] + else: + raise ValueError("LLM needs attention_mask") + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.to(device=device) + + return prompt_embeds, negative_prompt_embeds, prompt_attention_mask, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents( + self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None + ): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + (num_frames - 1) // self.vae_temporal_compression_ratio + 1, + height // self.vae_spatial_compression_ratio, + width // self.vae_spatial_compression_ratio, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # scale the initial noise by the standard deviation required by the scheduler + if hasattr(self.scheduler, "init_noise_sigma"): + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_control_latents( + self, control, control_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the control to latents shape as we concatenate the control to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + + if control is not None: + control = control.to(device=device, dtype=dtype) + bs = 1 + new_control = [] + for i in range(0, control.shape[0], bs): + control_bs = control[i : i + bs] + control_bs = self.vae.encode(control_bs)[0] + control_bs = control_bs.mode() + new_control.append(control_bs) + control = torch.cat(new_control, dim=0) + control = control * self.vae.config.scaling_factor + + if control_image is not None: + control_image = control_image.to(device=device, dtype=dtype) + bs = 1 + new_control_pixel_values = [] + for i in range(0, control_image.shape[0], bs): + control_pixel_values_bs = control_image[i : i + bs] + control_pixel_values_bs = self.vae.encode(control_pixel_values_bs)[0] + control_pixel_values_bs = control_pixel_values_bs.mode() + new_control_pixel_values.append(control_pixel_values_bs) + control_image_latents = torch.cat(new_control_pixel_values, dim=0) + control_image_latents = control_image_latents * self.vae.config.scaling_factor + else: + control_image_latents = None + + return control, control_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + num_frames: int | None = 49, + height: int | None = 512, + width: int | None = 512, + control_video: torch.FloatTensor = None, + control_camera_video: torch.FloatTensor = None, + ref_image: torch.FloatTensor = None, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + guidance_rescale: float = 0.0, + timesteps: list[int] | None = None, + ): + r""" + Generates images or video using the EasyAnimate pipeline based on the provided prompts. + + Args: + prompt (`str` or `list[str]`, *optional*): + Text prompts to guide the image or video generation. If not provided, use `prompt_embeds` instead. + num_frames (`int`, *optional*): + Length of the generated video (in frames). + height (`int`, *optional*): + Height of the generated image in pixels. + width (`int`, *optional*): + Width of the generated image in pixels. + control_video (`torch.FloatTensor`, *optional*): + Control video used to condition the generation. + control_camera_video (`torch.FloatTensor`, *optional*): + Control camera video used to condition the generation. + ref_image (`torch.FloatTensor`, *optional*): + Reference image used to condition the generation. + num_inference_steps (`int`, *optional*, defaults to 50): + Number of denoising steps during generation. More steps generally yield higher quality images but slow + down inference. + guidance_scale (`float`, *optional*, defaults to 5.0): + Encourages the model to align outputs with prompts. A higher value may decrease image quality. + negative_prompt (`str` or `list[str]`, *optional*): + Prompts indicating what to exclude in generation. If not specified, use `negative_prompt_embeds`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + Number of images to generate for each prompt. + eta (`float`, *optional*, defaults to 0.0): + Applies to DDIM scheduling. Controlled by the eta parameter from the related literature. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A generator to ensure reproducibility in image generation. + latents (`torch.Tensor`, *optional*): + Predefined latent tensors to condition generation. + prompt_embeds (`torch.Tensor`, *optional*): + Text embeddings for the prompts. Overrides prompt string inputs for more flexibility. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Embeddings for negative prompts. Overrides string inputs if defined. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the primary prompt embeddings. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for negative prompt embeddings. + output_type (`str`, *optional*, defaults to "latent"): + Format of the generated output, either as a PIL image or as a NumPy array. + return_dict (`bool`, *optional*, defaults to `True`): + If `True`, returns a structured output. Otherwise returns a simple tuple. + callback_on_step_end (`Callable`, *optional*): + Functions called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + Tensor names to be included in callback function calls. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Adjusts noise levels based on guidance scale. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, the scheduler's default schedule for + `num_inference_steps` is used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. default height and width + height = int((height // 16) * 16) + width = int((width // 16) * 16) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = self.transformer.dtype + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + device=device, + dtype=dtype, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + text_encoder_index=0, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler): + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, mu=1 + ) + else: + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps + ) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + num_frames, + height, + width, + dtype, + device, + generator, + latents, + ) + + if control_camera_video is not None: + control_video_latents = resize_mask(control_camera_video, latents, process_first_frame_only=True) + control_video_latents = control_video_latents * 6 + control_latents = ( + torch.cat([control_video_latents] * 2) if self.do_classifier_free_guidance else control_video_latents + ).to(device, dtype) + elif control_video is not None: + batch_size, channels, num_frames, height_video, width_video = control_video.shape + control_video = self.image_processor.preprocess( + control_video.permute(0, 2, 1, 3, 4).reshape( + batch_size * num_frames, channels, height_video, width_video + ), + height=height, + width=width, + ) + control_video = control_video.to(dtype=torch.float32) + control_video = control_video.reshape(batch_size, num_frames, channels, height, width).permute( + 0, 2, 1, 3, 4 + ) + control_video_latents = self.prepare_control_latents( + None, + control_video, + batch_size, + height, + width, + dtype, + device, + generator, + self.do_classifier_free_guidance, + )[1] + control_latents = ( + torch.cat([control_video_latents] * 2) if self.do_classifier_free_guidance else control_video_latents + ).to(device, dtype) + else: + control_video_latents = torch.zeros_like(latents).to(device, dtype) + control_latents = ( + torch.cat([control_video_latents] * 2) if self.do_classifier_free_guidance else control_video_latents + ).to(device, dtype) + + if ref_image is not None: + batch_size, channels, num_frames, height_video, width_video = ref_image.shape + ref_image = self.image_processor.preprocess( + ref_image.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height_video, width_video), + height=height, + width=width, + ) + ref_image = ref_image.to(dtype=torch.float32) + ref_image = ref_image.reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4) + + ref_image_latents = self.prepare_control_latents( + None, + ref_image, + batch_size, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + )[1] + + ref_image_latents_conv_in = torch.zeros_like(latents) + if latents.size()[2] != 1: + ref_image_latents_conv_in[:, :, :1] = ref_image_latents + ref_image_latents_conv_in = ( + torch.cat([ref_image_latents_conv_in] * 2) + if self.do_classifier_free_guidance + else ref_image_latents_conv_in + ).to(device, dtype) + control_latents = torch.cat([control_latents, ref_image_latents_conv_in], dim=1) + else: + ref_image_latents_conv_in = torch.zeros_like(latents) + ref_image_latents_conv_in = ( + torch.cat([ref_image_latents_conv_in] * 2) + if self.do_classifier_free_guidance + else ref_image_latents_conv_in + ).to(device, dtype) + control_latents = torch.cat([control_latents, ref_image_latents_conv_in], dim=1) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask]) + + # To latents.device + prompt_embeds = prompt_embeds.to(device=device) + prompt_attention_mask = prompt_attention_mask.to(device=device) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + if hasattr(self.scheduler, "scale_model_input"): + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # expand scalar t to 1-D tensor to match the 1st dim of latent_model_input + t_expand = torch.tensor([t] * latent_model_input.shape[0], device=device).to( + dtype=latent_model_input.dtype + ) + # predict the noise residual + noise_pred = self.transformer( + latent_model_input, + t_expand, + encoder_hidden_states=prompt_embeds, + control_latents=control_latents, + return_dict=False, + )[0] + if noise_pred.size()[1] != self.vae.config.latent_channels: + noise_pred, _ = noise_pred.chunk(2, dim=1) + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # Convert to tensor + if not output_type == "latent": + video = self.decode_latents(latents) + video = self.video_processor.postprocess_video(video=video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return EasyAnimatePipelineOutput(frames=video) diff --git a/diffusers/pipelines/easyanimate/pipeline_easyanimate_inpaint.py b/diffusers/pipelines/easyanimate/pipeline_easyanimate_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..69bb332944d60b883637f1c06c8601a6ed1c6599 --- /dev/null +++ b/diffusers/pipelines/easyanimate/pipeline_easyanimate_inpaint.py @@ -0,0 +1,1244 @@ +# Copyright 2025 The EasyAnimate team and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import torch +import torch.nn.functional as F +from PIL import Image +from transformers import ( + BertModel, + BertTokenizer, + Qwen2Tokenizer, + Qwen2VLForConditionalGeneration, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKLMagvit, EasyAnimateTransformer3DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from .pipeline_output import EasyAnimatePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import EasyAnimateInpaintPipeline + >>> from diffusers.pipelines.easyanimate.pipeline_easyanimate_inpaint import get_image_to_video_latent + >>> from diffusers.utils import export_to_video, load_image + + >>> pipe = EasyAnimateInpaintPipeline.from_pretrained( + ... "alibaba-pai/EasyAnimateV5.1-12b-zh-InP-diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> prompt = "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot." + >>> validation_image_start = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + ... ) + + >>> validation_image_end = None + >>> sample_size = (448, 576) + >>> num_frames = 49 + >>> input_video, input_video_mask = get_image_to_video_latent( + ... [validation_image_start], validation_image_end, num_frames, sample_size + ... ) + + >>> video = pipe( + ... prompt, + ... num_frames=num_frames, + ... negative_prompt="Twisted body, limb deformities, text subtitles, comics, stillness, ugliness, errors, garbled text.", + ... height=sample_size[0], + ... width=sample_size[1], + ... video=input_video, + ... mask_video=input_video_mask, + ... ) + >>> export_to_video(video.frames[0], "output.mp4", fps=8) + ``` +""" + + +def preprocess_image(image, sample_size): + """ + Preprocess a single image (PIL.Image, numpy.ndarray, or torch.Tensor) to a resized tensor. + """ + if isinstance(image, torch.Tensor): + # If input is a tensor, assume it's in CHW format and resize using interpolation + image = torch.nn.functional.interpolate( + image.unsqueeze(0), size=sample_size, mode="bilinear", align_corners=False + ).squeeze(0) + elif isinstance(image, Image.Image): + # If input is a PIL image, resize and convert to numpy array + image = image.resize((sample_size[1], sample_size[0])) + image = np.array(image) + elif isinstance(image, np.ndarray): + # If input is a numpy array, resize using PIL + image = Image.fromarray(image).resize((sample_size[1], sample_size[0])) + image = np.array(image) + else: + raise ValueError("Unsupported input type. Expected PIL.Image, numpy.ndarray, or torch.Tensor.") + + # Convert to tensor if not already + if not isinstance(image, torch.Tensor): + image = torch.from_numpy(image).permute(2, 0, 1).float() / 255.0 # HWC -> CHW, normalize to [0, 1] + + return image + + +def get_image_to_video_latent(validation_image_start, validation_image_end, num_frames, sample_size): + """ + Generate latent representations for video from start and end images. Inputs can be PIL.Image, numpy.ndarray, or + torch.Tensor. + """ + input_video = None + input_video_mask = None + + if validation_image_start is not None: + # Preprocess the starting image(s) + if isinstance(validation_image_start, list): + image_start = [preprocess_image(img, sample_size) for img in validation_image_start] + else: + image_start = preprocess_image(validation_image_start, sample_size) + + # Create video tensor from the starting image(s) + if isinstance(image_start, list): + start_video = torch.cat( + [img.unsqueeze(1).unsqueeze(0) for img in image_start], + dim=2, + ) + input_video = torch.tile(start_video[:, :, :1], [1, 1, num_frames, 1, 1]) + input_video[:, :, : len(image_start)] = start_video + else: + input_video = torch.tile( + image_start.unsqueeze(1).unsqueeze(0), + [1, 1, num_frames, 1, 1], + ) + + # Normalize input video (already normalized in preprocess_image) + + # Create mask for the input video + input_video_mask = torch.zeros_like(input_video[:, :1]) + if isinstance(image_start, list): + input_video_mask[:, :, len(image_start) :] = 255 + else: + input_video_mask[:, :, 1:] = 255 + + # Handle ending image(s) if provided + if validation_image_end is not None: + if isinstance(validation_image_end, list): + image_end = [preprocess_image(img, sample_size) for img in validation_image_end] + end_video = torch.cat( + [img.unsqueeze(1).unsqueeze(0) for img in image_end], + dim=2, + ) + input_video[:, :, -len(end_video) :] = end_video + input_video_mask[:, :, -len(image_end) :] = 0 + else: + image_end = preprocess_image(validation_image_end, sample_size) + input_video[:, :, -1:] = image_end.unsqueeze(1).unsqueeze(0) + input_video_mask[:, :, -1:] = 0 + + elif validation_image_start is None: + # If no starting image is provided, initialize empty tensors + input_video = torch.zeros([1, 3, num_frames, sample_size[0], sample_size[1]]) + input_video_mask = torch.ones([1, 1, num_frames, sample_size[0], sample_size[1]]) * 255 + + return input_video, input_video_mask + + +# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid +def get_resize_crop_region_for_grid(src, tgt_width, tgt_height): + tw = tgt_width + th = tgt_height + h, w = src + r = h / w + if r > (th / tw): + resize_height = th + resize_width = int(round(th / h * w)) + else: + resize_width = tw + resize_height = int(round(tw / w * h)) + + crop_top = int(round((th - resize_height) / 2.0)) + crop_left = int(round((tw - resize_width) / 2.0)) + + return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Resize mask information in magvit +def resize_mask(mask, latent, process_first_frame_only=True): + latent_size = latent.size() + + if process_first_frame_only: + target_size = list(latent_size[2:]) + target_size[0] = 1 + first_frame_resized = F.interpolate( + mask[:, :, 0:1, :, :], size=target_size, mode="trilinear", align_corners=False + ) + + target_size = list(latent_size[2:]) + target_size[0] = target_size[0] - 1 + if target_size[0] != 0: + remaining_frames_resized = F.interpolate( + mask[:, :, 1:, :, :], size=target_size, mode="trilinear", align_corners=False + ) + resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2) + else: + resized_mask = first_frame_resized + else: + target_size = list(latent_size[2:]) + resized_mask = F.interpolate(mask, size=target_size, mode="trilinear", align_corners=False) + return resized_mask + + +## Add noise to reference video +def add_noise_to_reference_video(image, ratio=None, generator=None): + if ratio is None: + sigma = torch.normal(mean=-3.0, std=0.5, size=(image.shape[0],)).to(image.device) + sigma = torch.exp(sigma).to(image.dtype) + else: + sigma = torch.ones((image.shape[0],)).to(image.device, image.dtype) * ratio + + if generator is not None: + image_noise = ( + torch.randn(image.size(), generator=generator, dtype=image.dtype, device=image.device) + * sigma[:, None, None, None, None] + ) + else: + image_noise = torch.randn_like(image) * sigma[:, None, None, None, None] + image_noise = torch.where(image == -1, torch.zeros_like(image), image_noise) + image = image + image_noise + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class EasyAnimateInpaintPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-video generation using EasyAnimate. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + EasyAnimate uses one text encoder [qwen2 vl](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) in V5.1. + + Args: + vae ([`AutoencoderKLMagvit`]): + Variational Auto-Encoder (VAE) Model to encode and decode video to and from latent representations. + text_encoder (`~transformers.Qwen2VLForConditionalGeneration`, `~transformers.BertModel` | None): + EasyAnimate uses [qwen2 vl](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct) in V5.1. + tokenizer (`~transformers.Qwen2Tokenizer`, `~transformers.BertTokenizer` | None): + A `Qwen2Tokenizer` or `BertTokenizer` to tokenize text. + transformer ([`EasyAnimateTransformer3DModel`]): + The EasyAnimate model designed by EasyAnimate Team. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with EasyAnimate to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKLMagvit, + text_encoder: Qwen2VLForConditionalGeneration | BertModel, + tokenizer: Qwen2Tokenizer | BertTokenizer, + transformer: EasyAnimateTransformer3DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.enable_text_attention_mask = ( + self.transformer.config.enable_text_attention_mask + if getattr(self, "transformer", None) is not None + else True + ) + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 4 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_spatial_compression_ratio, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + + # Copied from diffusers.pipelines.easyanimate.pipeline_easyanimate.EasyAnimatePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + dtype (`torch.dtype`): + torch dtype + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds` is passed directly. + max_sequence_length (`int`, *optional*): maximum sequence length to use for the prompt. + """ + dtype = dtype or self.text_encoder.dtype + device = device or self.text_encoder.device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + if isinstance(prompt, str): + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": prompt}], + } + ] + else: + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": _prompt}], + } + for _prompt in prompt + ] + text = [ + self.tokenizer.apply_chat_template([m], tokenize=False, add_generation_prompt=True) for m in messages + ] + + text_inputs = self.tokenizer( + text=text, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_attention_mask=True, + padding_side="right", + return_tensors="pt", + ) + text_inputs = text_inputs.to(self.text_encoder.device) + + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + if self.enable_text_attention_mask: + # Inference: Generation of the output + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ).hidden_states[-2] + else: + raise ValueError("LLM needs attention_mask") + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.to(device=device) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + if negative_prompt is not None and isinstance(negative_prompt, str): + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": negative_prompt}], + } + ] + else: + messages = [ + { + "role": "user", + "content": [{"type": "text", "text": _negative_prompt}], + } + for _negative_prompt in negative_prompt + ] + text = [ + self.tokenizer.apply_chat_template([m], tokenize=False, add_generation_prompt=True) for m in messages + ] + + text_inputs = self.tokenizer( + text=text, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_attention_mask=True, + padding_side="right", + return_tensors="pt", + ) + text_inputs = text_inputs.to(self.text_encoder.device) + + text_input_ids = text_inputs.input_ids + negative_prompt_attention_mask = text_inputs.attention_mask + if self.enable_text_attention_mask: + # Inference: Generation of the output + negative_prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=negative_prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-2] + else: + raise ValueError("LLM needs attention_mask") + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.to(device=device) + + return prompt_embeds, negative_prompt_embeds, prompt_attention_mask, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + height, + width, + dtype, + device, + generator, + do_classifier_free_guidance, + noise_aug_strength, + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + if mask is not None: + mask = mask.to(device=device, dtype=dtype) + new_mask = [] + bs = 1 + for i in range(0, mask.shape[0], bs): + mask_bs = mask[i : i + bs] + mask_bs = self.vae.encode(mask_bs)[0] + mask_bs = mask_bs.mode() + new_mask.append(mask_bs) + mask = torch.cat(new_mask, dim=0) + mask = mask * self.vae.config.scaling_factor + + if masked_image is not None: + masked_image = masked_image.to(device=device, dtype=dtype) + if self.transformer.config.add_noise_in_inpaint_model: + masked_image = add_noise_to_reference_video( + masked_image, ratio=noise_aug_strength, generator=generator + ) + new_mask_pixel_values = [] + bs = 1 + for i in range(0, masked_image.shape[0], bs): + mask_pixel_values_bs = masked_image[i : i + bs] + mask_pixel_values_bs = self.vae.encode(mask_pixel_values_bs)[0] + mask_pixel_values_bs = mask_pixel_values_bs.mode() + new_mask_pixel_values.append(mask_pixel_values_bs) + masked_image_latents = torch.cat(new_mask_pixel_values, dim=0) + masked_image_latents = masked_image_latents * self.vae.config.scaling_factor + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + else: + masked_image_latents = None + + return mask, masked_image_latents + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + num_frames, + dtype, + device, + generator, + latents=None, + video=None, + timestep=None, + is_strength_max=True, + return_noise=False, + return_video_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + (num_frames - 1) // self.vae_temporal_compression_ratio + 1, + height // self.vae_spatial_compression_ratio, + width // self.vae_spatial_compression_ratio, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if return_video_latents or (latents is None and not is_strength_max): + video = video.to(device=device, dtype=dtype) + bs = 1 + new_video = [] + for i in range(0, video.shape[0], bs): + video_bs = video[i : i + bs] + video_bs = self.vae.encode(video_bs)[0] + video_bs = video_bs.sample() + new_video.append(video_bs) + video = torch.cat(new_video, dim=0) + video = video * self.vae.config.scaling_factor + + video_latents = video.repeat(batch_size // video.shape[0], 1, 1, 1, 1) + video_latents = video_latents.to(device=device, dtype=dtype) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler): + latents = noise if is_strength_max else self.scheduler.scale_noise(video_latents, timestep, noise) + else: + latents = noise if is_strength_max else self.scheduler.add_noise(video_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + if hasattr(self.scheduler, "init_noise_sigma"): + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + else: + if hasattr(self.scheduler, "init_noise_sigma"): + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_video_latents: + outputs += (video_latents,) + + return outputs + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + num_frames: int | None = 49, + video: torch.FloatTensor = None, + mask_video: torch.FloatTensor = None, + masked_video_latents: torch.FloatTensor = None, + height: int | None = 512, + width: int | None = 512, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + guidance_rescale: float = 0.0, + strength: float = 1.0, + noise_aug_strength: float = 0.0563, + timesteps: list[int] | None = None, + ): + r""" + The call function to the pipeline for generation with HunyuanDiT. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + num_frames (`int`, *optional*): + Length of the video to be generated in seconds. This parameter influences the number of frames and + continuity of generated content. + video (`torch.FloatTensor`, *optional*): + A tensor representing an input video, which can be modified depending on the prompts provided. + mask_video (`torch.FloatTensor`, *optional*): + A tensor to specify areas of the video to be masked (omitted from generation). + masked_video_latents (`torch.FloatTensor`, *optional*): + Latents from masked portions of the video, utilized during image generation. + height (`int`, *optional*): + The height in pixels of the generated image or video frames. + width (`int`, *optional*): + The width in pixels of the generated image or video frames. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image but slower + inference time. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 5.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is effective when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to exclude in image generation. If not defined, you need to provide + `negative_prompt_embeds`. This parameter is ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + A parameter defined in the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only applies to the + [`~schedulers.DDIMScheduler`] and is ignored in other schedulers. It adjusts noise level during the + inference process. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) for setting + random seeds which helps in making generation deterministic. + latents (`torch.Tensor`, *optional*): + A pre-computed latent representation which can be used to guide the generation process. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings, aiding in fine-tuning what should not be represented in the + outputs. If not provided, embeddings are generated from the `negative_prompt` argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask guiding the focus of the model on specific parts of the prompt text. Required when using + `prompt_embeds`. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt, needed when `negative_prompt_embeds` are used. + output_type (`str`, *optional*, defaults to `"latent"`): + The output format of the generated image. Choose between `PIL.Image` and `np.array` to define how you + want the results to be formatted. + return_dict (`bool`, *optional*, defaults to `True`): + If set to `True`, a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] will be returned; + otherwise, a tuple containing the generated images and safety flags will be returned. + callback_on_step_end (`Callable[[int, int], None]`, `PipelineCallback`, `MultiPipelineCallbacks`, + *optional*): + A callback function (or a list of them) that will be executed at the end of each denoising step, + allowing for custom processing during generation. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + Specifies which tensor inputs should be included in the callback function. If not defined, all tensor + inputs will be passed, facilitating enhanced logging or monitoring of the generation process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Rescale parameter for adjusting noise configuration based on guidance rescale. Based on findings from + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + strength (`float`, *optional*, defaults to 1.0): + Affects the overall styling or quality of the generated output. Values closer to 1 usually provide + direct adherence to prompts. + noise_aug_strength (`float`, *optional*, defaults to 0.0563): + Strength of the noise augmentation applied to the conditioning video latents. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, the scheduler's default schedule for + `num_inference_steps` is used. + + Examples: + # Example usage of the function for generating images based on prompts. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + Returns either a structured output containing generated images and their metadata when `return_dict` is + `True`, or a simpler tuple, where the first element is a list of generated images and the second + element indicates if any of them contain "not-safe-for-work" (NSFW) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. default height and width + height = int(height // 16 * 16) + width = int(width // 16 * 16) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = self.transformer.dtype + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + device=device, + dtype=dtype, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + ) + + # 4. set timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler): + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, mu=1 + ) + else: + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps=num_inference_steps, strength=strength, device=device + ) + + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + if video is not None: + batch_size, channels, num_frames, height_video, width_video = video.shape + init_video = self.image_processor.preprocess( + video.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height_video, width_video), + height=height, + width=width, + ) + init_video = init_video.to(dtype=torch.float32) + init_video = init_video.reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4) + else: + init_video = None + + # Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_transformer = self.transformer.config.in_channels + return_image_latents = num_channels_transformer == num_channels_latents + + # 5. Prepare latents. + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + num_frames, + dtype, + device, + generator, + latents, + video=init_video, + timestep=latent_timestep, + is_strength_max=is_strength_max, + return_noise=True, + return_video_latents=return_image_latents, + ) + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 6. Prepare inpaint latents if it needs. + if mask_video is not None: + if (mask_video == 255).all(): + mask = torch.zeros_like(latents).to(device, dtype) + # Use zero latents if we want to t2v. + if self.transformer.config.resize_inpaint_mask_directly: + mask_latents = torch.zeros_like(latents)[:, :1].to(device, dtype) + else: + mask_latents = torch.zeros_like(latents).to(device, dtype) + masked_video_latents = torch.zeros_like(latents).to(device, dtype) + + mask_input = torch.cat([mask_latents] * 2) if self.do_classifier_free_guidance else mask_latents + masked_video_latents_input = ( + torch.cat([masked_video_latents] * 2) if self.do_classifier_free_guidance else masked_video_latents + ) + inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=1).to(dtype) + else: + # Prepare mask latent variables + batch_size, channels, num_frames, height_video, width_video = mask_video.shape + mask_condition = self.mask_processor.preprocess( + mask_video.permute(0, 2, 1, 3, 4).reshape( + batch_size * num_frames, channels, height_video, width_video + ), + height=height, + width=width, + ) + mask_condition = mask_condition.to(dtype=torch.float32) + mask_condition = mask_condition.reshape(batch_size, num_frames, channels, height, width).permute( + 0, 2, 1, 3, 4 + ) + + if num_channels_transformer != num_channels_latents: + mask_condition_tile = torch.tile(mask_condition, [1, 3, 1, 1, 1]) + if masked_video_latents is None: + masked_video = ( + init_video * (mask_condition_tile < 0.5) + + torch.ones_like(init_video) * (mask_condition_tile > 0.5) * -1 + ) + else: + masked_video = masked_video_latents + + if self.transformer.config.resize_inpaint_mask_directly: + _, masked_video_latents = self.prepare_mask_latents( + None, + masked_video, + batch_size, + height, + width, + dtype, + device, + generator, + self.do_classifier_free_guidance, + noise_aug_strength=noise_aug_strength, + ) + mask_latents = resize_mask( + 1 - mask_condition, masked_video_latents, self.vae.config.cache_mag_vae + ) + mask_latents = mask_latents.to(device, dtype) * self.vae.config.scaling_factor + else: + mask_latents, masked_video_latents = self.prepare_mask_latents( + mask_condition_tile, + masked_video, + batch_size, + height, + width, + dtype, + device, + generator, + self.do_classifier_free_guidance, + noise_aug_strength=noise_aug_strength, + ) + + mask_input = torch.cat([mask_latents] * 2) if self.do_classifier_free_guidance else mask_latents + masked_video_latents_input = ( + torch.cat([masked_video_latents] * 2) + if self.do_classifier_free_guidance + else masked_video_latents + ) + inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=1).to(dtype) + else: + inpaint_latents = None + + mask = torch.tile(mask_condition, [1, num_channels_latents, 1, 1, 1]) + mask = F.interpolate(mask, size=latents.size()[-3:], mode="trilinear", align_corners=True).to( + device, dtype + ) + else: + if num_channels_transformer != num_channels_latents: + mask = torch.zeros_like(latents).to(device, dtype) + if self.transformer.config.resize_inpaint_mask_directly: + mask_latents = torch.zeros_like(latents)[:, :1].to(device, dtype) + else: + mask_latents = torch.zeros_like(latents).to(device, dtype) + masked_video_latents = torch.zeros_like(latents).to(device, dtype) + + mask_input = torch.cat([mask_latents] * 2) if self.do_classifier_free_guidance else mask_latents + masked_video_latents_input = ( + torch.cat([masked_video_latents] * 2) if self.do_classifier_free_guidance else masked_video_latents + ) + inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=1).to(dtype) + else: + mask = torch.zeros_like(init_video[:, :1]) + mask = torch.tile(mask, [1, num_channels_latents, 1, 1, 1]) + mask = F.interpolate(mask, size=latents.size()[-3:], mode="trilinear", align_corners=True).to( + device, dtype + ) + + inpaint_latents = None + + # Check that sizes of mask, masked image and latents match + if num_channels_transformer != num_channels_latents: + num_channels_mask = mask_latents.shape[1] + num_channels_masked_image = masked_video_latents.shape[1] + if ( + num_channels_latents + num_channels_mask + num_channels_masked_image + != self.transformer.config.in_channels + ): + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.transformer`: {self.transformer.config} expects" + f" {self.transformer.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.transformer` or your `mask_image` or `image` input." + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask]) + + # To latents.device + prompt_embeds = prompt_embeds.to(device=device) + prompt_attention_mask = prompt_attention_mask.to(device=device) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + if hasattr(self.scheduler, "scale_model_input"): + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # expand scalar t to 1-D tensor to match the 1st dim of latent_model_input + t_expand = torch.tensor([t] * latent_model_input.shape[0], device=device).to( + dtype=latent_model_input.dtype + ) + + # predict the noise residual + noise_pred = self.transformer( + latent_model_input, + t_expand, + encoder_hidden_states=prompt_embeds, + inpaint_latents=inpaint_latents, + return_dict=False, + )[0] + if noise_pred.size()[1] != self.vae.config.latent_channels: + noise_pred, _ = noise_pred.chunk(2, dim=1) + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if num_channels_transformer == num_channels_latents: + init_latents_proper = image_latents + init_mask = mask + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler): + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep], noise) + ) + else: + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + latents = 1 / self.vae.config.scaling_factor * latents + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video=video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return EasyAnimatePipelineOutput(frames=video) diff --git a/diffusers/pipelines/easyanimate/pipeline_output.py b/diffusers/pipelines/easyanimate/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..4fd904ae7dfeb3e842cbb62844df643c7e5b9889 --- /dev/null +++ b/diffusers/pipelines/easyanimate/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class EasyAnimatePipelineOutput(BaseOutput): + r""" + Output class for EasyAnimate pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/diffusers/pipelines/ernie_image/__init__.py b/diffusers/pipelines/ernie_image/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..97355fb609f3155732c7a7a9720ad4c77eb675a8 --- /dev/null +++ b/diffusers/pipelines/ernie_image/__init__.py @@ -0,0 +1,47 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa: F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_ernie_image"] = ["ErnieImagePipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_ernie_image import ErnieImagePipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/ernie_image/pipeline_ernie_image.py b/diffusers/pipelines/ernie_image/pipeline_ernie_image.py new file mode 100644 index 0000000000000000000000000000000000000000..11fce6a204bf28a24c70fbfe25861e7e33282a00 --- /dev/null +++ b/diffusers/pipelines/ernie_image/pipeline_ernie_image.py @@ -0,0 +1,399 @@ +# Copyright 2025 Baidu ERNIE-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Ernie-Image Pipeline for HuggingFace Diffusers. +""" + +import json +from typing import Callable, List, Optional, Union + +import torch +from transformers import AutoTokenizer, Mistral3Model + +from ...image_processor import VaeImageProcessor +from ...loaders import ErnieImageLoraLoaderMixin +from ...models import AutoencoderKLFlux2 +from ...models.transformers import ErnieImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils.import_utils import is_transformers_version +from ...utils.torch_utils import randn_tensor +from .pipeline_output import ErnieImagePipelineOutput + + +if is_transformers_version("<", "5.0.0"): + raise ImportError("`ErnieImagePipeline` requires `transformers>=5.0.0` for `Ministral3ForCausalLM`.") + +from transformers import Ministral3ForCausalLM # noqa: E402 + + +class ErnieImagePipeline(DiffusionPipeline, ErnieImageLoraLoaderMixin): + """ + Pipeline for text-to-image generation using ErnieImageTransformer2DModel. + + This pipeline uses: + - A custom DiT transformer model + - A Flux2-style VAE for encoding/decoding latents + - A text encoder (e.g., Qwen) for text conditioning + - Flow Matching Euler Discrete Scheduler + """ + + model_cpu_offload_seq = "pe->text_encoder->transformer->vae" + # For SGLang fallback ... + _optional_components = ["pe", "pe_tokenizer"] + _callback_tensor_inputs = ["latents"] + + def __init__( + self, + transformer: ErnieImageTransformer2DModel, + vae: AutoencoderKLFlux2, + text_encoder: Mistral3Model, + tokenizer: AutoTokenizer, + scheduler: FlowMatchEulerDiscreteScheduler, + pe: Optional[Ministral3ForCausalLM] = None, + pe_tokenizer: Optional[AutoTokenizer] = None, + ): + super().__init__() + self.register_modules( + transformer=transformer, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + pe=pe, + pe_tokenizer=pe_tokenizer, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels)) if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @torch.no_grad() + def _enhance_prompt_with_pe( + self, + prompt: str, + device: torch.device, + width: int = 1024, + height: int = 1024, + system_prompt: Optional[str] = None, + temperature: float = 0.6, + top_p: float = 0.95, + ) -> str: + """Use PE model to rewrite/enhance a short prompt via chat_template.""" + # Build user message as JSON carrying prompt text and target resolution + user_content = json.dumps( + {"prompt": prompt, "width": width, "height": height}, + ensure_ascii=False, + ) + messages = [] + if system_prompt is not None: + messages.append({"role": "system", "content": system_prompt}) + messages.append({"role": "user", "content": user_content}) + + # apply_chat_template picks up the chat_template.jinja loaded with pe_tokenizer + input_text = self.pe_tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=False, # "Output:" is already in the user block + ) + inputs = self.pe_tokenizer(input_text, return_tensors="pt").to(device) + output_ids = self.pe.generate( + **inputs, + max_new_tokens=self.pe_tokenizer.model_max_length, + do_sample=temperature != 1.0 or top_p != 1.0, + temperature=temperature, + top_p=top_p, + pad_token_id=self.pe_tokenizer.pad_token_id, + eos_token_id=self.pe_tokenizer.eos_token_id, + ) + # Decode only newly generated tokens + generated_ids = output_ids[0][inputs["input_ids"].shape[1] :] + return self.pe_tokenizer.decode(generated_ids, skip_special_tokens=True).strip() + + @torch.no_grad() + def encode_prompt( + self, + prompt: Union[str, List[str]], + device: torch.device, + num_images_per_prompt: int = 1, + ) -> List[torch.Tensor]: + """Encode text prompts to embeddings.""" + if isinstance(prompt, str): + prompt = [prompt] + + text_hiddens = [] + + for p in prompt: + ids = self.tokenizer( + p, + add_special_tokens=True, + truncation=True, + padding=False, + )["input_ids"] + + if len(ids) == 0: + if self.tokenizer.bos_token_id is not None: + ids = [self.tokenizer.bos_token_id] + else: + ids = [0] + + input_ids = torch.tensor([ids], device=device) + with torch.no_grad(): + outputs = self.text_encoder( + input_ids=input_ids, + output_hidden_states=True, + ) + # Use second to last hidden state (matches training) + hidden = outputs.hidden_states[-2][0] # [T, H] + + # Repeat for num_images_per_prompt + for _ in range(num_images_per_prompt): + text_hiddens.append(hidden) + + return text_hiddens + + @staticmethod + def _patchify_latents(latents: torch.Tensor) -> torch.Tensor: + """2x2 patchify: [B, 32, H, W] -> [B, 128, H/2, W/2]""" + b, c, h, w = latents.shape + latents = latents.view(b, c, h // 2, 2, w // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + return latents.reshape(b, c * 4, h // 2, w // 2) + + @staticmethod + def _unpatchify_latents(latents: torch.Tensor) -> torch.Tensor: + """Reverse patchify: [B, 128, H/2, W/2] -> [B, 32, H, W]""" + b, c, h, w = latents.shape + latents = latents.reshape(b, c // 4, 2, 2, h, w) + latents = latents.permute(0, 1, 4, 2, 5, 3) + return latents.reshape(b, c // 4, h * 2, w * 2) + + @staticmethod + def _pad_text(text_hiddens: List[torch.Tensor], device: torch.device, dtype: torch.dtype, text_in_dim: int): + B = len(text_hiddens) + if B == 0: + return torch.zeros((0, 0, text_in_dim), device=device, dtype=dtype), torch.zeros( + (0,), device=device, dtype=torch.long + ) + normalized = [ + th.squeeze(1).to(device).to(dtype) if th.dim() == 3 else th.to(device).to(dtype) for th in text_hiddens + ] + lens = torch.tensor([t.shape[0] for t in normalized], device=device, dtype=torch.long) + Tmax = int(lens.max().item()) + text_bth = torch.zeros((B, Tmax, text_in_dim), device=device, dtype=dtype) + for i, t in enumerate(normalized): + text_bth[i, : t.shape[0], :] = t + return text_bth, lens + + @torch.no_grad() + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + negative_prompt: Optional[Union[str, List[str]]] = "", + height: int = 1024, + width: int = 1024, + num_inference_steps: int = 50, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: list[torch.FloatTensor] | None = None, + output_type: str = "pil", + return_dict: bool = True, + callback_on_step_end: Optional[Callable[[int, int, dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + use_pe: bool = True, # 默认使用PE进行改写 + ): + """ + Generate images from text prompts. + + Args: + prompt: Text prompt(s) + negative_prompt: Negative prompt(s) for CFG. Default is "". + height: Image height in pixels (must be divisible by 16). Default: 1024. + width: Image width in pixels (must be divisible by 16). Default: 1024. + num_inference_steps: Number of denoising steps + guidance_scale: CFG scale (1.0 = no guidance). Default: 4.0. + num_images_per_prompt: Number of images per prompt + generator: Random generator for reproducibility + latents: Pre-generated latents (optional) + prompt_embeds: Pre-computed text embeddings for positive prompts (optional). + If provided, `encode_prompt` is skipped for positive prompts. + negative_prompt_embeds: Pre-computed text embeddings for negative prompts (optional). + If provided, `encode_prompt` is skipped for negative prompts. + output_type: "pil" or "latent" + return_dict: Whether to return a dataclass + callback_on_step_end: Optional callback invoked at the end of each denoising step. + Called as `callback_on_step_end(pipeline, step, timestep, callback_kwargs)` where `callback_kwargs` + contains the tensors listed in `callback_on_step_end_tensor_inputs`. The callback may return a dict to + override those tensors for subsequent steps. + callback_on_step_end_tensor_inputs: List of tensor names passed into the callback kwargs. + Must be a subset of `_callback_tensor_inputs` (default: `["latents"]`). + use_pe: Whether to use the PE model to enhance prompts before generation. + + Returns: + :class:`ErnieImagePipelineOutput` with `images` and `revised_prompts`. + """ + device = self._execution_device + dtype = self.transformer.dtype + + self._guidance_scale = guidance_scale + + # Validate prompt / prompt_embeds + if prompt is None and prompt_embeds is None: + raise ValueError("Must provide either `prompt` or `prompt_embeds`.") + if prompt is not None and prompt_embeds is not None: + raise ValueError("Cannot provide both `prompt` and `prompt_embeds` at the same time.") + + # Validate dimensions + if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0: + raise ValueError(f"Height and width must be divisible by {self.vae_scale_factor}") + + # Handle prompts + if prompt is not None: + if isinstance(prompt, str): + prompt = [prompt] + + # [Phase 1] PE: enhance prompts + revised_prompts: Optional[List[str]] = None + if prompt is not None and use_pe and self.pe is not None and self.pe_tokenizer is not None: + prompt = [self._enhance_prompt_with_pe(p, device, width=width, height=height) for p in prompt] + revised_prompts = list(prompt) + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = len(prompt_embeds) + total_batch_size = batch_size * num_images_per_prompt + + # Handle negative prompt + if negative_prompt is None: + negative_prompt = "" + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + if len(negative_prompt) != batch_size: + raise ValueError(f"negative_prompt must have same length as prompt ({batch_size})") + + # [Phase 2] Text encoding + if prompt_embeds is not None: + text_hiddens = [h for h in prompt_embeds for _ in range(num_images_per_prompt)] + else: + text_hiddens = self.encode_prompt(prompt, device, num_images_per_prompt) + + # CFG with negative prompt + if self.do_classifier_free_guidance: + if negative_prompt_embeds is not None: + uncond_text_hiddens = [h for h in negative_prompt_embeds for _ in range(num_images_per_prompt)] + else: + uncond_text_hiddens = self.encode_prompt(negative_prompt, device, num_images_per_prompt) + + # Latent dimensions + latent_h = height // self.vae_scale_factor + latent_w = width // self.vae_scale_factor + latent_channels = self.transformer.config.in_channels # After patchify + + # Initialize latents + if latents is None: + latents = randn_tensor( + (total_batch_size, latent_channels, latent_h, latent_w), + generator=generator, + device=device, + dtype=dtype, + ) + + # Setup scheduler + sigmas = torch.linspace(1.0, 0.0, num_inference_steps + 1) + self.scheduler.set_timesteps(sigmas=sigmas[:-1], device=device) + + # Denoising loop + if self.do_classifier_free_guidance: + cfg_text_hiddens = list(uncond_text_hiddens) + list(text_hiddens) + else: + cfg_text_hiddens = text_hiddens + text_bth, text_lens = self._pad_text( + text_hiddens=cfg_text_hiddens, device=device, dtype=dtype, text_in_dim=self.transformer.config.text_in_dim + ) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(self.scheduler.timesteps): + if self.do_classifier_free_guidance: + latent_model_input = torch.cat([latents, latents], dim=0) + t_batch = torch.full((total_batch_size * 2,), t.item(), device=device, dtype=dtype) + else: + latent_model_input = latents + t_batch = torch.full((total_batch_size,), t.item(), device=device, dtype=dtype) + + # Model prediction + pred = self.transformer( + hidden_states=latent_model_input, + timestep=t_batch, + text_bth=text_bth, + text_lens=text_lens, + return_dict=False, + )[0] + + # Apply CFG + if self.do_classifier_free_guidance: + pred_uncond, pred_cond = pred.chunk(2, dim=0) + pred = pred_uncond + guidance_scale * (pred_cond - pred_uncond) + + # Scheduler step + latents = self.scheduler.step(pred, t, latents).prev_sample + + # Callback + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + latents = callback_outputs.pop("latents", latents) + + progress_bar.update() + + if output_type == "latent": + images = latents + else: + # Decode latents to images + # Unnormalize latents using VAE's BN stats + # TODO: switch to `self.vae.config.batch_norm_eps` once the hub config is updated to match the trained value (1e-5). + bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(device=device, dtype=latents.dtype) + bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + 1e-5).to( + device=device, dtype=latents.dtype + ) + latents = latents * bn_std + bn_mean + + # Unpatchify + latents = self._unpatchify_latents(latents) + + # Decode + images = self.vae.decode(latents, return_dict=False)[0] + + # Post-process + images = self.image_processor.postprocess(images, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (images,) + + return ErnieImagePipelineOutput(images=images, revised_prompts=revised_prompts) diff --git a/diffusers/pipelines/ernie_image/pipeline_output.py b/diffusers/pipelines/ernie_image/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..8919db0c0acac07a7e093ee7e40e76f84d714bf7 --- /dev/null +++ b/diffusers/pipelines/ernie_image/pipeline_output.py @@ -0,0 +1,36 @@ +# Copyright 2025 Baidu ERNIE-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import List, Optional + +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class ErnieImagePipelineOutput(BaseOutput): + """ + Output class for ERNIE-Image pipelines. + + Args: + images (`List[PIL.Image.Image]`): + List of generated images. + revised_prompts (`List[str]`, *optional*): + List of PE-revised prompts. `None` when PE is disabled or unavailable. + """ + + images: List[PIL.Image.Image] + revised_prompts: Optional[List[str]] diff --git a/diffusers/pipelines/flux/__init__.py b/diffusers/pipelines/flux/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ea25c148e2f1d09f3a7fcf124020d6d24bd76935 --- /dev/null +++ b/diffusers/pipelines/flux/__init__.py @@ -0,0 +1,73 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["FluxPipelineOutput", "FluxPriorReduxPipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["modeling_flux"] = ["ReduxImageEncoder"] + _import_structure["pipeline_flux"] = ["FluxPipeline"] + _import_structure["pipeline_flux_control"] = ["FluxControlPipeline"] + _import_structure["pipeline_flux_control_img2img"] = ["FluxControlImg2ImgPipeline"] + _import_structure["pipeline_flux_control_inpaint"] = ["FluxControlInpaintPipeline"] + _import_structure["pipeline_flux_controlnet"] = ["FluxControlNetPipeline"] + _import_structure["pipeline_flux_controlnet_image_to_image"] = ["FluxControlNetImg2ImgPipeline"] + _import_structure["pipeline_flux_controlnet_inpainting"] = ["FluxControlNetInpaintPipeline"] + _import_structure["pipeline_flux_fill"] = ["FluxFillPipeline"] + _import_structure["pipeline_flux_img2img"] = ["FluxImg2ImgPipeline"] + _import_structure["pipeline_flux_inpaint"] = ["FluxInpaintPipeline"] + _import_structure["pipeline_flux_kontext"] = ["FluxKontextPipeline"] + _import_structure["pipeline_flux_kontext_inpaint"] = ["FluxKontextInpaintPipeline"] + _import_structure["pipeline_flux_prior_redux"] = ["FluxPriorReduxPipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .modeling_flux import ReduxImageEncoder + from .pipeline_flux import FluxPipeline + from .pipeline_flux_control import FluxControlPipeline + from .pipeline_flux_control_img2img import FluxControlImg2ImgPipeline + from .pipeline_flux_control_inpaint import FluxControlInpaintPipeline + from .pipeline_flux_controlnet import FluxControlNetPipeline + from .pipeline_flux_controlnet_image_to_image import FluxControlNetImg2ImgPipeline + from .pipeline_flux_controlnet_inpainting import FluxControlNetInpaintPipeline + from .pipeline_flux_fill import FluxFillPipeline + from .pipeline_flux_img2img import FluxImg2ImgPipeline + from .pipeline_flux_inpaint import FluxInpaintPipeline + from .pipeline_flux_kontext import FluxKontextPipeline + from .pipeline_flux_kontext_inpaint import FluxKontextInpaintPipeline + from .pipeline_flux_prior_redux import FluxPriorReduxPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/flux/modeling_flux.py b/diffusers/pipelines/flux/modeling_flux.py new file mode 100644 index 0000000000000000000000000000000000000000..916e3fbe59533d39af85cba3adeae3c771726b02 --- /dev/null +++ b/diffusers/pipelines/flux/modeling_flux.py @@ -0,0 +1,46 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from dataclasses import dataclass + +import torch +import torch.nn as nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin +from ...utils import BaseOutput + + +@dataclass +class ReduxImageEncoderOutput(BaseOutput): + image_embeds: torch.Tensor | None = None + + +class ReduxImageEncoder(ModelMixin, ConfigMixin): + @register_to_config + def __init__( + self, + redux_dim: int = 1152, + txt_in_features: int = 4096, + ) -> None: + super().__init__() + + self.redux_up = nn.Linear(redux_dim, txt_in_features * 3) + self.redux_down = nn.Linear(txt_in_features * 3, txt_in_features) + + def forward(self, x: torch.Tensor) -> ReduxImageEncoderOutput: + projected_x = self.redux_down(nn.functional.silu(self.redux_up(x))) + + return ReduxImageEncoderOutput(image_embeds=projected_x) diff --git a/diffusers/pipelines/flux/pipeline_flux.py b/diffusers/pipelines/flux/pipeline_flux.py new file mode 100644 index 0000000000000000000000000000000000000000..d3e0682c5419d66187889189866c70dde237e2a8 --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux.py @@ -0,0 +1,970 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import FluxPipeline + + >>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0] + >>> image.save("flux.png") + ``` +""" + + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux pipeline for text-to-image generation. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + return latents.to(device=device, dtype=dtype), latent_image_ids + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] | None = None, + true_cfg_scale: float = 1.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 3.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_ip_adapter_image: PipelineImageInput | None = None, + negative_ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and + `negative_prompt` is provided. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 3.5): + Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with `prompt` at the expense of lower image quality. + + Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + if do_true_cfg and prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + negative_text_ids, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas: + sigmas = None + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + negative_ip_adapter_image = [negative_ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + ip_adapter_image = [ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 6. Denoising loop + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_control.py b/diffusers/pipelines/flux/pipeline_flux_control.py new file mode 100644 index 0000000000000000000000000000000000000000..46671c44cca8c25b45daaf6e5253d6cc97f5d97a --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_control.py @@ -0,0 +1,861 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from controlnet_aux import CannyDetector + >>> from diffusers import FluxControlPipeline + >>> from diffusers.utils import load_image + + >>> pipe = FluxControlPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-Canny-dev", torch_dtype=torch.bfloat16 + ... ).to("cuda") + + >>> prompt = "A robot made of exotic candies and chocolates of different kinds. The background is filled with confetti and celebratory gifts." + >>> control_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png" + ... ) + + >>> processor = CannyDetector() + >>> control_image = processor( + ... control_image, low_threshold=50, high_threshold=200, detect_resolution=1024, image_resolution=1024 + ... ) + + >>> image = pipe( + ... prompt=prompt, + ... control_image=control_image, + ... height=1024, + ... width=1024, + ... num_inference_steps=50, + ... guidance_scale=30.0, + ... ).images[0] + >>> image.save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxControlPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The Flux pipeline for controllable text-to-image generation with image conditions. + + Reference: https://bfl.ai/flux-1-tools + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.vae_latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.vae_latent_channels + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + return latents.to(device=device, dtype=dtype), latent_image_ids + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + return latents, latent_image_ids + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + control_image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 3.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 3.5): + Embedded guidance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with prompt at the expense of lower image quality. + + Guidance-distilled models approximates true classifier-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 8 + + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + + if control_image.ndim == 4: + control_image = self.vae.encode(control_image).latent_dist.sample(generator=generator) + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + height_control_image, width_control_image = control_image.shape[2:] + control_image = self._pack_latents( + control_image, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents, control_image], dim=2) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_control_img2img.py b/diffusers/pipelines/flux/pipeline_flux_control_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..b455c611e0ae527e25ff95899b70dd71dee6944d --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_control_img2img.py @@ -0,0 +1,945 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from controlnet_aux import CannyDetector + >>> from diffusers import FluxControlImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = FluxControlImg2ImgPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-Canny-dev", torch_dtype=torch.bfloat16 + ... ).to("cuda") + + >>> prompt = "A robot made of exotic candies and chocolates of different kinds. Abstract background" + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/watercolor-painting.jpg" + ... ) + >>> control_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png" + ... ) + + >>> processor = CannyDetector() + >>> control_image = processor( + ... control_image, low_threshold=50, high_threshold=200, detect_resolution=1024, image_resolution=1024 + ... ) + + >>> image = pipe( + ... prompt=prompt, + ... image=image, + ... control_image=control_image, + ... strength=0.8, + ... height=1024, + ... width=1024, + ... num_inference_steps=50, + ... guidance_scale=30.0, + ... ).images[0] + >>> image.save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxControlImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin): + r""" + The Flux pipeline for image inpainting. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is not None: + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, latent_image_ids + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + control_image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Preprocess image + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4.Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 8 + + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + + if control_image.ndim == 4: + control_image = self.vae.encode(control_image).latent_dist.sample(generator=generator) + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + height_control_image, width_control_image = control_image.shape[2:] + control_image = self._pack_latents( + control_image, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents, control_image], dim=2) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_control_inpaint.py b/diffusers/pipelines/flux/pipeline_flux_control_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..15e27653c3e2b70aa49ef21142e553e1d29e3087 --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_control_inpaint.py @@ -0,0 +1,1114 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ( + CLIPTextModel, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +) +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + import torch + from diffusers import FluxControlInpaintPipeline + from diffusers.models.transformers import FluxTransformer2DModel + from transformers import T5EncoderModel + from diffusers.utils import load_image, make_image_grid + from image_gen_aux import DepthPreprocessor # https://github.com/huggingface/image_gen_aux + from PIL import Image + import numpy as np + + pipe = FluxControlInpaintPipeline.from_pretrained( + "black-forest-labs/FLUX.1-Depth-dev", + torch_dtype=torch.bfloat16, + ) + # use following lines if you have GPU constraints + # --------------------------------------------------------------- + transformer = FluxTransformer2DModel.from_pretrained( + "sayakpaul/FLUX.1-Depth-dev-nf4", subfolder="transformer", torch_dtype=torch.bfloat16 + ) + text_encoder_2 = T5EncoderModel.from_pretrained( + "sayakpaul/FLUX.1-Depth-dev-nf4", subfolder="text_encoder_2", torch_dtype=torch.bfloat16 + ) + pipe.transformer = transformer + pipe.text_encoder_2 = text_encoder_2 + pipe.enable_model_cpu_offload() + # --------------------------------------------------------------- + pipe.to("cuda") + + prompt = "a blue robot singing opera with human-like expressions" + image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png") + + head_mask = np.zeros_like(image) + head_mask[65:580, 300:642] = 255 + mask_image = Image.fromarray(head_mask) + + processor = DepthPreprocessor.from_pretrained("LiheYoung/depth-anything-large-hf") + control_image = processor(image)[0].convert("RGB") + + output = pipe( + prompt=prompt, + image=image, + control_image=control_image, + mask_image=mask_image, + num_inference_steps=30, + strength=0.9, + guidance_scale=10.0, + generator=torch.Generator().manual_seed(42), + ).images[0] + make_image_grid([image, control_image, mask_image, output.resize(image.size)], rows=1, cols=4).save( + "output.png" + ) + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxControlInpaintPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The Flux pipeline for image inpainting using Flux-dev-Depth/Canny. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is not None: + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, noise, image_latents, latent_image_ids + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + def prepare_mask_latents( + self, + image, + mask_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + image = self.image_processor.preprocess(image, height=height, width=width) + mask_image = self.mask_processor.preprocess(mask_image, height=height, width=width) + + masked_image = image * (1 - mask_image) + masked_image = masked_image.to(device=device, dtype=dtype) + + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask_image = torch.nn.functional.interpolate(mask_image, size=(height, width)) + mask_image = mask_image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == num_channels_latents: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask_image.shape[0] < batch_size: + if not batch_size % mask_image.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask_image.shape[0]} mask_image were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask_image = mask_image.repeat(batch_size // mask_image.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask_image = self._pack_latents( + mask_image.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + masked_image_latents = torch.cat((masked_image_latents, mask_image), dim=-1) + + return mask_image, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + control_image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, `list[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will be generated by `mask_image`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + device = self._execution_device + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 3. Preprocess mask and image + num_channels_latents = self.vae.config.latent_channels + if masked_image_latents is not None: + # pre computed masked_image_latents and mask_image + masked_image_latents = masked_image_latents.to(latents.device) + mask = mask_image.to(latents.device) + else: + mask, masked_image_latents = self.prepare_mask_latents( + image, + mask_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 4.Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 8 + + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + + if control_image.ndim == 4: + control_image = self.vae.encode(control_image).latent_dist.sample(generator=generator) + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + height_control_image, width_control_image = control_image.shape[2:] + control_image = self._pack_latents( + control_image, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + latents, noise, image_latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height_8 = 2 * (int(height) // (self.vae_scale_factor * 2)) + width_8 = 2 * (int(width) // (self.vae_scale_factor * 2)) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents, control_image], dim=2) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + # for 64 channel transformer only. + init_mask = mask + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + image_latents, torch.tensor([noise_timestep]), noise + ) + else: + init_latents_proper = image_latents + init_latents_proper = self._pack_latents( + init_latents_proper, batch_size * num_images_per_prompt, num_channels_latents, height_8, width_8 + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_controlnet.py b/diffusers/pipelines/flux/pipeline_flux_controlnet.py new file mode 100644 index 0000000000000000000000000000000000000000..e7792d667f16ea709b2da4d2b8c07e334f98b2dc --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_controlnet.py @@ -0,0 +1,1204 @@ +# Copyright 2025 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.controlnets.controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers.utils import load_image + >>> from diffusers import FluxControlNetPipeline + >>> from diffusers import FluxControlNetModel + + >>> base_model = "black-forest-labs/FLUX.1-dev" + >>> controlnet_model = "InstantX/FLUX.1-dev-controlnet-canny" + >>> controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16) + >>> pipe = FluxControlNetPipeline.from_pretrained( + ... base_model, controlnet=controlnet, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> control_image = load_image("https://huggingface.co/InstantX/SD3-Controlnet-Canny/resolve/main/canny.jpg") + >>> prompt = "A girl in city, 25 years old, cool, futuristic" + >>> image = pipe( + ... prompt, + ... control_image=control_image, + ... control_guidance_start=0.2, + ... control_guidance_end=0.8, + ... controlnet_conditioning_scale=1.0, + ... num_inference_steps=28, + ... guidance_scale=3.5, + ... ).images[0] + >>> image.save("flux.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin, FluxIPAdapterMixin): + r""" + The Flux pipeline for text-to-image generation. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "control_image"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + controlnet: FluxControlNetModel + | list[FluxControlNetModel] + | tuple[FluxControlNetModel] + | FluxMultiControlNetModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + if isinstance(controlnet, (list, tuple)): + controlnet = FluxMultiControlNetModel(controlnet) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + controlnet=controlnet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + return latents.to(device=device, dtype=dtype), latent_image_ids + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + return latents, latent_image_ids + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] | None = None, + true_cfg_scale: float = 1.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + control_image: PipelineImageInput = None, + control_mode: int | list[int] | None = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_ip_adapter_image: PipelineImageInput | None = None, + negative_ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and + `negative_prompt` is provided. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set + the corresponding scale as a list. + control_mode (`int` or `list[int]`,, *optional*, defaults to None): + The control mode when applying ControlNet-Union. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(self.controlnet.nets) if isinstance(self.controlnet, FluxMultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + dtype = self.transformer.dtype + + # 3. Prepare text embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + _, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 3. Prepare control image + num_channels_latents = self.transformer.config.in_channels // 4 + if isinstance(self.controlnet, FluxControlNetModel): + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image.shape[-2:] + + # xlab controlnet has a input_hint_block and instantx controlnet does not + controlnet_blocks_repeat = False if self.controlnet.input_hint_block is None else True + if self.controlnet.input_hint_block is None: + # vae encode + control_image = retrieve_latents(self.vae.encode(control_image), generator=generator) + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # pack + height_control_image, width_control_image = control_image.shape[2:] + control_image = self._pack_latents( + control_image, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + # Here we ensure that `control_mode` has the same length as the control_image. + if control_mode is not None: + if not isinstance(control_mode, int): + raise ValueError(" For `FluxControlNet`, `control_mode` should be an `int` or `None`") + control_mode = torch.tensor(control_mode).to(device, dtype=torch.long) + control_mode = control_mode.view(-1, 1).expand(control_image.shape[0], 1) + + elif isinstance(self.controlnet, FluxMultiControlNetModel): + control_images = [] + # xlab controlnet has a input_hint_block and instantx controlnet does not + controlnet_blocks_repeat = False if self.controlnet.nets[0].input_hint_block is None else True + for i, control_image_ in enumerate(control_image): + control_image_ = self.prepare_image( + image=control_image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image_.shape[-2:] + + if self.controlnet.nets[0].input_hint_block is None: + # vae encode + control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator) + control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # pack + height_control_image, width_control_image = control_image_.shape[2:] + control_image_ = self._pack_latents( + control_image_, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + control_images.append(control_image_) + + control_image = control_images + + # Here we ensure that `control_mode` has the same length as the control_image. + if isinstance(control_mode, list) and len(control_mode) != len(control_image): + raise ValueError( + "For Multi-ControlNet, `control_mode` must be a list of the same " + + " length as the number of controlnets (control images) specified" + ) + if not isinstance(control_mode, list): + control_mode = [control_mode] * len(control_image) + # set control mode + control_modes = [] + for cmode in control_mode: + if cmode is None: + cmode = -1 + control_mode = torch.tensor(cmode).expand(control_images[0].shape[0]).to(device, dtype=torch.long) + control_modes.append(control_mode) + control_mode = control_modes + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(self.controlnet, FluxControlNetModel) else keeps) + + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + if isinstance(self.controlnet, FluxMultiControlNetModel): + use_guidance = self.controlnet.nets[0].config.guidance_embeds + else: + use_guidance = self.controlnet.config.guidance_embeds + + guidance = torch.tensor([guidance_scale], device=device) if use_guidance else None + guidance = guidance.expand(latents.shape[0]) if guidance is not None else None + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + # controlnet + controlnet_block_samples, controlnet_single_block_samples = self.controlnet( + hidden_states=latents, + controlnet_cond=control_image, + controlnet_mode=control_mode, + conditioning_scale=cond_scale, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + ) + + guidance = ( + torch.tensor([guidance_scale], device=device) if self.transformer.config.guidance_embeds else None + ) + guidance = guidance.expand(latents.shape[0]) if guidance is not None else None + + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + controlnet_block_samples=controlnet_block_samples, + controlnet_single_block_samples=controlnet_single_block_samples, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + controlnet_blocks_repeat=controlnet_blocks_repeat, + )[0] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + controlnet_block_samples=controlnet_block_samples, + controlnet_single_block_samples=controlnet_single_block_samples, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + controlnet_blocks_repeat=controlnet_blocks_repeat, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + control_image = callback_outputs.pop("control_image", control_image) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py b/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py new file mode 100644 index 0000000000000000000000000000000000000000..61c9da0c949679f93d4ec97449deb193f45dc38d --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py @@ -0,0 +1,1013 @@ +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ( + CLIPTextModel, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.controlnets.controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import FluxControlNetImg2ImgPipeline, FluxControlNetModel + >>> from diffusers.utils import load_image + + >>> device = "cuda" if torch.cuda.is_available() else "cpu" + + >>> controlnet = FluxControlNetModel.from_pretrained( + ... "InstantX/FLUX.1-dev-Controlnet-Canny-alpha", torch_dtype=torch.bfloat16 + ... ) + + >>> pipe = FluxControlNetImg2ImgPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-schnell", controlnet=controlnet, torch_dtype=torch.float16 + ... ) + + >>> pipe.text_encoder.to(torch.float16) + >>> pipe.controlnet.to(torch.float16) + >>> pipe.to("cuda") + + >>> control_image = load_image("https://huggingface.co/InstantX/SD3-Controlnet-Canny/resolve/main/canny.jpg") + >>> init_image = load_image( + ... "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + ... ) + + >>> prompt = "A girl in city, 25 years old, cool, futuristic" + >>> image = pipe( + ... prompt, + ... image=init_image, + ... control_image=control_image, + ... control_guidance_start=0.2, + ... control_guidance_end=0.8, + ... controlnet_conditioning_scale=1.0, + ... strength=0.7, + ... num_inference_steps=2, + ... guidance_scale=3.5, + ... ).images[0] + >>> image.save("flux_controlnet_img2img.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxControlNetImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin): + r""" + The Flux controlnet pipeline for image-to-image generation. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "control_image"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + controlnet: FluxControlNetModel + | list[FluxControlNetModel] + | tuple[FluxControlNetModel] + | FluxMultiControlNetModel, + ): + super().__init__() + if isinstance(controlnet, (list, tuple)): + controlnet = FluxMultiControlNetModel(controlnet) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + controlnet=controlnet, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + prompt_2, + strength, + height, + width, + callback_on_step_end_tensor_inputs, + prompt_embeds=None, + pooled_prompt_embeds=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is not None: + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, latent_image_ids + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + control_image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + control_mode: int | list[int] | None = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. + image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.FloatTensor`): + The image(s) to modify with the pipeline. + control_image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.FloatTensor`): + The ControlNet input condition. Image to control the generation. + height (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + strength (`float`, *optional*, defaults to 0.6): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. + num_inference_steps (`int`, *optional*, defaults to 28): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). + control_mode (`int` or `list[int]`, *optional*): + The mode for the ControlNet. If multiple ControlNets are used, this should be a list. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original transformer. + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or more [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) to + make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + Additional keyword arguments to be passed to the joint attention mechanism. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising step during the inference. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. + max_sequence_length (`int`, *optional*, defaults to 512): + The maximum length of the sequence to be generated. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(self.controlnet.nets) if isinstance(self.controlnet, FluxMultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + self.check_inputs( + prompt, + prompt_2, + strength, + height, + width, + callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + dtype = self.transformer.dtype + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + num_channels_latents = self.transformer.config.in_channels // 4 + + if isinstance(self.controlnet, FluxControlNetModel): + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image.shape[-2:] + + # xlab controlnet has a input_hint_block and instantx controlnet does not + controlnet_blocks_repeat = False if self.controlnet.input_hint_block is None else True + if self.controlnet.input_hint_block is None: + control_image = retrieve_latents(self.vae.encode(control_image), generator=generator) + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + height_control_image, width_control_image = control_image.shape[2:] + control_image = self._pack_latents( + control_image, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + if control_mode is not None: + control_mode = torch.tensor(control_mode).to(device, dtype=torch.long) + control_mode = control_mode.reshape([-1, 1]) + + elif isinstance(self.controlnet, FluxMultiControlNetModel): + control_images = [] + + # xlab controlnet has a input_hint_block and instantx controlnet does not + controlnet_blocks_repeat = False if self.controlnet.nets[0].input_hint_block is None else True + for i, control_image_ in enumerate(control_image): + control_image_ = self.prepare_image( + image=control_image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image_.shape[-2:] + + if self.controlnet.nets[0].input_hint_block is None: + control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator) + control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + height_control_image, width_control_image = control_image_.shape[2:] + control_image_ = self._pack_latents( + control_image_, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + control_images.append(control_image_) + + control_image = control_images + + control_mode_ = [] + if isinstance(control_mode, list): + for cmode in control_mode: + if cmode is None: + control_mode_.append(-1) + else: + control_mode_.append(cmode) + control_mode = torch.tensor(control_mode_).to(device, dtype=torch.long) + control_mode = control_mode.reshape([-1, 1]) + + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(self.controlnet, FluxControlNetModel) else keeps) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + if isinstance(self.controlnet, FluxMultiControlNetModel): + use_guidance = self.controlnet.nets[0].config.guidance_embeds + else: + use_guidance = self.controlnet.config.guidance_embeds + + guidance = torch.tensor([guidance_scale], device=device) if use_guidance else None + guidance = guidance.expand(latents.shape[0]) if guidance is not None else None + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + controlnet_block_samples, controlnet_single_block_samples = self.controlnet( + hidden_states=latents, + controlnet_cond=control_image, + controlnet_mode=control_mode, + conditioning_scale=cond_scale, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + ) + + guidance = ( + torch.tensor([guidance_scale], device=device) if self.transformer.config.guidance_embeds else None + ) + guidance = guidance.expand(latents.shape[0]) if guidance is not None else None + + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + controlnet_block_samples=controlnet_block_samples, + controlnet_single_block_samples=controlnet_single_block_samples, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + controlnet_blocks_repeat=controlnet_blocks_repeat, + )[0] + + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + control_image = callback_outputs.pop("control_image", control_image) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_controlnet_inpainting.py b/diffusers/pipelines/flux/pipeline_flux_controlnet_inpainting.py new file mode 100644 index 0000000000000000000000000000000000000000..eed671152bc9b88fa6374e3fa603968c39ef5d91 --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_controlnet_inpainting.py @@ -0,0 +1,1210 @@ +import inspect +from typing import Any, Callable + +import numpy as np +import PIL +import torch +from transformers import ( + CLIPTextModel, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.controlnets.controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import FluxControlNetInpaintPipeline + >>> from diffusers.models import FluxControlNetModel + >>> from diffusers.utils import load_image + + >>> controlnet = FluxControlNetModel.from_pretrained( + ... "InstantX/FLUX.1-dev-controlnet-canny", torch_dtype=torch.float16 + ... ) + >>> pipe = FluxControlNetInpaintPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-schnell", controlnet=controlnet, torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + + >>> control_image = load_image( + ... "https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny-alpha/resolve/main/canny.jpg" + ... ) + >>> init_image = load_image( + ... "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + ... ) + >>> mask_image = load_image( + ... "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + ... ) + + >>> prompt = "A girl holding a sign that says InstantX" + >>> image = pipe( + ... prompt, + ... image=init_image, + ... mask_image=mask_image, + ... control_image=control_image, + ... control_guidance_start=0.2, + ... control_guidance_end=0.8, + ... controlnet_conditioning_scale=0.7, + ... strength=0.7, + ... num_inference_steps=28, + ... guidance_scale=3.5, + ... ).images[0] + >>> image.save("flux_controlnet_inpaint.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxControlNetInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin): + r""" + The Flux controlnet pipeline for inpainting. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "control_image", "mask", "masked_image_latents"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + controlnet: FluxControlNetModel + | list[FluxControlNetModel] + | tuple[FluxControlNetModel] + | FluxMultiControlNetModel, + ): + super().__init__() + if isinstance(controlnet, (list, tuple)): + controlnet = FluxMultiControlNetModel(controlnet) + + self.register_modules( + scheduler=scheduler, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + transformer=transformer, + controlnet=controlnet, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + strength, + height, + width, + output_type, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + noise = self._pack_latents(noise, batch_size, num_channels_latents, height, width) + image_latents = self._pack_latents(image_latents, batch_size, num_channels_latents, height, width) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, noise, image_latents, latent_image_ids + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate(mask, size=(height, width)) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 16: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask = self._pack_latents( + mask.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + + return mask, masked_image_latents + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + control_image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + strength: float = 0.6, + padding_mask_crop: int | None = None, + sigmas: list[float] | None = None, + num_inference_steps: int = 28, + guidance_scale: float = 7.0, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + control_mode: int | list[int] | None = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. + image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.FloatTensor`): + The image(s) to inpaint. + mask_image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.FloatTensor`): + The mask image(s) to use for inpainting. White pixels in the mask will be repainted, while black pixels + will be preserved. + masked_image_latents (`torch.FloatTensor`, *optional*): + Pre-generated masked image latents. + control_image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.FloatTensor`): + The ControlNet input condition. Image to control the generation. + height (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + strength (`float`, *optional*, defaults to 0.6): + Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. + padding_mask_crop (`int`, *optional*): + The size of the padding to use when cropping the mask. + num_inference_steps (`int`, *optional*, defaults to 28): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + control_mode (`int` or `list[int]`, *optional*): + The mode for the ControlNet. If multiple ControlNets are used, this should be a list. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original transformer. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or more [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) to + make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + Additional keyword arguments to be passed to the joint attention mechanism. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising step during the inference. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. + max_sequence_length (`int`, *optional*, defaults to 512): + The maximum length of the sequence to be generated. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + global_height = height + global_width = width + + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(self.controlnet.nets) if isinstance(self.controlnet, FluxMultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs + self.check_inputs( + prompt, + prompt_2, + image, + mask_image, + strength, + height, + width, + output_type=output_type, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + padding_mask_crop=padding_mask_crop, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + dtype = self.transformer.dtype + + # 3. Encode input prompt + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region( + mask_image, global_width, global_height, pad=padding_mask_crop + ) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=global_height, width=global_width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 5. Prepare control image + num_channels_latents = self.transformer.config.in_channels // 4 + if isinstance(self.controlnet, FluxControlNetModel): + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image.shape[-2:] + + # xlab controlnet has a input_hint_block and instantx controlnet does not + controlnet_blocks_repeat = False if self.controlnet.input_hint_block is None else True + if self.controlnet.input_hint_block is None: + # vae encode + control_image = retrieve_latents(self.vae.encode(control_image), generator=generator) + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # pack + height_control_image, width_control_image = control_image.shape[2:] + control_image = self._pack_latents( + control_image, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + # set control mode + if control_mode is not None: + control_mode = torch.tensor(control_mode).to(device, dtype=torch.long) + control_mode = control_mode.reshape([-1, 1]) + + elif isinstance(self.controlnet, FluxMultiControlNetModel): + control_images = [] + + # xlab controlnet has a input_hint_block and instantx controlnet does not + controlnet_blocks_repeat = False if self.controlnet.nets[0].input_hint_block is None else True + for i, control_image_ in enumerate(control_image): + control_image_ = self.prepare_image( + image=control_image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image_.shape[-2:] + + if self.controlnet.nets[0].input_hint_block is None: + # vae encode + control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator) + control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # pack + height_control_image, width_control_image = control_image_.shape[2:] + control_image_ = self._pack_latents( + control_image_, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + control_images.append(control_image_) + + control_image = control_images + + # set control mode + control_mode_ = [] + if isinstance(control_mode, list): + for cmode in control_mode: + if cmode is None: + control_mode_.append(-1) + else: + control_mode_.append(cmode) + control_mode = torch.tensor(control_mode_).to(device, dtype=torch.long) + control_mode = control_mode.reshape([-1, 1]) + + # 6. Prepare timesteps + + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(global_height) // self.vae_scale_factor // 2) * ( + int(global_width) // self.vae_scale_factor // 2 + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 7. Prepare latent variables + + latents, noise, image_latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + global_height, + global_width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 8. Prepare mask latents + mask_condition = self.mask_processor.preprocess( + mask_image, height=global_height, width=global_width, resize_mode=resize_mode, crops_coords=crops_coords + ) + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + global_height, + global_width, + prompt_embeds.dtype, + device, + generator, + ) + + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(self.controlnet, FluxControlNetModel) else keeps) + + # 9. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + # predict the noise residual + if isinstance(self.controlnet, FluxMultiControlNetModel): + use_guidance = self.controlnet.nets[0].config.guidance_embeds + else: + use_guidance = self.controlnet.config.guidance_embeds + if use_guidance: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + controlnet_block_samples, controlnet_single_block_samples = self.controlnet( + hidden_states=latents, + controlnet_cond=control_image, + controlnet_mode=control_mode, + conditioning_scale=cond_scale, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + ) + + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + controlnet_block_samples=controlnet_block_samples, + controlnet_single_block_samples=controlnet_single_block_samples, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + controlnet_blocks_repeat=controlnet_blocks_repeat, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + # For inpainting, we need to apply the mask and add the masked image latents + init_latents_proper = image_latents + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + # call the callback, if provided + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + control_image = callback_outputs.pop("control_image", control_image) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # Post-processing + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, global_height, global_width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_fill.py b/diffusers/pipelines/flux/pipeline_flux_fill.py new file mode 100644 index 0000000000000000000000000000000000000000..ab4431b5b7682af196685ad7889377ec4d58dff4 --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_fill.py @@ -0,0 +1,1012 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from ...image_processor import VaeImageProcessor +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import FluxFillPipeline + >>> from diffusers.utils import load_image + + >>> image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/cup.png") + >>> mask = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/cup_mask.png") + + >>> pipe = FluxFillPipeline.from_pretrained("black-forest-labs/FLUX.1-Fill-dev", torch_dtype=torch.bfloat16) + >>> pipe.enable_model_cpu_offload() # save some VRAM by offloading the model to CPU + + >>> image = pipe( + ... prompt="a white paper cup", + ... image=image, + ... mask_image=mask, + ... height=1632, + ... width=1232, + ... guidance_scale=30, + ... num_inference_steps=50, + ... max_sequence_length=512, + ... generator=torch.Generator("cpu").manual_seed(0), + ... ).images[0] + >>> image.save("flux_fill.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class FluxFillPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The Flux Fill pipeline for image inpainting/outpainting. + + Reference: https://blackforestlabs.ai/flux-1-tools/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # 1. calculate the height and width of the latents + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + # 2. encode the masked image + if masked_image.shape[1] == num_channels_latents: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + # 3. duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + batch_size = batch_size * num_images_per_prompt + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + # 4. pack the masked_image_latents + # batch_size, num_channels_latents, height, width -> batch_size, height//2 * width//2 , num_channels_latents*4 + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + + # 5.resize mask to latents shape we we concatenate the mask to the latents + mask = mask[:, 0, :, :] # batch_size, 8 * height, 8 * width (mask has not been 8x compressed) + mask = mask.view( + batch_size, height, self.vae_scale_factor, width, self.vae_scale_factor + ) # batch_size, height, 8, width, 8 + mask = mask.permute(0, 2, 4, 1, 3) # batch_size, 8, 8, height, width + mask = mask.reshape( + batch_size, self.vae_scale_factor * self.vae_scale_factor, height, width + ) # batch_size, 8*8, height, width + + # 6. pack the mask: + # batch_size, 64, height, width -> batch_size, height//2 * width//2 , 64*2*2 + mask = self._pack_latents( + mask, + batch_size, + self.vae_scale_factor * self.vae_scale_factor, + height, + width, + ) + mask = mask.to(device=device, dtype=dtype) + + return mask, masked_image_latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + image=None, + mask_image=None, + masked_image_latents=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + if image is not None and masked_image_latents is not None: + raise ValueError( + "Please provide either `image` or `masked_image_latents`, `masked_image_latents` should not be passed." + ) + + if image is not None and mask_image is None: + raise ValueError("Please provide `mask_image` when passing `image`.") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + # Copied from diffusers.pipelines.flux.pipeline_flux_img2img.FluxImg2ImgPipeline.prepare_latents + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is not None: + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: torch.FloatTensor | None = None, + mask_image: torch.FloatTensor | None = None, + masked_image_latents: torch.FloatTensor | None = None, + height: int | None = None, + width: int | None = None, + strength: float = 1.0, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 30.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, `list[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will be generated by `mask_image`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 30.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + image=image, + mask_image=mask_image, + masked_image_latents=masked_image_latents, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare prompt embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare mask and masked image latents + if masked_image_latents is not None: + masked_image_latents = masked_image_latents.to(latents.device) + else: + mask_image = self.mask_processor.preprocess(mask_image, height=height, width=width) + + masked_image = init_image * (1 - mask_image) + masked_image = masked_image.to(device=device, dtype=prompt_embeds.dtype) + + height, width = init_image.shape[-2:] + mask, masked_image_latents = self.prepare_mask_latents( + mask_image, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + masked_image_latents = torch.cat((masked_image_latents, mask), dim=-1) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=torch.cat((latents, masked_image_latents), dim=2), + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post-process the image + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_img2img.py b/diffusers/pipelines/flux/pipeline_flux_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..94582a84cf84e1396ecc3fcd8f64446ac8ba8bfa --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_img2img.py @@ -0,0 +1,1059 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + + >>> from diffusers import FluxImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> device = "cuda" + >>> pipe = FluxImg2ImgPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) + >>> pipe = pipe.to(device) + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> init_image = load_image(url).resize((1024, 1024)) + + >>> prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k" + + >>> images = pipe( + ... prompt=prompt, image=init_image, num_inference_steps=4, strength=0.95, guidance_scale=0.0 + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxImg2ImgPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin, FluxIPAdapterMixin): + r""" + The Flux pipeline for image inpainting. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + prompt_2, + strength, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is not None: + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] | None = None, + true_cfg_scale: float = 1.0, + image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_ip_adapter_image: PipelineImageInput | None = None, + negative_ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and + `negative_prompt` is provided. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Preprocess image + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + _, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4.Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + + latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_inpaint.py b/diffusers/pipelines/flux/pipeline_flux_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..4c35ffefe0887d76109edf343ae25a26c8b7002d --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_inpaint.py @@ -0,0 +1,1232 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import FluxInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = FluxInpaintPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe(prompt=prompt, image=source, mask_image=mask).images[0] + >>> image.save("flux_inpainting.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxInpaintPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FluxIPAdapterMixin): + r""" + The Flux pipeline for image inpainting. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + strength, + height, + width, + output_type, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + noise = self._pack_latents(noise, batch_size, num_channels_latents, height, width) + image_latents = self._pack_latents(image_latents, batch_size, num_channels_latents, height, width) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, noise, image_latents, latent_image_ids + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate(mask, size=(height, width)) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 16: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = ( + masked_image_latents - self.vae.config.shift_factor + ) * self.vae.config.scaling_factor + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask = self._pack_latents( + mask.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + + return mask, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] | None = None, + true_cfg_scale: float = 1.0, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_ip_adapter_image: PipelineImageInput | None = None, + negative_ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and + `negative_prompt` is provided. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, `list[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will be generated by `mask_image`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + image, + mask_image, + strength, + height, + width, + output_type=output_type, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + padding_mask_crop=padding_mask_crop, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + _, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4.Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + num_channels_transformer = self.transformer.config.in_channels + + latents, noise, image_latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + # for 64 channel transformer only. + init_latents_proper = image_latents + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [ + self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image + ] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_kontext.py b/diffusers/pipelines/flux/pipeline_flux_kontext.py new file mode 100644 index 0000000000000000000000000000000000000000..849d9686de625aa7342df0e7d77ddb62c3a6c1e7 --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_kontext.py @@ -0,0 +1,1110 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import FluxKontextPipeline + >>> from diffusers.utils import load_image + + >>> pipe = FluxKontextPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-Kontext-dev", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png" + ... ).convert("RGB") + >>> prompt = "Make Pikachu hold a sign that says 'Black Forest Labs is awesome', yarn art style, detailed, vibrant colors" + >>> image = pipe( + ... image=image, + ... prompt=prompt, + ... guidance_scale=2.5, + ... generator=torch.Generator().manual_seed(42), + ... ).images[0] + >>> image.save("output.png") + ``` +""" + +PREFERRED_KONTEXT_RESOLUTIONS = [ + (672, 1568), + (688, 1504), + (720, 1456), + (752, 1392), + (800, 1328), + (832, 1248), + (880, 1184), + (944, 1104), + (1024, 1024), + (1104, 944), + (1184, 880), + (1248, 832), + (1328, 800), + (1392, 752), + (1456, 720), + (1504, 688), + (1568, 672), +] + + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class FluxKontextPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux Kontext pipeline for image-to-image and text-to-image generation. + + Reference: https://bfl.ai/announcements/flux-1-kontext-dev + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax") + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + def prepare_latents( + self, + image: torch.Tensor | None, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + + image_latents = image_ids = None + if image is not None: + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latent_height, image_latent_width = image_latents.shape[2:] + image_latents = self._pack_latents( + image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_ids = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + # image ids are the same as latent ids with the first dimension set to 1 instead of 0 + image_ids[..., 0] = 1 + + latent_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents, latent_ids, image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput | None = None, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] | None = None, + true_cfg_scale: float = 1.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 3.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_ip_adapter_image: PipelineImageInput | None = None, + negative_ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + max_area: int = 1024**2, + _auto_resize: bool = True, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 3.5): + Embedded guidance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with prompt at the expense of lower image quality. + + Guidance-distilled models approximates true classifier-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): + Maximum sequence length to use with the `prompt`. + max_area (`int`, defaults to `1024 ** 2`): + The maximum area of the generated image in pixels. The height and width will be adjusted to fit this + area while maintaining the aspect ratio. + _auto_resize (`bool`, *optional*, defaults to `True`): + Whether to automatically resize the input image to the preferred resolutions. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_height, original_width = height, width + aspect_ratio = width / height + width = round((max_area * aspect_ratio) ** 0.5) + height = round((max_area / aspect_ratio) ** 0.5) + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + if height != original_height or width != original_width: + logger.warning( + f"Generation `height` and `width` have been adjusted to {height} and {width} to fit the model requirements." + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + if do_true_cfg and prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + negative_text_ids, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 3. Preprocess image + if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): + img = image[0] if isinstance(image, list) else image + image_height, image_width = self.image_processor.get_default_height_width(img) + aspect_ratio = image_width / image_height + if _auto_resize: + # Kontext is trained on specific resolutions, using one of them is recommended + _, image_width, image_height = min( + (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS + ) + image_width = image_width // multiple_of * multiple_of + image_height = image_height // multiple_of * multiple_of + image = self.image_processor.resize(image, image_height, image_width) + image = self.image_processor.preprocess(image, image_height, image_width) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents, latent_ids, image_ids = self.prepare_latents( + image, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + if image_ids is not None: + latent_ids = torch.cat([latent_ids, image_ids], dim=0) # dim 0 is sequence dimension + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + negative_ip_adapter_image = [negative_ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + ip_adapter_image = [ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 6. Denoising loop + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_kontext_inpaint.py b/diffusers/pipelines/flux/pipeline_flux_kontext_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..9a79aafd8760f437be9a812df36efe45ac7e2456 --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_kontext_inpaint.py @@ -0,0 +1,1429 @@ +# Copyright 2025 ZenAI. All rights reserved. +# author: @vuongminh1907 + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import FluxPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + # Inpainting with text only + ```py + >>> import torch + >>> from diffusers import FluxKontextInpaintPipeline + >>> from diffusers.utils import load_image + + >>> prompt = "Change the yellow dinosaur to green one" + >>> img_url = ( + ... "https://github.com/ZenAI-Vietnam/Flux-Kontext-pipelines/blob/main/assets/dinosaur_input.jpeg?raw=true" + ... ) + >>> mask_url = ( + ... "https://github.com/ZenAI-Vietnam/Flux-Kontext-pipelines/blob/main/assets/dinosaur_mask.png?raw=true" + ... ) + + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + + >>> pipe = FluxKontextInpaintPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-Kontext-dev", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = pipe(prompt=prompt, image=source, mask_image=mask, strength=1.0).images[0] + >>> image.save("kontext_inpainting_normal.png") + ``` + + # Inpainting with image conditioning + ```py + >>> import torch + >>> from diffusers import FluxKontextInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = FluxKontextInpaintPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-Kontext-dev", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> prompt = "Replace this ball" + >>> img_url = "https://images.pexels.com/photos/39362/the-ball-stadion-football-the-pitch-39362.jpeg?auto=compress&cs=tinysrgb&dpr=1&w=500" + >>> mask_url = ( + ... "https://github.com/ZenAI-Vietnam/Flux-Kontext-pipelines/blob/main/assets/ball_mask.png?raw=true" + ... ) + >>> image_reference_url = ( + ... "https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcTah3x6OL_ECMBaZ5ZlJJhNsyC-OSMLWAI-xw&s" + ... ) + + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image_reference = load_image(image_reference_url) + + >>> mask = pipe.mask_processor.blur(mask, blur_factor=12) + >>> image = pipe( + ... prompt=prompt, image=source, mask_image=mask, image_reference=image_reference, strength=1.0 + ... ).images[0] + >>> image.save("kontext_inpainting_ref.png") + ``` +""" + +PREFERRED_KONTEXT_RESOLUTIONS = [ + (672, 1568), + (688, 1504), + (720, 1456), + (752, 1392), + (800, 1328), + (832, 1248), + (880, 1184), + (944, 1104), + (1024, 1024), + (1104, 944), + (1184, 880), + (1248, 832), + (1328, 800), + (1392, 752), + (1456, 720), + (1504, 688), + (1568, 672), +] + + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class FluxKontextInpaintPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux Kontext pipeline for text-to-image generation. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.flux.pipeline_flux_inpaint.FluxInpaintPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + strength, + height, + width, + output_type, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax") + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + def prepare_latents( + self, + image: torch.Tensor | None, + timestep: int, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + image_reference: torch.Tensor | None = None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + + # Prepare image latents + image_latents = image_ids = None + if image is not None: + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + # Prepare image reference latents + image_reference_latents = image_reference_ids = None + if image_reference is not None: + image_reference = image_reference.to(device=device, dtype=dtype) + if image_reference.shape[1] != self.latent_channels: + image_reference_latents = self._encode_vae_image(image=image_reference, generator=generator) + else: + image_reference_latents = image_reference + if batch_size > image_reference_latents.shape[0] and batch_size % image_reference_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_reference_latents.shape[0] + image_reference_latents = torch.cat([image_reference_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_reference_latents.shape[0] and batch_size % image_reference_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_reference` of batch size {image_reference_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_reference_latents = torch.cat([image_reference_latents], dim=0) + + latent_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device=device, dtype=dtype) + latents = noise + + image_latent_height, image_latent_width = image_latents.shape[2:] + image_latents = self._pack_latents( + image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_ids = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + # image ids are the same as latent ids with the first dimension set to 1 instead of 0 + image_ids[..., 0] = 1 + + if image_reference_latents is not None: + image_reference_latent_height, image_reference_latent_width = image_reference_latents.shape[2:] + image_reference_latents = self._pack_latents( + image_reference_latents, + batch_size, + num_channels_latents, + image_reference_latent_height, + image_reference_latent_width, + ) + image_reference_ids = self._prepare_latent_image_ids( + batch_size, image_reference_latent_height // 2, image_reference_latent_width // 2, device, dtype + ) + # image_reference_ids are the same as latent ids with the first dimension set to 1 instead of 0 + image_reference_ids[..., 0] = 1 + + noise = self._pack_latents(noise, batch_size, num_channels_latents, height, width) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents, image_latents, image_reference_latents, latent_ids, image_ids, image_reference_ids, noise + + # Copied from diffusers.pipelines.flux.pipeline_flux_inpaint.FluxInpaintPipeline.prepare_mask_latents + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate(mask, size=(height, width)) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 16: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = ( + masked_image_latents - self.vae.config.shift_factor + ) * self.vae.config.scaling_factor + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask = self._pack_latents( + mask.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + + return mask, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput | None = None, + image_reference: PipelineImageInput | None = None, + mask_image: PipelineImageInput = None, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] | None = None, + true_cfg_scale: float = 1.0, + height: int | None = None, + width: int | None = None, + strength: float = 1.0, + padding_mask_crop: int | None = None, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 3.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_ip_adapter_image: PipelineImageInput | None = None, + negative_ip_adapter_image_embeds: list[torch.Tensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + max_area: int = 1024**2, + _auto_resize: bool = True, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be be inpainted (which parts of the image + to be masked out with `mask_image` and repainted according to `prompt` and `image_reference`). For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + image_reference (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point for the + masked area. For both numpy array and pytorch tensor, the expected value range is between `[0, 1]` If + it's a tensor or a list or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)` If it is + a numpy array or a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can + also accept image latents as `image`, but if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and + `negative_prompt` is provided. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 3.5): + Embedded guidance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with `prompt` at the expense of lower image quality. + + Guidance-distilled models approximates true classifier-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): + Maximum sequence length to use with the `prompt`. + max_area (`int`, defaults to `1024 ** 2`): + The maximum area of the generated image in pixels. The height and width will be adjusted to fit this + area while maintaining the aspect ratio. + _auto_resize (`bool`, *optional*, defaults to `True`): + Whether to automatically resize the input image to the preferred resolutions. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_height, original_width = height, width + aspect_ratio = width / height + width = round((max_area * aspect_ratio) ** 0.5) + height = round((max_area / aspect_ratio) ** 0.5) + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + if height != original_height or width != original_width: + logger.warning( + f"Generation `height` and `width` have been adjusted to {height} and {width} to fit the model requirements." + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + image, + mask_image, + strength, + height, + width, + output_type=output_type, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + padding_mask_crop=padding_mask_crop, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Preprocess image + if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): + if isinstance(image, list) and isinstance(image[0], torch.Tensor) and image[0].ndim == 4: + image = torch.cat(image, dim=0) + img = image[0] if isinstance(image, list) else image + image_height, image_width = self.image_processor.get_default_height_width(img) + aspect_ratio = image_width / image_height + if _auto_resize: + # Kontext is trained on specific resolutions, using one of them is recommended + _, image_width, image_height = min( + (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS + ) + image_width = image_width // multiple_of * multiple_of + image_height = image_height // multiple_of * multiple_of + image = self.image_processor.resize(image, image_height, image_width) + + # Choose the resolution of the image to be the same as the image + width = image_width + height = image_height + + # 2.1 Preprocess mask + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + image = self.image_processor.preprocess( + image, image_height, image_width, crops_coords=crops_coords, resize_mode=resize_mode + ) + else: + raise ValueError("image must be provided correctly for inpainting") + + init_image = image.to(dtype=torch.float32) + + # 2.1 Preprocess image_reference + if image_reference is not None and not ( + isinstance(image_reference, torch.Tensor) and image_reference.size(1) == self.latent_channels + ): + if ( + isinstance(image_reference, list) + and isinstance(image_reference[0], torch.Tensor) + and image_reference[0].ndim == 4 + ): + image_reference = torch.cat(image_reference, dim=0) + img_reference = image_reference[0] if isinstance(image_reference, list) else image_reference + image_reference_height, image_reference_width = self.image_processor.get_default_height_width( + img_reference + ) + aspect_ratio = image_reference_width / image_reference_height + if _auto_resize: + # Kontext is trained on specific resolutions, using one of them is recommended + _, image_reference_width, image_reference_height = min( + (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS + ) + image_reference_width = image_reference_width // multiple_of * multiple_of + image_reference_height = image_reference_height // multiple_of * multiple_of + image_reference = self.image_processor.resize( + image_reference, image_reference_height, image_reference_width + ) + image_reference = self.image_processor.preprocess( + image_reference, + image_reference_height, + image_reference_width, + crops_coords=crops_coords, + resize_mode=resize_mode, + ) + else: + image_reference = None + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + negative_text_ids, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents, image_reference_latents, latent_ids, image_ids, image_reference_ids, noise = ( + self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image_reference, + ) + ) + + if image_reference_ids is not None: + latent_ids = torch.cat([latent_ids, image_reference_ids], dim=0) # dim 0 is sequence dimension + elif image_ids is not None: + latent_ids = torch.cat([latent_ids, image_ids], dim=0) # dim 0 is sequence dimension + + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + masked_image = init_image * (mask_condition < 0.5) + + mask, _ = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + negative_ip_adapter_image = [negative_ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + ip_adapter_image = [ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + + latent_model_input = latents + if image_reference_latents is not None: + latent_model_input = torch.cat([latents, image_reference_latents], dim=1) + elif image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + init_latents_proper = image_latents + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/flux/pipeline_flux_prior_redux.py b/diffusers/pipelines/flux/pipeline_flux_prior_redux.py new file mode 100644 index 0000000000000000000000000000000000000000..f173fdef88c6523d2380caed234a71ab10736e76 --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_flux_prior_redux.py @@ -0,0 +1,502 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import torch +from PIL import Image +from transformers import ( + CLIPTextModel, + CLIPTokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput +from ...loaders import FluxLoraLoaderMixin, TextualInversionLoaderMixin +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ..pipeline_utils import DiffusionPipeline +from .modeling_flux import ReduxImageEncoder +from .pipeline_output import FluxPriorReduxPipelineOutput + + +if is_torch_xla_available(): + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import FluxPriorReduxPipeline, FluxPipeline + >>> from diffusers.utils import load_image + + >>> device = "cuda" + >>> dtype = torch.bfloat16 + + >>> repo_redux = "black-forest-labs/FLUX.1-Redux-dev" + >>> repo_base = "black-forest-labs/FLUX.1-dev" + >>> pipe_prior_redux = FluxPriorReduxPipeline.from_pretrained(repo_redux, torch_dtype=dtype).to(device) + >>> pipe = FluxPipeline.from_pretrained( + ... repo_base, text_encoder=None, text_encoder_2=None, torch_dtype=torch.bfloat16 + ... ).to(device) + + >>> image = load_image( + ... "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/style_ziggy/img5.png" + ... ) + >>> pipe_prior_output = pipe_prior_redux(image) + >>> images = pipe( + ... guidance_scale=2.5, + ... num_inference_steps=50, + ... generator=torch.Generator("cpu").manual_seed(0), + ... **pipe_prior_output, + ... ).images + >>> images[0].save("flux-redux.png") + ``` +""" + + +class FluxPriorReduxPipeline(DiffusionPipeline): + r""" + The Flux Redux pipeline for image-to-image generation. + + Reference: https://blackforestlabs.ai/flux-1-tools/ + + Args: + image_encoder ([`SiglipVisionModel`]): + SIGLIP vision model to encode the input image. + feature_extractor ([`SiglipImageProcessor`]): + Image processor for preprocessing images for the SIGLIP model. + image_embedder ([`ReduxImageEncoder`]): + Redux image encoder to process the SIGLIP embeddings. + text_encoder ([`CLIPTextModel`], *optional*): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`], *optional*): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`, *optional*): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`, *optional*): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "image_encoder->image_embedder" + _optional_components = [ + "text_encoder", + "tokenizer", + "text_encoder_2", + "tokenizer_2", + ] + _callback_tensor_inputs = [] + + def __init__( + self, + image_encoder: SiglipVisionModel, + feature_extractor: SiglipImageProcessor, + image_embedder: ReduxImageEncoder, + text_encoder: CLIPTextModel = None, + tokenizer: CLIPTokenizer = None, + text_encoder_2: T5EncoderModel = None, + tokenizer_2: T5TokenizerFast = None, + ): + super().__init__() + + self.register_modules( + image_encoder=image_encoder, + feature_extractor=feature_extractor, + image_embedder=image_embedder, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + + def check_inputs( + self, + image, + prompt, + prompt_2, + prompt_embeds=None, + pooled_prompt_embeds=None, + prompt_embeds_scale=1.0, + pooled_prompt_embeds_scale=1.0, + ): + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + if prompt is not None and (isinstance(prompt, list) and isinstance(image, list) and len(prompt) != len(image)): + raise ValueError( + f"number of prompts must be equal to number of images, but {len(prompt)} prompts were provided and {len(image)} images" + ) + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + image_batch_size = ( + image.shape[0] if isinstance(image, torch.Tensor) else len(image) if isinstance(image, list) else 1 + ) + for scale_name, scale in ( + ("prompt_embeds_scale", prompt_embeds_scale), + ("pooled_prompt_embeds_scale", pooled_prompt_embeds_scale), + ): + if isinstance(scale, list) and len(scale) != image_batch_size: + raise ValueError( + f"number of weights in `{scale_name}` must be equal to number of images, but " + f"{len(scale)} weights were provided and {image_batch_size} images" + ) + + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + image = self.feature_extractor.preprocess( + images=image, do_resize=True, return_tensors="pt", do_convert_rgb=True + ) + image = image.to(device=device, dtype=dtype) + + image_enc_hidden_states = self.image_encoder(**image).last_hidden_state + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + + return image_enc_hidden_states + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + prompt_embeds_scale: float | list[float] | None = 1.0, + pooled_prompt_embeds_scale: float | list[float] | None = 1.0, + return_dict: bool = True, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. **experimental feature**: to use this feature, + make sure to explicitly load text encoders to the pipeline. Prompts will be ignored if text encoders + are not loaded. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. + prompt_embeds_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + Scale factor (or per-image list of scale factors) applied to the redux prompt embeddings before they + are returned. + pooled_prompt_embeds_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + Scale factor (or per-image list of scale factors) applied to the redux pooled prompt embeddings before + they are returned. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPriorReduxPipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.flux.FluxPriorReduxPipelineOutput`] or `tuple`: + [`~pipelines.flux.FluxPriorReduxPipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + image, + prompt, + prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + prompt_embeds_scale=prompt_embeds_scale, + pooled_prompt_embeds_scale=pooled_prompt_embeds_scale, + ) + + # 2. Define call parameters + if image is not None and isinstance(image, Image.Image): + batch_size = 1 + elif image is not None and isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + if prompt is not None and isinstance(prompt, str): + prompt = batch_size * [prompt] + if isinstance(prompt_embeds_scale, float): + prompt_embeds_scale = batch_size * [prompt_embeds_scale] + if isinstance(pooled_prompt_embeds_scale, float): + pooled_prompt_embeds_scale = batch_size * [pooled_prompt_embeds_scale] + + device = self._execution_device + + # 3. Prepare image embeddings + image_latents = self.encode_image(image, device, 1) + + image_embeds = self.image_embedder(image_latents).image_embeds + image_embeds = image_embeds.to(device=device) + + # 3. Prepare (dummy) text embeddings + if hasattr(self, "text_encoder") and self.text_encoder is not None: + ( + prompt_embeds, + pooled_prompt_embeds, + _, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=1, + max_sequence_length=512, + lora_scale=None, + ) + else: + if prompt is not None: + logger.warning( + "prompt input is ignored when text encoders are not loaded to the pipeline. " + "Make sure to explicitly load the text encoders to enable prompt input. " + ) + # max_sequence_length is 512, t5 encoder hidden size is 4096 + prompt_embeds = torch.zeros((batch_size, 512, 4096), device=device, dtype=image_embeds.dtype) + # pooled_prompt_embeds is 768, clip text encoder hidden size + pooled_prompt_embeds = torch.zeros((batch_size, 768), device=device, dtype=image_embeds.dtype) + + # scale & concatenate image and text embeddings + prompt_embeds = torch.cat([prompt_embeds, image_embeds], dim=1) + + prompt_embeds *= torch.tensor(prompt_embeds_scale, device=device, dtype=image_embeds.dtype)[:, None, None] + pooled_prompt_embeds *= torch.tensor(pooled_prompt_embeds_scale, device=device, dtype=image_embeds.dtype)[ + :, None + ] + + # weighted sum + prompt_embeds = torch.sum(prompt_embeds, dim=0, keepdim=True) + pooled_prompt_embeds = torch.sum(pooled_prompt_embeds, dim=0, keepdim=True) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (prompt_embeds, pooled_prompt_embeds) + + return FluxPriorReduxPipelineOutput(prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds) diff --git a/diffusers/pipelines/flux/pipeline_output.py b/diffusers/pipelines/flux/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..7f24bdbf5d60da1f0bd44971d10616ddc3dc655e --- /dev/null +++ b/diffusers/pipelines/flux/pipeline_output.py @@ -0,0 +1,38 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image +import torch + +from ...utils import BaseOutput + + +@dataclass +class FluxPipelineOutput(BaseOutput): + """ + Output class for Flux image generation pipelines. + + Args: + images (`list[PIL.Image.Image]` or `torch.Tensor` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array or torch tensor of shape `(batch_size, + height, width, num_channels)`. PIL images or numpy array present the denoised images of the diffusion + pipeline. Torch tensors can represent either the denoised images or the intermediate latents ready to be + passed to the decoder. + """ + + images: list[PIL.Image.Image] | np.ndarray + + +@dataclass +class FluxPriorReduxPipelineOutput(BaseOutput): + """ + Output class for Flux Prior Redux pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + prompt_embeds: torch.Tensor + pooled_prompt_embeds: torch.Tensor diff --git a/diffusers/pipelines/flux2/__init__.py b/diffusers/pipelines/flux2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4be2b69f49a952b53de511b7aedb0d2177272a51 --- /dev/null +++ b/diffusers/pipelines/flux2/__init__.py @@ -0,0 +1,53 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["Flux2PipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_flux2"] = ["Flux2Pipeline"] + _import_structure["pipeline_flux2_klein"] = ["Flux2KleinPipeline"] + _import_structure["pipeline_flux2_klein_inpaint"] = ["Flux2KleinInpaintPipeline"] + _import_structure["pipeline_flux2_klein_kv"] = ["Flux2KleinKVPipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_flux2 import Flux2Pipeline + from .pipeline_flux2_klein import Flux2KleinPipeline + from .pipeline_flux2_klein_inpaint import Flux2KleinInpaintPipeline + from .pipeline_flux2_klein_kv import Flux2KleinKVPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/flux2/image_processor.py b/diffusers/pipelines/flux2/image_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..c153386951dd0200fda54f37e6617bf3cf6c07a6 --- /dev/null +++ b/diffusers/pipelines/flux2/image_processor.py @@ -0,0 +1,185 @@ +# Copyright 2025 The Black Forest Labs Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math + +import PIL.Image + +from ...configuration_utils import register_to_config +from ...image_processor import VaeImageProcessor + + +class Flux2ImageProcessor(VaeImageProcessor): + r""" + Image processor to preprocess the reference (character) image for the Flux2 model. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept + `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method. + vae_scale_factor (`int`, *optional*, defaults to `16`): + VAE (spatial) scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of + this factor. + vae_latent_channels (`int`, *optional*, defaults to `32`): + VAE latent channels. + do_normalize (`bool`, *optional*, defaults to `True`): + Whether to normalize the image to [-1,1]. + do_binarize (`bool`, *optional*, defaults to `False`): + Whether to binarize the image to 0/1. + do_convert_rgb (`bool`, *optional*, defaults to be `True`): + Whether to convert the images to RGB format. + do_convert_grayscale (`bool`, *optional*, defaults to be `False`): + Whether to convert the images to grayscale format. + """ + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 16, + vae_latent_channels: int = 32, + do_normalize: bool = True, + do_binarize: bool = False, + do_convert_rgb: bool = True, + do_convert_grayscale: bool = False, + ): + super().__init__( + do_resize=do_resize, + vae_scale_factor=vae_scale_factor, + vae_latent_channels=vae_latent_channels, + do_normalize=do_normalize, + do_binarize=do_binarize, + do_convert_rgb=do_convert_rgb, + do_convert_grayscale=do_convert_grayscale, + ) + + @staticmethod + def check_image_input( + image: PIL.Image.Image, max_aspect_ratio: int = 8, min_side_length: int = 64, max_area: int = 1024 * 1024 + ) -> PIL.Image.Image: + """ + Check if image meets minimum size and aspect ratio requirements. + + Args: + image: PIL Image to validate + max_aspect_ratio: Maximum allowed aspect ratio (width/height or height/width) + min_side_length: Minimum pixels required for width and height + max_area: Maximum allowed area in pixels² + + Returns: + The input image if valid + + Raises: + ValueError: If image is too small or aspect ratio is too extreme + """ + if not isinstance(image, PIL.Image.Image): + raise ValueError(f"Image must be a PIL.Image.Image, got {type(image)}") + + width, height = image.size + + # Check minimum dimensions + if width < min_side_length or height < min_side_length: + raise ValueError( + f"Image too small: {width}×{height}. Both dimensions must be at least {min_side_length}px" + ) + + # Check aspect ratio + aspect_ratio = max(width / height, height / width) + if aspect_ratio > max_aspect_ratio: + raise ValueError( + f"Aspect ratio too extreme: {width}×{height} (ratio: {aspect_ratio:.1f}:1). " + f"Maximum allowed ratio is {max_aspect_ratio}:1" + ) + + return image + + @staticmethod + def _resize_to_target_area(image: PIL.Image.Image, target_area: int = 1024 * 1024) -> PIL.Image.Image: + image_width, image_height = image.size + + scale = math.sqrt(target_area / (image_width * image_height)) + width = int(image_width * scale) + height = int(image_height * scale) + + return image.resize((width, height), PIL.Image.Resampling.LANCZOS) + + @staticmethod + def _resize_if_exceeds_area(image, target_area=1024 * 1024) -> PIL.Image.Image: + image_width, image_height = image.size + pixel_count = image_width * image_height + if pixel_count <= target_area: + return image + return Flux2ImageProcessor._resize_to_target_area(image, target_area) + + def _resize_and_crop( + self, + image: PIL.Image.Image, + width: int, + height: int, + ) -> PIL.Image.Image: + r""" + center crop the image to the specified width and height. + + Args: + image (`PIL.Image.Image`): + The image to resize and crop. + width (`int`): + The width to resize the image to. + height (`int`): + The height to resize the image to. + + Returns: + `PIL.Image.Image`: + The resized and cropped image. + """ + image_width, image_height = image.size + + left = (image_width - width) // 2 + top = (image_height - height) // 2 + right = left + width + bottom = top + height + + return image.crop((left, top, right, bottom)) + + # Taken from + # https://github.com/black-forest-labs/flux2/blob/5a5d316b1b42f6b59a8c9194b77c8256be848432/src/flux2/sampling.py#L310C1-L339C19 + @staticmethod + def concatenate_images(images: list[PIL.Image.Image]) -> PIL.Image.Image: + """ + Concatenate a list of PIL images horizontally with center alignment and white background. + """ + + # If only one image, return a copy of it + if len(images) == 1: + return images[0].copy() + + # Convert all images to RGB if not already + images = [img.convert("RGB") if img.mode != "RGB" else img for img in images] + + # Calculate dimensions for horizontal concatenation + total_width = sum(img.width for img in images) + max_height = max(img.height for img in images) + + # Create new image with white background + background_color = (255, 255, 255) + new_img = PIL.Image.new("RGB", (total_width, max_height), background_color) + + # Paste images with center alignment + x_offset = 0 + for img in images: + y_offset = (max_height - img.height) // 2 + new_img.paste(img, (x_offset, y_offset)) + x_offset += img.width + + return new_img diff --git a/diffusers/pipelines/flux2/pipeline_flux2.py b/diffusers/pipelines/flux2/pipeline_flux2.py new file mode 100644 index 0000000000000000000000000000000000000000..b1645b4ae244d3fb4a80a477a9fb473100c45afe --- /dev/null +++ b/diffusers/pipelines/flux2/pipeline_flux2.py @@ -0,0 +1,1034 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL +import torch +from transformers import AutoProcessor, Mistral3ForConditionalGeneration + +from ...loaders import Flux2LoraLoaderMixin +from ...models import AutoencoderKLFlux2, Flux2Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import Flux2ImageProcessor +from .pipeline_output import Flux2PipelineOutput +from .system_messages import SYSTEM_MESSAGE, SYSTEM_MESSAGE_UPSAMPLING_I2I, SYSTEM_MESSAGE_UPSAMPLING_T2I + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import Flux2Pipeline + + >>> pipe = Flux2Pipeline.from_pretrained("black-forest-labs/FLUX.2-dev", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=50, guidance_scale=2.5).images[0] + >>> image.save("flux.png") + ``` +""" + +UPSAMPLING_MAX_IMAGE_SIZE = 768**2 + + +# Adapted from +# https://github.com/black-forest-labs/flux2/blob/5a5d316b1b42f6b59a8c9194b77c8256be848432/src/flux2/text_encoder.py#L68 +def format_input( + prompts: list[str], + system_message: str = SYSTEM_MESSAGE, + images: list[PIL.Image.Image, list[list[PIL.Image.Image]]] | None = None, +): + """ + Format a batch of text prompts into the conversation format expected by apply_chat_template. Optionally, add images + to the input. + + Args: + prompts: List of text prompts + system_message: System message to use (default: CREATIVE_SYSTEM_MESSAGE) + images (optional): List of images to add to the input. + + Returns: + List of conversations, where each conversation is a list of message dicts + """ + # Remove [IMG] tokens from prompts to avoid Pixtral validation issues + # when truncation is enabled. The processor counts [IMG] tokens and fails + # if the count changes after truncation. + cleaned_txt = [prompt.replace("[IMG]", "") for prompt in prompts] + + if images is None or len(images) == 0: + return [ + [ + { + "role": "system", + "content": [{"type": "text", "text": system_message}], + }, + {"role": "user", "content": [{"type": "text", "text": prompt}]}, + ] + for prompt in cleaned_txt + ] + else: + assert len(images) == len(prompts), "Number of images must match number of prompts" + messages = [ + [ + { + "role": "system", + "content": [{"type": "text", "text": system_message}], + }, + ] + for _ in cleaned_txt + ] + + for i, (el, images) in enumerate(zip(messages, images)): + # optionally add the images per batch element. + if images is not None: + el.append( + { + "role": "user", + "content": [{"type": "image", "image": image_obj} for image_obj in images], + } + ) + # add the text. + el.append( + { + "role": "user", + "content": [{"type": "text", "text": cleaned_txt[i]}], + } + ) + + return messages + + +# Adapted from +# https://github.com/black-forest-labs/flux2/blob/5a5d316b1b42f6b59a8c9194b77c8256be848432/src/flux2/text_encoder.py#L49C5-L66C19 +def _validate_and_process_images( + images: list[list[PIL.Image.Image]] | list[PIL.Image.Image], + image_processor: Flux2ImageProcessor, + upsampling_max_image_size: int, +) -> list[list[PIL.Image.Image]]: + # Simple validation: ensure it's a list of PIL images or list of lists of PIL images + if not images: + return [] + + # Check if it's a list of lists or a list of images + if isinstance(images[0], PIL.Image.Image): + # It's a list of images, convert to list of lists + images = [[im] for im in images] + + # potentially concatenate multiple images to reduce the size + images = [[image_processor.concatenate_images(img_i)] if len(img_i) > 1 else img_i for img_i in images] + + # cap the pixels + images = [ + [image_processor._resize_if_exceeds_area(img_i, upsampling_max_image_size) for img_i in img_i] + for img_i in images + ] + return images + + +# Taken from +# https://github.com/black-forest-labs/flux2/blob/5a5d316b1b42f6b59a8c9194b77c8256be848432/src/flux2/sampling.py#L251 +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class Flux2Pipeline(DiffusionPipeline, Flux2LoraLoaderMixin): + r""" + The Flux2 pipeline for text-to-image generation. + + Reference: [https://bfl.ai/blog/flux-2](https://bfl.ai/blog/flux-2) + + Args: + transformer ([`Flux2Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLFlux2`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Mistral3ForConditionalGeneration`]): + [Mistral3ForConditionalGeneration](https://huggingface.co/docs/transformers/en/model_doc/mistral3#transformers.Mistral3ForConditionalGeneration) + tokenizer (`AutoProcessor`): + Tokenizer of class + [PixtralProcessor](https://huggingface.co/docs/transformers/en/model_doc/pixtral#transformers.PixtralProcessor). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLFlux2, + text_encoder: Mistral3ForConditionalGeneration, + tokenizer: AutoProcessor, + transformer: Flux2Transformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = Flux2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 512 + self.default_sample_size = 128 + + self.system_message = SYSTEM_MESSAGE + self.system_message_upsampling_t2i = SYSTEM_MESSAGE_UPSAMPLING_T2I + self.system_message_upsampling_i2i = SYSTEM_MESSAGE_UPSAMPLING_I2I + self.upsampling_max_image_size = UPSAMPLING_MAX_IMAGE_SIZE + + @staticmethod + def _get_mistral_3_small_prompt_embeds( + text_encoder: Mistral3ForConditionalGeneration, + tokenizer: AutoProcessor, + prompt: str | list[str], + dtype: torch.dtype | None = None, + device: torch.device | None = None, + max_sequence_length: int = 512, + system_message: str = SYSTEM_MESSAGE, + hidden_states_layers: list[int] = (10, 20, 30), + ): + dtype = text_encoder.dtype if dtype is None else dtype + device = text_encoder.device if device is None else device + + prompt = [prompt] if isinstance(prompt, str) else prompt + + # Format input messages + messages_batch = format_input(prompts=prompt, system_message=system_message) + + # Process all messages at once + inputs = tokenizer.apply_chat_template( + messages_batch, + add_generation_prompt=False, + tokenize=True, + return_dict=True, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=max_sequence_length, + ) + + # Move to device + input_ids = inputs["input_ids"].to(device) + attention_mask = inputs["attention_mask"].to(device) + + # Forward pass through the model + output = text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + use_cache=False, + ) + + # Only use outputs from intermediate layers and stack them + out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1) + out = out.to(dtype=dtype, device=device) + + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim) + + return prompt_embeds + + @staticmethod + def _prepare_text_ids( + x: torch.Tensor, # (B, L, D) or (L, D) + t_coord: torch.Tensor | None = None, + ): + B, L, _ = x.shape + out_ids = [] + + for i in range(B): + t = torch.arange(1) if t_coord is None else t_coord[i] + h = torch.arange(1) + w = torch.arange(1) + l = torch.arange(L) + + coords = torch.cartesian_prod(t, h, w, l) + out_ids.append(coords) + + return torch.stack(out_ids) + + @staticmethod + def _prepare_latent_ids( + latents: torch.Tensor, # (B, C, H, W) + ): + r""" + Generates 4D position coordinates (T, H, W, L) for latent tensors. + + Args: + latents (torch.Tensor): + Latent tensor of shape (B, C, H, W) + + Returns: + torch.Tensor: + Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0, + H=[0..H-1], W=[0..W-1], L=0 + """ + + batch_size, _, height, width = latents.shape + + t = torch.arange(1) # [0] - time dimension + h = torch.arange(height) + w = torch.arange(width) + l = torch.arange(1) # [0] - layer dimension + + # Create position IDs: (H*W, 4) + latent_ids = torch.cartesian_prod(t, h, w, l) + + # Expand to batch: (B, H*W, 4) + latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1) + + return latent_ids + + @staticmethod + def _prepare_image_ids( + image_latents: list[torch.Tensor], # [(1, C, H, W), (1, C, H, W), ...] + scale: int = 10, + ): + r""" + Generates 4D time-space coordinates (T, H, W, L) for a sequence of image latents. + + This function creates a unique coordinate for every pixel/patch across all input latent with different + dimensions. + + Args: + image_latents (list[torch.Tensor]): + A list of image latent feature tensors, typically of shape (C, H, W). + scale (int, optional): + A factor used to define the time separation (T-coordinate) between latents. T-coordinate for the i-th + latent is: 'scale + scale * i'. Defaults to 10. + + Returns: + torch.Tensor: + The combined coordinate tensor. Shape: (1, N_total, 4) Where N_total is the sum of (H * W) for all + input latents. + + Coordinate Components (Dimension 4): + - T (Time): The unique index indicating which latent image the coordinate belongs to. + - H (Height): The row index within that latent image. + - W (Width): The column index within that latent image. + - L (Seq. Length): A sequence length dimension, which is always fixed at 0 (size 1) + """ + + if not isinstance(image_latents, list): + raise ValueError(f"Expected `image_latents` to be a list, got {type(image_latents)}.") + + # create time offset for each reference image + t_coords = [scale + scale * t for t in torch.arange(0, len(image_latents))] + t_coords = [t.view(-1) for t in t_coords] + + image_latent_ids = [] + for x, t in zip(image_latents, t_coords): + x = x.squeeze(0) + _, height, width = x.shape + + x_ids = torch.cartesian_prod(t, torch.arange(height), torch.arange(width), torch.arange(1)) + image_latent_ids.append(x_ids) + + image_latent_ids = torch.cat(image_latent_ids, dim=0) + image_latent_ids = image_latent_ids.unsqueeze(0) + + return image_latent_ids + + @staticmethod + def _patchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2) + return latents + + @staticmethod + def _unpatchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width) + latents = latents.permute(0, 1, 4, 2, 5, 3) + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2) + return latents + + @staticmethod + def _pack_latents(latents): + """ + pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels) + """ + + batch_size, num_channels, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1) + + return latents + + @staticmethod + def _unpack_latents_with_ids(x: torch.Tensor, x_ids: torch.Tensor) -> list[torch.Tensor]: + """ + using position ids to scatter tokens into place + """ + x_list = [] + for data, pos in zip(x, x_ids): + _, ch = data.shape # noqa: F841 + h_ids = pos[:, 1].to(torch.int64) + w_ids = pos[:, 2].to(torch.int64) + + h = torch.max(h_ids) + 1 + w = torch.max(w_ids) + 1 + + flat_ids = h_ids * w + w_ids + + out = torch.zeros((h * w, ch), device=data.device, dtype=data.dtype) + out.scatter_(0, flat_ids.unsqueeze(1).expand(-1, ch), data) + + # reshape from (H * W, C) to (H, W, C) and permute to (C, H, W) + + out = out.view(h, w, ch).permute(2, 0, 1) + x_list.append(out) + + return torch.stack(x_list, dim=0) + + def upsample_prompt( + self, + prompt: str | list[str], + images: list[PIL.Image.Image, list[list[PIL.Image.Image]]] = None, + temperature: float = 0.15, + device: torch.device = None, + ) -> list[str]: + prompt = [prompt] if isinstance(prompt, str) else prompt + device = self.text_encoder.device if device is None else device + + # Set system message based on whether images are provided + if images is None or len(images) == 0 or images[0] is None: + system_message = SYSTEM_MESSAGE_UPSAMPLING_T2I + else: + system_message = SYSTEM_MESSAGE_UPSAMPLING_I2I + + # Validate and process the input images + if images: + images = _validate_and_process_images(images, self.image_processor, self.upsampling_max_image_size) + + # Format input messages + messages_batch = format_input(prompts=prompt, system_message=system_message, images=images) + + # Process all messages at once + # with image processing a too short max length can throw an error in here. + inputs = self.tokenizer.apply_chat_template( + messages_batch, + add_generation_prompt=True, + tokenize=True, + return_dict=True, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=2048, + ) + + # Move to device + inputs["input_ids"] = inputs["input_ids"].to(device) + inputs["attention_mask"] = inputs["attention_mask"].to(device) + + if "pixel_values" in inputs: + inputs["pixel_values"] = inputs["pixel_values"].to(device, self.text_encoder.dtype) + + # Generate text using the model's generate method + generated_ids = self.text_encoder.generate( + **inputs, + max_new_tokens=512, + do_sample=True, + temperature=temperature, + use_cache=True, + ) + + # Decode only the newly generated tokens (skip input tokens) + # Extract only the generated portion + input_length = inputs["input_ids"].shape[1] + generated_tokens = generated_ids[:, input_length:] + + upsampled_prompt = self.tokenizer.tokenizer.batch_decode( + generated_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=True + ) + return upsampled_prompt + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int] = (10, 20, 30), + ): + device = device or self._execution_device + + if prompt is None: + prompt = "" + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds = self._get_mistral_3_small_prompt_embeds( + text_encoder=self.text_encoder, + tokenizer=self.tokenizer, + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + system_message=self.system_message, + hidden_states_layers=text_encoder_out_layers, + ) + + batch_size, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = self._prepare_text_ids(prompt_embeds) + text_ids = text_ids.to(device) + return prompt_embeds, text_ids + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if image.ndim != 4: + raise ValueError(f"Expected image dims 4, got {image.ndim}.") + + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + image_latents = self._patchify_latents(image_latents) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.device, image_latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + image_latents.device, image_latents.dtype + ) + image_latents = (image_latents - latents_bn_mean) / latents_bn_std + + return image_latents + + def prepare_latents( + self, + batch_size, + num_latents_channels, + height, + width, + dtype, + device, + generator: torch.Generator, + latents: torch.Tensor | None = None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_latents_channels * 4, height // 2, width // 2) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + latent_ids = self._prepare_latent_ids(latents) + latent_ids = latent_ids.to(device) + + latents = self._pack_latents(latents) # [B, C, H, W] -> [B, H*W, C] + return latents, latent_ids + + def prepare_image_latents( + self, + images: list[torch.Tensor], + batch_size, + generator: torch.Generator, + device, + dtype, + ): + image_latents = [] + for image in images: + image = image.to(device=device, dtype=dtype) + imagge_latent = self._encode_vae_image(image=image, generator=generator) + image_latents.append(imagge_latent) # (1, 128, 32, 32) + + image_latent_ids = self._prepare_image_ids(image_latents) + + # Pack each latent and concatenate + packed_latents = [] + for latent in image_latents: + # latent: (1, 128, 32, 32) + packed = self._pack_latents(latent) # (1, 1024, 128) + packed = packed.squeeze(0) # (1024, 128) - remove batch dim + packed_latents.append(packed) + + # Concatenate all reference tokens along sequence dimension + image_latents = torch.cat(packed_latents, dim=0) # (N*1024, 128) + image_latents = image_latents.unsqueeze(0) # (1, N*1024, 128) + + image_latents = image_latents.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.to(device) + + return image_latents, image_latent_ids + + def check_inputs( + self, + prompt, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if ( + height is not None + and height % (self.vae_scale_factor * 2) != 0 + or width is not None + and width % (self.vae_scale_factor * 2) != 0 + ): + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PIL.Image.Image | list[PIL.Image.Image] | None = None, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = 4.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int] = (10, 20, 30), + caption_upsample_temperature: float = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + guidance_scale (`float`, *optional*, defaults to 1.0): + Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with `prompt` at the expense of lower image quality. + + Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + text_encoder_out_layers (`tuple[int]`): + Layer indices to use in the `text_encoder` to derive the final prompt embeddings. + caption_upsample_temperature (`float`): + When specified, we will try to perform caption upsampling for potentially improved outputs. We + recommend setting it to 0.15 if caption upsampling is to be performed. + + Examples: + + Returns: + [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`: [`~pipelines.flux2.Flux2PipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + prompt_embeds=prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. prepare text embeddings + if caption_upsample_temperature: + prompt = self.upsample_prompt( + prompt, images=image, temperature=caption_upsample_temperature, device=device + ) + prompt_embeds, text_ids = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + text_encoder_out_layers=text_encoder_out_layers, + ) + + # 4. process images + if image is not None and not isinstance(image, list): + image = [image] + + condition_images = None + if image is not None: + for img in image: + self.image_processor.check_image_input(img) + + condition_images = [] + for img in image: + image_width, image_height = img.size + if image_width * image_height > 1024 * 1024: + img = self.image_processor._resize_to_target_area(img, 1024 * 1024) + image_width, image_height = img.size + + multiple_of = self.vae_scale_factor * 2 + image_width = (image_width // multiple_of) * multiple_of + image_height = (image_height // multiple_of) * multiple_of + img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop") + condition_images.append(img) + height = height or image_height + width = width or image_width + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 5. prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_ids = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_latents_channels=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + image_latents = None + image_latent_ids = None + if condition_images is not None: + image_latents, image_latent_ids = self.prepare_image_latents( + images=condition_images, + batch_size=batch_size * num_images_per_prompt, + generator=generator, + device=device, + dtype=self.vae.dtype, + ) + + # 6. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas: + sigmas = None + image_seq_len = latents.shape[1] + mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + + # 7. Denoising loop + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + latent_model_input = latents.to(self.transformer.dtype) + latent_image_ids = latent_ids + + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1).to(self.transformer.dtype) + latent_image_ids = torch.cat([latent_ids, image_latent_ids], dim=1) + + noise_pred = self.transformer( + hidden_states=latent_model_input, # (B, image_seq_len, C) + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, # B, text_seq_len, 4 + img_ids=latent_image_ids, # B, image_seq_len, 4 + joint_attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + noise_pred = noise_pred[:, : latents.size(1) :] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents_with_ids(latents, latent_ids) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + latents.device, latents.dtype + ) + latents = latents * latents_bn_std + latents_bn_mean + latents = self._unpatchify_latents(latents) + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return Flux2PipelineOutput(images=image) diff --git a/diffusers/pipelines/flux2/pipeline_flux2_klein.py b/diffusers/pipelines/flux2/pipeline_flux2_klein.py new file mode 100644 index 0000000000000000000000000000000000000000..d768e6127f2682769529e465d253d16c9d78e153 --- /dev/null +++ b/diffusers/pipelines/flux2/pipeline_flux2_klein.py @@ -0,0 +1,927 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL +import torch +from transformers import Qwen2TokenizerFast, Qwen3ForCausalLM + +from ...loaders import Flux2LoraLoaderMixin +from ...models import AutoencoderKLFlux2, Flux2Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import maybe_adjust_dtype_for_device, randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import Flux2ImageProcessor +from .pipeline_output import Flux2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import Flux2KleinPipeline + + >>> pipe = Flux2KleinPipeline.from_pretrained( + ... "black-forest-labs/FLUX.2-klein-base-9B", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=50, guidance_scale=4.0).images[0] + >>> image.save("flux2_output.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux2.pipeline_flux2.compute_empirical_mu +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class Flux2KleinPipeline(DiffusionPipeline, Flux2LoraLoaderMixin): + r""" + The Flux2 Klein pipeline for text-to-image generation. + + Reference: + [https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence) + + Args: + transformer ([`Flux2Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLFlux2`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen3ForCausalLM`]): + [Qwen3ForCausalLM](https://huggingface.co/docs/transformers/en/model_doc/qwen3#transformers.Qwen3ForCausalLM) + tokenizer (`Qwen2TokenizerFast`): + Tokenizer of class + [Qwen2TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLFlux2, + text_encoder: Qwen3ForCausalLM, + tokenizer: Qwen2TokenizerFast, + transformer: Flux2Transformer2DModel, + is_distilled: bool = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + + self.register_to_config(is_distilled=is_distilled) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = Flux2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 512 + self.default_sample_size = 128 + + @staticmethod + def _get_qwen3_prompt_embeds( + text_encoder: Qwen3ForCausalLM, + tokenizer: Qwen2TokenizerFast, + prompt: str | list[str], + dtype: torch.dtype | None = None, + device: torch.device | None = None, + max_sequence_length: int = 512, + hidden_states_layers: list[int] = (9, 18, 27), + ): + dtype = text_encoder.dtype if dtype is None else dtype + device = text_encoder.device if device is None else device + + prompt = [prompt] if isinstance(prompt, str) else prompt + + all_input_ids = [] + all_attention_masks = [] + + for single_prompt in prompt: + messages = [{"role": "user", "content": single_prompt}] + text = tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=False, + ) + inputs = tokenizer( + text, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=max_sequence_length, + ) + + all_input_ids.append(inputs["input_ids"]) + all_attention_masks.append(inputs["attention_mask"]) + + input_ids = torch.cat(all_input_ids, dim=0).to(device) + attention_mask = torch.cat(all_attention_masks, dim=0).to(device) + + # Forward pass through the model + output = text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + use_cache=False, + ) + + # Only use outputs from intermediate layers and stack them + out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1) + out = out.to(dtype=dtype, device=device) + + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim) + + return prompt_embeds + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_text_ids + def _prepare_text_ids( + x: torch.Tensor, # (B, L, D) or (L, D) + t_coord: torch.Tensor | None = None, + ): + B, L, _ = x.shape + out_ids = [] + + for i in range(B): + t = torch.arange(1) if t_coord is None else t_coord[i] + h = torch.arange(1) + w = torch.arange(1) + l = torch.arange(L) + + coords = torch.cartesian_prod(t, h, w, l) + out_ids.append(coords) + + return torch.stack(out_ids) + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_latent_ids + def _prepare_latent_ids( + latents: torch.Tensor, # (B, C, H, W) + ): + r""" + Generates 4D position coordinates (T, H, W, L) for latent tensors. + + Args: + latents (torch.Tensor): + Latent tensor of shape (B, C, H, W) + + Returns: + torch.Tensor: + Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0, + H=[0..H-1], W=[0..W-1], L=0 + """ + + batch_size, _, height, width = latents.shape + + t = torch.arange(1) # [0] - time dimension + h = torch.arange(height) + w = torch.arange(width) + l = torch.arange(1) # [0] - layer dimension + + # Create position IDs: (H*W, 4) + latent_ids = torch.cartesian_prod(t, h, w, l) + + # Expand to batch: (B, H*W, 4) + latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1) + + return latent_ids + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_image_ids + def _prepare_image_ids( + image_latents: list[torch.Tensor], # [(1, C, H, W), (1, C, H, W), ...] + scale: int = 10, + ): + r""" + Generates 4D time-space coordinates (T, H, W, L) for a sequence of image latents. + + This function creates a unique coordinate for every pixel/patch across all input latent with different + dimensions. + + Args: + image_latents (list[torch.Tensor]): + A list of image latent feature tensors, typically of shape (C, H, W). + scale (int, optional): + A factor used to define the time separation (T-coordinate) between latents. T-coordinate for the i-th + latent is: 'scale + scale * i'. Defaults to 10. + + Returns: + torch.Tensor: + The combined coordinate tensor. Shape: (1, N_total, 4) Where N_total is the sum of (H * W) for all + input latents. + + Coordinate Components (Dimension 4): + - T (Time): The unique index indicating which latent image the coordinate belongs to. + - H (Height): The row index within that latent image. + - W (Width): The column index within that latent image. + - L (Seq. Length): A sequence length dimension, which is always fixed at 0 (size 1) + """ + + if not isinstance(image_latents, list): + raise ValueError(f"Expected `image_latents` to be a list, got {type(image_latents)}.") + + # create time offset for each reference image + t_coords = [scale + scale * t for t in torch.arange(0, len(image_latents))] + t_coords = [t.view(-1) for t in t_coords] + + image_latent_ids = [] + for x, t in zip(image_latents, t_coords): + x = x.squeeze(0) + _, height, width = x.shape + + x_ids = torch.cartesian_prod(t, torch.arange(height), torch.arange(width), torch.arange(1)) + image_latent_ids.append(x_ids) + + image_latent_ids = torch.cat(image_latent_ids, dim=0) + image_latent_ids = image_latent_ids.unsqueeze(0) + + return image_latent_ids + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._patchify_latents + def _patchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpatchify_latents + def _unpatchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width) + latents = latents.permute(0, 1, 4, 2, 5, 3) + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._pack_latents + def _pack_latents(latents): + """ + pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels) + """ + + batch_size, num_channels, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1) + + return latents + + @staticmethod + def _unpack_latents_with_ids( + x: torch.Tensor, x_ids: torch.Tensor, height: int | None = None, width: int | None = None + ) -> list[torch.Tensor]: + """ + using position ids to scatter tokens into place + """ + x_list = [] + for data, pos in zip(x, x_ids): + _, ch = data.shape # noqa: F841 + idx_dtype = maybe_adjust_dtype_for_device(torch.int64, data.device) + h_ids = pos[:, 1].to(idx_dtype) + w_ids = pos[:, 2].to(idx_dtype) + + # Use provided height/width to avoid DtoH sync from torch.max().item() + h = height if height is not None else torch.max(h_ids) + 1 + w = width if width is not None else torch.max(w_ids) + 1 + + flat_ids = h_ids * w + w_ids + + out = torch.zeros((h * w, ch), device=data.device, dtype=data.dtype) + out.scatter_(0, flat_ids.unsqueeze(1).expand(-1, ch), data) + + # reshape from (H * W, C) to (H, W, C) and permute to (C, H, W) + + out = out.view(h, w, ch).permute(2, 0, 1) + x_list.append(out) + + return torch.stack(x_list, dim=0) + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int] = (9, 18, 27), + ): + device = device or self._execution_device + + if prompt is None: + prompt = "" + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds = self._get_qwen3_prompt_embeds( + text_encoder=self.text_encoder, + tokenizer=self.tokenizer, + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + hidden_states_layers=text_encoder_out_layers, + ) + + batch_size, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = self._prepare_text_ids(prompt_embeds) + text_ids = text_ids.to(device) + return prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if image.ndim != 4: + raise ValueError(f"Expected image dims 4, got {image.ndim}.") + + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + image_latents = self._patchify_latents(image_latents) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.device, image_latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + image_latents.device, image_latents.dtype + ) + image_latents = (image_latents - latents_bn_mean) / latents_bn_std + + return image_latents + + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_latents_channels, + height, + width, + dtype, + device, + generator: torch.Generator, + latents: torch.Tensor | None = None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_latents_channels * 4, height // 2, width // 2) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + latent_ids = self._prepare_latent_ids(latents) + latent_ids = latent_ids.to(device) + + latents = self._pack_latents(latents) # [B, C, H, W] -> [B, H*W, C] + return latents, latent_ids + + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline.prepare_image_latents + def prepare_image_latents( + self, + images: list[torch.Tensor], + batch_size, + generator: torch.Generator, + device, + dtype, + ): + image_latents = [] + for image in images: + image = image.to(device=device, dtype=dtype) + imagge_latent = self._encode_vae_image(image=image, generator=generator) + image_latents.append(imagge_latent) # (1, 128, 32, 32) + + image_latent_ids = self._prepare_image_ids(image_latents) + + # Pack each latent and concatenate + packed_latents = [] + for latent in image_latents: + # latent: (1, 128, 32, 32) + packed = self._pack_latents(latent) # (1, 1024, 128) + packed = packed.squeeze(0) # (1024, 128) - remove batch dim + packed_latents.append(packed) + + # Concatenate all reference tokens along sequence dimension + image_latents = torch.cat(packed_latents, dim=0) # (N*1024, 128) + image_latents = image_latents.unsqueeze(0) # (1, N*1024, 128) + + image_latents = image_latents.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.to(device) + + return image_latents, image_latent_ids + + def check_inputs( + self, + prompt, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + guidance_scale=None, + ): + if ( + height is not None + and height % (self.vae_scale_factor * 2) != 0 + or width is not None + and width % (self.vae_scale_factor * 2) != 0 + ): + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if guidance_scale > 1.0 and self.config.is_distilled: + logger.warning(f"Guidance scale {guidance_scale} is ignored for step-wise distilled models.") + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and not self.config.is_distilled + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: list[PIL.Image.Image] | PIL.Image.Image | None = None, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: str | list[str] | None = None, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int, dict], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int] = (9, 18, 27), + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. For step-wise distilled models, + `guidance_scale` is ignored. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Note that "" is used as the negative prompt in this pipeline. + If not provided, will be generated from "". + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + text_encoder_out_layers (`tuple[int]`): + Layer indices to use in the `text_encoder` to derive the final prompt embeddings. + + Examples: + + Returns: + [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`: [`~pipelines.flux2.Flux2PipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + prompt_embeds=prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + guidance_scale=guidance_scale, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. prepare text embeddings + prompt_embeds, text_ids = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + text_encoder_out_layers=text_encoder_out_layers, + ) + + if self.do_classifier_free_guidance: + negative_prompt = "" + if prompt is not None and isinstance(prompt, list): + negative_prompt = [negative_prompt] * len(prompt) + negative_prompt_embeds, negative_text_ids = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + text_encoder_out_layers=text_encoder_out_layers, + ) + + # 4. process images + if image is not None and not isinstance(image, list): + image = [image] + + condition_images = None + if image is not None: + for img in image: + self.image_processor.check_image_input(img) + + condition_images = [] + for img in image: + image_width, image_height = img.size + if image_width * image_height > 1024 * 1024: + img = self.image_processor._resize_to_target_area(img, 1024 * 1024) + image_width, image_height = img.size + + multiple_of = self.vae_scale_factor * 2 + image_width = (image_width // multiple_of) * multiple_of + image_height = (image_height // multiple_of) * multiple_of + img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop") + condition_images.append(img) + height = height or image_height + width = width or image_width + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 5. prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_ids = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_latents_channels=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + image_latents = None + image_latent_ids = None + if condition_images is not None: + image_latents, image_latent_ids = self.prepare_image_latents( + images=condition_images, + batch_size=batch_size * num_images_per_prompt, + generator=generator, + device=device, + dtype=self.vae.dtype, + ) + + # 6. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas: + sigmas = None + image_seq_len = latents.shape[1] + mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 7. Denoising loop + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + latent_model_input = latents.to(self.transformer.dtype) + latent_image_ids = latent_ids + + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1).to(self.transformer.dtype) + latent_image_ids = torch.cat([latent_ids, image_latent_ids], dim=1) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, # (B, image_seq_len, C) + timestep=timestep / 1000, + guidance=None, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, # B, text_seq_len, 4 + img_ids=latent_image_ids, # B, image_seq_len, 4 + joint_attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + noise_pred = noise_pred[:, : latents.size(1) :] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=None, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self._attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1) :] + noise_pred = neg_noise_pred + guidance_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + # Pass pre-computed latent height/width to avoid DtoH sync from torch.max().item() + latent_height = 2 * (int(height) // (self.vae_scale_factor * 2)) + latent_width = 2 * (int(width) // (self.vae_scale_factor * 2)) + latents = self._unpack_latents_with_ids(latents, latent_ids, latent_height // 2, latent_width // 2) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + latents.device, latents.dtype + ) + latents = latents * latents_bn_std + latents_bn_mean + latents = self._unpatchify_latents(latents) + if output_type == "latent": + image = latents + else: + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return Flux2PipelineOutput(images=image) diff --git a/diffusers/pipelines/flux2/pipeline_flux2_klein_inpaint.py b/diffusers/pipelines/flux2/pipeline_flux2_klein_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..fd9467003a7129ad165da8b69570cacfc09a67b3 --- /dev/null +++ b/diffusers/pipelines/flux2/pipeline_flux2_klein_inpaint.py @@ -0,0 +1,1272 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import numpy as np +import PIL +import torch +from transformers import Qwen2TokenizerFast, Qwen3ForCausalLM + +from ...image_processor import PipelineImageInput +from ...loaders import Flux2LoraLoaderMixin +from ...models import AutoencoderKLFlux2, Flux2Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import Flux2ImageProcessor +from .pipeline_output import Flux2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + # Inpainting with text only + ```py + >>> import torch + >>> from diffusers import Flux2KleinInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = Flux2KleinInpaintPipeline.from_pretrained( + ... "black-forest-labs/FLUX.2-klein-base-9B", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe(prompt=prompt, image=source, mask_image=mask).images[0] + >>> image.save("flux2klein_inpainting.png") + ``` + + # Inpainting with image reference conditioning + ```py + >>> import torch + >>> from diffusers import Flux2KleinInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = Flux2KleinInpaintPipeline.from_pretrained( + ... "black-forest-labs/FLUX.2-klein-base-9B", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> prompt = "Replace this ball" + >>> img_url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/klein_inpaint/the-ball-stadion-football-the-pitch-39362.jpeg" + >>> mask_url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/klein_inpaint/ball_mask.png" + >>> image_reference_url = ( + ... "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/klein_inpaint/ball.jpg" + ... ) + + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image_reference = load_image(image_reference_url) + + >>> mask = pipe.mask_processor.blur(mask, blur_factor=12) + >>> image = pipe( + ... prompt=prompt, image=source, mask_image=mask, image_reference=image_reference, strength=1.0 + ... ).images[0] + >>> image.save("flux2klein_inpainting_ref.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux2.pipeline_flux2.compute_empirical_mu +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class Flux2KleinInpaintPipeline(DiffusionPipeline, Flux2LoraLoaderMixin): + r""" + Flux2 Klein pipeline for image inpainting with optional reference image conditioning. + + Reference: + [https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence) + + Args: + transformer ([`Flux2Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLFlux2`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen3ForCausalLM`]): + [Qwen3ForCausalLM](https://huggingface.co/docs/transformers/en/model_doc/qwen3#transformers.Qwen3ForCausalLM) + tokenizer (`Qwen2TokenizerFast`): + Tokenizer of class + [Qwen2TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLFlux2, + text_encoder: Qwen3ForCausalLM, + tokenizer: Qwen2TokenizerFast, + transformer: Flux2Transformer2DModel, + is_distilled: bool = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + + self.register_to_config(is_distilled=is_distilled) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 32 + self.image_processor = Flux2ImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.mask_processor = Flux2ImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_rgb=False, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = 512 + self.default_sample_size = 128 + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2_klein.Flux2KleinPipeline._get_qwen3_prompt_embeds + def _get_qwen3_prompt_embeds( + text_encoder: Qwen3ForCausalLM, + tokenizer: Qwen2TokenizerFast, + prompt: str | list[str], + dtype: torch.dtype | None = None, + device: torch.device | None = None, + max_sequence_length: int = 512, + hidden_states_layers: list[int] = (9, 18, 27), + ): + dtype = text_encoder.dtype if dtype is None else dtype + device = text_encoder.device if device is None else device + + prompt = [prompt] if isinstance(prompt, str) else prompt + + all_input_ids = [] + all_attention_masks = [] + + for single_prompt in prompt: + messages = [{"role": "user", "content": single_prompt}] + text = tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=False, + ) + inputs = tokenizer( + text, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=max_sequence_length, + ) + + all_input_ids.append(inputs["input_ids"]) + all_attention_masks.append(inputs["attention_mask"]) + + input_ids = torch.cat(all_input_ids, dim=0).to(device) + attention_mask = torch.cat(all_attention_masks, dim=0).to(device) + + # Forward pass through the model + output = text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + use_cache=False, + ) + + # Only use outputs from intermediate layers and stack them + out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1) + out = out.to(dtype=dtype, device=device) + + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim) + + return prompt_embeds + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_text_ids + def _prepare_text_ids( + x: torch.Tensor, # (B, L, D) or (L, D) + t_coord: torch.Tensor | None = None, + ): + B, L, _ = x.shape + out_ids = [] + + for i in range(B): + t = torch.arange(1) if t_coord is None else t_coord[i] + h = torch.arange(1) + w = torch.arange(1) + l = torch.arange(L) + + coords = torch.cartesian_prod(t, h, w, l) + out_ids.append(coords) + + return torch.stack(out_ids) + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_latent_ids + def _prepare_latent_ids( + latents: torch.Tensor, # (B, C, H, W) + ): + r""" + Generates 4D position coordinates (T, H, W, L) for latent tensors. + + Args: + latents (torch.Tensor): + Latent tensor of shape (B, C, H, W) + + Returns: + torch.Tensor: + Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0, + H=[0..H-1], W=[0..W-1], L=0 + """ + + batch_size, _, height, width = latents.shape + + t = torch.arange(1) # [0] - time dimension + h = torch.arange(height) + w = torch.arange(width) + l = torch.arange(1) # [0] - layer dimension + + # Create position IDs: (H*W, 4) + latent_ids = torch.cartesian_prod(t, h, w, l) + + # Expand to batch: (B, H*W, 4) + latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1) + + return latent_ids + + @staticmethod + def _prepare_image_ids( + image_latents: list[torch.Tensor], # list of (B_i, C, H, W) before packing + batch_size: int, + scale: int = 10, + ): + r""" + Generates 4D time-space coordinates (T, H, W, L) for a sequence of image latents. + + This function creates a unique coordinate for every pixel/patch across all input latent with different + dimensions. + + Args: + image_latents (list[torch.Tensor]): + A list of image latent feature tensors, typically of shape (C, H, W). + scale (int, optional): + A factor used to define the time separation (T-coordinate) between latents. T-coordinate for the i-th + latent is: 'scale + scale * i'. Defaults to 10. + + Returns: + torch.Tensor: + The combined coordinate tensor. Shape: (1, N_total, 4) Where N_total is the sum of (H * W) for all + input latents. + + Coordinate Components (Dimension 4): + - T (Time): The unique index indicating which latent image the coordinate belongs to. + - H (Height): The row index within that latent image. + - W (Width): The column index within that latent image. + - L (Seq. Length): A sequence length dimension, which is always fixed at 0 (size 1) + """ + + if not isinstance(image_latents, list): + raise ValueError(f"Expected `image_latents` to be a list, got {type(image_latents)}.") + + all_image_latent_ids = [] + t_offset = scale + for x in image_latents: + b_i, _, height, width = x.shape + + # Create IDs for a single image at this t_offset + t = torch.tensor([t_offset]).view(-1) + x_ids = torch.cartesian_prod(t, torch.arange(height), torch.arange(width), torch.arange(1)) + + if b_i == 1 or b_i == batch_size: + x_ids = x_ids.unsqueeze(0).expand(batch_size, -1, -1) + all_image_latent_ids.append(x_ids) + t_offset += scale + else: + # multiple images per sample in the batch + item_ids = [x_ids] + for _ in range(1, b_i): + t_offset += scale + t = torch.tensor([t_offset]).view(-1) + item_ids.append( + torch.cartesian_prod(t, torch.arange(height), torch.arange(width), torch.arange(1)) + ) + x_ids = torch.cat(item_ids, dim=0) # (b_i * h * w, 4) + x_ids = x_ids.unsqueeze(0).expand(batch_size, -1, -1) + all_image_latent_ids.append(x_ids) + t_offset += scale + + image_latent_ids = torch.cat(all_image_latent_ids, dim=1) + + return image_latent_ids + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._patchify_latents + def _patchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpatchify_latents + def _unpatchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width) + latents = latents.permute(0, 1, 4, 2, 5, 3) + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._pack_latents + def _pack_latents(latents): + """ + pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels) + """ + + batch_size, num_channels, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpack_latents_with_ids + def _unpack_latents_with_ids(x: torch.Tensor, x_ids: torch.Tensor) -> list[torch.Tensor]: + """ + using position ids to scatter tokens into place + """ + x_list = [] + for data, pos in zip(x, x_ids): + _, ch = data.shape # noqa: F841 + h_ids = pos[:, 1].to(torch.int64) + w_ids = pos[:, 2].to(torch.int64) + + h = torch.max(h_ids) + 1 + w = torch.max(w_ids) + 1 + + flat_ids = h_ids * w + w_ids + + out = torch.zeros((h * w, ch), device=data.device, dtype=data.dtype) + out.scatter_(0, flat_ids.unsqueeze(1).expand(-1, ch), data) + + # reshape from (H * W, C) to (H, W, C) and permute to (C, H, W) + + out = out.view(h, w, ch).permute(2, 0, 1) + x_list.append(out) + + return torch.stack(x_list, dim=0) + + @staticmethod + def _get_raw_image_size(image: PipelineImageInput) -> tuple[int, int]: + """Helper to get (height, width) without rounding/scaling.""" + if isinstance(image, list): + image = image[0] + + if isinstance(image, PIL.Image.Image): + return image.height, image.width + elif isinstance(image, torch.Tensor): + return image.shape[-2], image.shape[-1] + elif isinstance(image, np.ndarray): + if image.ndim >= 3: + return image.shape[-3], image.shape[-2] + return image.shape[-2], image.shape[-1] + + if hasattr(image, "shape"): + return image.shape[-2], image.shape[-1] + + raise ValueError(f"Unsupported image type: {type(image)}") + + # Copied from diffusers.pipelines.flux2.pipeline_flux2_klein.Flux2KleinPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int] = (9, 18, 27), + ): + device = device or self._execution_device + + if prompt is None: + prompt = "" + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds = self._get_qwen3_prompt_embeds( + text_encoder=self.text_encoder, + tokenizer=self.tokenizer, + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + hidden_states_layers=text_encoder_out_layers, + ) + + batch_size, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = self._prepare_text_ids(prompt_embeds) + text_ids = text_ids.to(device) + return prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if image.ndim != 4: + raise ValueError(f"Expected image dims 4, got {image.ndim}.") + + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + image_latents = self._patchify_latents(image_latents) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.device, image_latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + image_latents.device, image_latents.dtype + ) + image_latents = (image_latents - latents_bn_mean) / latents_bn_std + + return image_latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_latents_channels, + height, + width, + dtype, + device, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_latents_channels * 4, height // 2, width // 2) + + # Create a dummy tensor for _prepare_latent_ids + dummy_latents = torch.zeros(shape, device=device, dtype=dtype) + latent_image_ids = self._prepare_latent_ids(dummy_latents) + latent_image_ids = latent_image_ids.to(device) + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels * 4: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.device, image_latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + image_latents.device, image_latents.dtype + ) + image_latents = (image_latents - latents_bn_mean) / latents_bn_std + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + noise = self._pack_latents(noise) + packed_image_latents = self._pack_latents(image_latents) + latents = self._pack_latents(latents) + return latents, noise, packed_image_latents, image_latents, latent_image_ids + + def prepare_image_latents( + self, + images: list[torch.Tensor], + batch_size, + generator: torch.Generator, + device, + dtype, + ): + image_latents = [] + for image in images: + image = image.to(device=device, dtype=dtype) + + if image.shape[1] != self.latent_channels * 4: + image_latent = self._encode_vae_image(image=image, generator=generator) + else: + image_latent = image + image_latents.append(image_latent) + + image_latent_ids = self._prepare_image_ids(image_latents, batch_size) + + # Pack each latent and combine batch properly + final_latents = [] + for latent in image_latents: + packed = self._pack_latents(latent) # (B_i, seq_len, 128) + b_i = packed.shape[0] + + if b_i == 1 and batch_size > 1: + packed = packed.repeat(batch_size, 1, 1) + elif b_i == batch_size: + pass + else: + # Concatenate all reference tokens along sequence dimension for each sample + seq_len = packed.shape[1] + packed = packed.reshape(1, b_i * seq_len, -1) + if batch_size > 1: + packed = packed.repeat(batch_size, 1, 1) + final_latents.append(packed) + + image_latents = torch.cat(final_latents, dim=1) # (batch_size, total_seq_len, 128) + + image_latent_ids = image_latent_ids.to(device) + + return image_latents, image_latent_ids + + def prepare_mask_latents( + self, + mask, + batch_size, + num_images_per_prompt, + height, + width, + dtype, + device, + ): + # Interpolate the mask directly to the final packed spatial size. + target_h = int(height) // (self.vae_scale_factor * 2) + target_w = int(width) // (self.vae_scale_factor * 2) + mask = torch.nn.functional.interpolate(mask, size=(target_h, target_w), mode="bilinear") + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + # duplicate mask for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + + # Pack to (B, seq_len, 1), will broadcast against (B, seq_len, C) latents + mask = self._pack_latents(mask) + + return mask + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + image, + mask_image, + image_reference, + strength, + height, + width, + output_type, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + guidance_scale=None, + ): + if image is None: + raise ValueError("`image` has to be provided for inpainting.") + + if mask_image is None: + raise ValueError("`mask_image` has to be provided for inpainting.") + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if ( + height is not None + and height % (self.vae_scale_factor * 2) != 0 + or width is not None + and width % (self.vae_scale_factor * 2) != 0 + ): + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + else: + if image is not None: + if not isinstance(image, (PIL.Image.Image, torch.Tensor, np.ndarray, list)): + raise ValueError( + f"`image` has to be of type `PIL.Image.Image`, `torch.Tensor`, `np.ndarray`, or `list`" + f" but is {type(image)}." + ) + if mask_image is not None: + if not isinstance(mask_image, (PIL.Image.Image, torch.Tensor, np.ndarray, list)): + raise ValueError( + f"`mask_image` has to be of type `PIL.Image.Image`, `torch.Tensor`, `np.ndarray`, or `list`" + f" but is {type(mask_image)}." + ) + + if image_reference is not None: + if not isinstance(image_reference, (PIL.Image.Image, torch.Tensor, np.ndarray, list)): + raise ValueError( + f"`image_reference` has to be of type `PIL.Image.Image`, `torch.Tensor`, `np.ndarray`, or `list`" + f" but is {type(image_reference)}." + ) + + if guidance_scale > 1.0 and self.config.is_distilled: + logger.warning(f"Guidance scale {guidance_scale} is ignored for step-wise distilled models.") + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and not self.config.is_distilled + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + image: PipelineImageInput | None = None, + image_reference: PipelineImageInput | None = None, + mask_image: PipelineImageInput | None = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 0.8, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 8.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int, dict], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int, ...] = (9, 18, 27), + ): + r""" + Function invoked when calling the pipeline for inpainting. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a + list of tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or + a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. It can also accept image + latents directly, in which case encoding is skipped. Latents must be in patchified form of shape `(B, + latent_channels * 4, H // 2, W // 2)`, where each 2×2 spatial patch has been folded into the channel + dimension. + image_reference (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`, *optional*): + `Image`, numpy array or tensor representing an image batch to be used as the reference for the masked + area. This allows conditioning the inpainted region on a specific reference image. For both numpy array + and pytorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a list of + tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a list + of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. It can also accept image latents + directly, in which case encoding is skipped. Latents must be in patchified form of shape `(B, + latent_channels * 4, H // 2, W // 2)`, where each 2×2 spatial patch has been folded into the channel + dimension. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + height (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.default_sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 8.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. For step-wise distilled models, + `guidance_scale` is ignored. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Note that "" is used as the negative prompt in this pipeline. + If not provided, will be generated from "". + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux2.Flux2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + text_encoder_out_layers (`Tuple[int]`): + Layer indices to use in the `text_encoder` to derive the final prompt embeddings. + + Examples: + + Returns: + [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`: [`~pipelines.flux2.Flux2PipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + image=image, + mask_image=mask_image, + image_reference=image_reference, + strength=strength, + height=height, + width=width, + output_type=output_type, + prompt_embeds=prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + padding_mask_crop=padding_mask_crop, + guidance_scale=guidance_scale, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Preprocess image + multiple_of = self.vae_scale_factor * 2 + if isinstance(image, torch.Tensor) and image.ndim == 4 and image.size(1) == self.latent_channels * 4: + init_image = image + original_image = image + crops_coords = None + resize_mode = "default" + height = image.shape[2] * self.vae_scale_factor * 2 + width = image.shape[3] * self.vae_scale_factor * 2 + elif image is not None: + if isinstance(image, list) and isinstance(image[0], torch.Tensor) and image[0].ndim == 4: + image = torch.cat(image, dim=0) + img = image[0] if isinstance(image, list) else image + raw_h, raw_w = self._get_raw_image_size(img) + + if raw_h * raw_w > 1024 * 1024: + scale = math.sqrt(1024 * 1024 / (raw_h * raw_w)) + image = self.image_processor.resize(image, int(raw_h * scale), int(raw_w * scale)) + img = image[0] if isinstance(image, list) else image + raw_h, raw_w = self._get_raw_image_size(img) + + image_width = (raw_w // multiple_of) * multiple_of + image_height = (raw_h // multiple_of) * multiple_of + + # Use the resolution of the input image + width = image_width + height = image_height + + # 2.1 Preprocess mask + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, image_height, image_width, crops_coords=crops_coords, resize_mode=resize_mode + ) + + # 2.2 Preprocess reference image + processed_image_reference = None + if image_reference is not None and not ( + isinstance(image_reference, torch.Tensor) and image_reference.size(1) == self.latent_channels * 4 + ): + if ( + isinstance(image_reference, list) + and isinstance(image_reference[0], torch.Tensor) + and image_reference[0].ndim == 4 + ): + image_reference = torch.cat(image_reference, dim=0) + + img_reference = image_reference[0] if isinstance(image_reference, list) else image_reference + raw_ref_h, raw_ref_w = self._get_raw_image_size(img_reference) + + if raw_ref_h * raw_ref_w > 1024 * 1024: + scale = math.sqrt(1024 * 1024 / (raw_ref_h * raw_ref_w)) + image_reference = self.image_processor.resize( + image_reference, int(raw_ref_h * scale), int(raw_ref_w * scale) + ) + img_reference = image_reference[0] if isinstance(image_reference, list) else image_reference + raw_ref_h, raw_ref_w = self._get_raw_image_size(img_reference) + + image_reference_width = (raw_ref_w // multiple_of) * multiple_of + image_reference_height = (raw_ref_h // multiple_of) * multiple_of + + processed_image_reference = self.image_processor.preprocess( + image_reference, + image_reference_height, + image_reference_width, + resize_mode="crop", + ) + else: + if image_reference is not None: + bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_reference.device, image_reference.dtype) + bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + image_reference.device, image_reference.dtype + ) + processed_image_reference = (image_reference - bn_mean) / bn_std + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 4. Prepare text embeddings + prompt_embeds, text_ids = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + text_encoder_out_layers=text_encoder_out_layers, + ) + + if self.do_classifier_free_guidance: + negative_prompt = "" + if prompt is not None and isinstance(prompt, list): + negative_prompt = [negative_prompt] * len(prompt) + negative_prompt_embeds, negative_text_ids = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + text_encoder_out_layers=text_encoder_out_layers, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas: + sigmas = None + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline " + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + + latents, noise, image_latents, image_latents_encoded, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + ref_images = [image_latents_encoded] + if processed_image_reference is not None: + ref_images.append(processed_image_reference) + + condition_image_latents, condition_image_ids = self.prepare_image_latents( + ref_images, + batch_size * num_images_per_prompt, + generator, + device, + prompt_embeds.dtype, + ) + + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + mask = self.prepare_mask_latents( + mask_condition, + batch_size, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # Combine base latent position IDs with condition image position IDs. + combined_image_ids = torch.cat([latent_image_ids, condition_image_ids], dim=1) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + latent_model_input = torch.cat([latents, condition_image_latents], dim=1) + img_ids = combined_image_ids + + latent_model_input = latent_model_input.to(self.transformer.dtype) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, # (B, image_seq_len, C) + timestep=timestep / 1000, + guidance=None, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, # B, text_seq_len, 4 + img_ids=img_ids, # B, image_seq_len, 4 + joint_attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=None, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=img_ids, + joint_attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + noise_pred = neg_noise_pred + guidance_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + init_latents_proper = image_latents + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + # 8. Post-processing + latents = self._unpack_latents_with_ids(latents, latent_image_ids) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + latents.device, latents.dtype + ) + latents = latents * latents_bn_std + latents_bn_mean + latents = self._unpatchify_latents(latents) + + if output_type == "latent": + image = latents + else: + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [ + self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image + ] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return Flux2PipelineOutput(images=image) diff --git a/diffusers/pipelines/flux2/pipeline_flux2_klein_kv.py b/diffusers/pipelines/flux2/pipeline_flux2_klein_kv.py new file mode 100644 index 0000000000000000000000000000000000000000..78ed42f20afbe1d6349839151dbbb2710abffb08 --- /dev/null +++ b/diffusers/pipelines/flux2/pipeline_flux2_klein_kv.py @@ -0,0 +1,888 @@ +# Copyright 2025 Black Forest Labs and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL +import torch +from transformers import Qwen2TokenizerFast, Qwen3ForCausalLM + +from ...loaders import Flux2LoraLoaderMixin +from ...models import AutoencoderKLFlux2, Flux2Transformer2DModel +from ...models.transformers.transformer_flux2 import Flux2KVAttnProcessor, Flux2KVParallelSelfAttnProcessor +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import Flux2ImageProcessor +from .pipeline_output import Flux2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from PIL import Image + >>> from diffusers import Flux2KleinKVPipeline + + >>> pipe = Flux2KleinKVPipeline.from_pretrained( + ... "black-forest-labs/FLUX.2-klein-9b-kv", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> ref_image = Image.open("reference.png") + >>> image = pipe("A cat dressed like a wizard", image=ref_image, num_inference_steps=4).images[0] + >>> image.save("flux2_kv_output.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux2.pipeline_flux2.compute_empirical_mu +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class Flux2KleinKVPipeline(DiffusionPipeline, Flux2LoraLoaderMixin): + r""" + The Flux2 Klein KV pipeline for text-to-image generation with KV-cached reference image conditioning. + + On the first denoising step, reference image tokens are included in the forward pass and their attention K/V + projections are cached. On subsequent steps, the cached K/V are reused without recomputing, providing faster + inference when using reference images. + + Reference: + [https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence) + + Args: + transformer ([`Flux2Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLFlux2`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen3ForCausalLM`]): + [Qwen3ForCausalLM](https://huggingface.co/docs/transformers/en/model_doc/qwen3#transformers.Qwen3ForCausalLM) + tokenizer (`Qwen2TokenizerFast`): + Tokenizer of class + [Qwen2TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLFlux2, + text_encoder: Qwen3ForCausalLM, + tokenizer: Qwen2TokenizerFast, + transformer: Flux2Transformer2DModel, + is_distilled: bool = True, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = Flux2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 512 + self.default_sample_size = 128 + + # Set KV-cache-aware attention processors + self._set_kv_attn_processors() + + @staticmethod + def _get_qwen3_prompt_embeds( + text_encoder: Qwen3ForCausalLM, + tokenizer: Qwen2TokenizerFast, + prompt: str | list[str], + dtype: torch.dtype | None = None, + device: torch.device | None = None, + max_sequence_length: int = 512, + hidden_states_layers: list[int] = (9, 18, 27), + ): + dtype = text_encoder.dtype if dtype is None else dtype + device = text_encoder.device if device is None else device + + prompt = [prompt] if isinstance(prompt, str) else prompt + + all_input_ids = [] + all_attention_masks = [] + + for single_prompt in prompt: + messages = [{"role": "user", "content": single_prompt}] + text = tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=False, + ) + inputs = tokenizer( + text, + return_tensors="pt", + padding="max_length", + truncation=True, + max_length=max_sequence_length, + ) + + all_input_ids.append(inputs["input_ids"]) + all_attention_masks.append(inputs["attention_mask"]) + + input_ids = torch.cat(all_input_ids, dim=0).to(device) + attention_mask = torch.cat(all_attention_masks, dim=0).to(device) + + # Forward pass through the model + output = text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + use_cache=False, + ) + + # Only use outputs from intermediate layers and stack them + out = torch.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1) + out = out.to(dtype=dtype, device=device) + + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim) + + return prompt_embeds + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_text_ids + def _prepare_text_ids( + x: torch.Tensor, # (B, L, D) or (L, D) + t_coord: torch.Tensor | None = None, + ): + B, L, _ = x.shape + out_ids = [] + + for i in range(B): + t = torch.arange(1) if t_coord is None else t_coord[i] + h = torch.arange(1) + w = torch.arange(1) + l = torch.arange(L) + + coords = torch.cartesian_prod(t, h, w, l) + out_ids.append(coords) + + return torch.stack(out_ids) + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_latent_ids + def _prepare_latent_ids( + latents: torch.Tensor, # (B, C, H, W) + ): + r""" + Generates 4D position coordinates (T, H, W, L) for latent tensors. + + Args: + latents (torch.Tensor): + Latent tensor of shape (B, C, H, W) + + Returns: + torch.Tensor: + Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0, + H=[0..H-1], W=[0..W-1], L=0 + """ + + batch_size, _, height, width = latents.shape + + t = torch.arange(1) # [0] - time dimension + h = torch.arange(height) + w = torch.arange(width) + l = torch.arange(1) # [0] - layer dimension + + # Create position IDs: (H*W, 4) + latent_ids = torch.cartesian_prod(t, h, w, l) + + # Expand to batch: (B, H*W, 4) + latent_ids = latent_ids.unsqueeze(0).expand(batch_size, -1, -1) + + return latent_ids + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._prepare_image_ids + def _prepare_image_ids( + image_latents: list[torch.Tensor], # [(1, C, H, W), (1, C, H, W), ...] + scale: int = 10, + ): + r""" + Generates 4D time-space coordinates (T, H, W, L) for a sequence of image latents. + + This function creates a unique coordinate for every pixel/patch across all input latent with different + dimensions. + + Args: + image_latents (list[torch.Tensor]): + A list of image latent feature tensors, typically of shape (C, H, W). + scale (int, optional): + A factor used to define the time separation (T-coordinate) between latents. T-coordinate for the i-th + latent is: 'scale + scale * i'. Defaults to 10. + + Returns: + torch.Tensor: + The combined coordinate tensor. Shape: (1, N_total, 4) Where N_total is the sum of (H * W) for all + input latents. + + Coordinate Components (Dimension 4): + - T (Time): The unique index indicating which latent image the coordinate belongs to. + - H (Height): The row index within that latent image. + - W (Width): The column index within that latent image. + - L (Seq. Length): A sequence length dimension, which is always fixed at 0 (size 1) + """ + + if not isinstance(image_latents, list): + raise ValueError(f"Expected `image_latents` to be a list, got {type(image_latents)}.") + + # create time offset for each reference image + t_coords = [scale + scale * t for t in torch.arange(0, len(image_latents))] + t_coords = [t.view(-1) for t in t_coords] + + image_latent_ids = [] + for x, t in zip(image_latents, t_coords): + x = x.squeeze(0) + _, height, width = x.shape + + x_ids = torch.cartesian_prod(t, torch.arange(height), torch.arange(width), torch.arange(1)) + image_latent_ids.append(x_ids) + + image_latent_ids = torch.cat(image_latent_ids, dim=0) + image_latent_ids = image_latent_ids.unsqueeze(0) + + return image_latent_ids + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._patchify_latents + def _patchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4) + latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpatchify_latents + def _unpatchify_latents(latents): + batch_size, num_channels_latents, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width) + latents = latents.permute(0, 1, 4, 2, 5, 3) + latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._pack_latents + def _pack_latents(latents): + """ + pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels) + """ + + batch_size, num_channels, height, width = latents.shape + latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpack_latents_with_ids + def _unpack_latents_with_ids(x: torch.Tensor, x_ids: torch.Tensor) -> list[torch.Tensor]: + """ + using position ids to scatter tokens into place + """ + x_list = [] + for data, pos in zip(x, x_ids): + _, ch = data.shape # noqa: F841 + h_ids = pos[:, 1].to(torch.int64) + w_ids = pos[:, 2].to(torch.int64) + + h = torch.max(h_ids) + 1 + w = torch.max(w_ids) + 1 + + flat_ids = h_ids * w + w_ids + + out = torch.zeros((h * w, ch), device=data.device, dtype=data.dtype) + out.scatter_(0, flat_ids.unsqueeze(1).expand(-1, ch), data) + + # reshape from (H * W, C) to (H, W, C) and permute to (C, H, W) + + out = out.view(h, w, ch).permute(2, 0, 1) + x_list.append(out) + + return torch.stack(x_list, dim=0) + + def _set_kv_attn_processors(self): + """Replace default attention processors with KV-cache-aware variants.""" + for block in self.transformer.transformer_blocks: + block.attn.set_processor(Flux2KVAttnProcessor()) + for block in self.transformer.single_transformer_blocks: + block.attn.set_processor(Flux2KVParallelSelfAttnProcessor()) + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int] = (9, 18, 27), + ): + device = device or self._execution_device + + if prompt is None: + prompt = "" + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds = self._get_qwen3_prompt_embeds( + text_encoder=self.text_encoder, + tokenizer=self.tokenizer, + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + hidden_states_layers=text_encoder_out_layers, + ) + + batch_size, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = self._prepare_text_ids(prompt_embeds) + text_ids = text_ids.to(device) + return prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if image.ndim != 4: + raise ValueError(f"Expected image dims 4, got {image.ndim}.") + + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + image_latents = self._patchify_latents(image_latents) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.device, image_latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + image_latents.device, image_latents.dtype + ) + image_latents = (image_latents - latents_bn_mean) / latents_bn_std + + return image_latents + + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_latents_channels, + height, + width, + dtype, + device, + generator: torch.Generator, + latents: torch.Tensor | None = None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_latents_channels * 4, height // 2, width // 2) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + latent_ids = self._prepare_latent_ids(latents) + latent_ids = latent_ids.to(device) + + latents = self._pack_latents(latents) # [B, C, H, W] -> [B, H*W, C] + return latents, latent_ids + + # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline.prepare_image_latents + def prepare_image_latents( + self, + images: list[torch.Tensor], + batch_size, + generator: torch.Generator, + device, + dtype, + ): + image_latents = [] + for image in images: + image = image.to(device=device, dtype=dtype) + imagge_latent = self._encode_vae_image(image=image, generator=generator) + image_latents.append(imagge_latent) # (1, 128, 32, 32) + + image_latent_ids = self._prepare_image_ids(image_latents) + + # Pack each latent and concatenate + packed_latents = [] + for latent in image_latents: + # latent: (1, 128, 32, 32) + packed = self._pack_latents(latent) # (1, 1024, 128) + packed = packed.squeeze(0) # (1024, 128) - remove batch dim + packed_latents.append(packed) + + # Concatenate all reference tokens along sequence dimension + image_latents = torch.cat(packed_latents, dim=0) # (N*1024, 128) + image_latents = image_latents.unsqueeze(0) # (1, N*1024, 128) + + image_latents = image_latents.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.repeat(batch_size, 1, 1) + image_latent_ids = image_latent_ids.to(device) + + return image_latents, image_latent_ids + + def check_inputs( + self, + prompt, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if ( + height is not None + and height % (self.vae_scale_factor * 2) != 0 + or width is not None + and width % (self.vae_scale_factor * 2) != 0 + ): + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: list[PIL.Image.Image] | PIL.Image.Image | None = None, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 4, + sigmas: list[float] | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int, dict], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + text_encoder_out_layers: tuple[int] = (9, 18, 27), + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `List[PIL.Image.Image]`, *optional*): + Reference image(s) for conditioning. On the first denoising step, reference tokens are included in the + forward pass and their attention K/V are cached. On subsequent steps, the cached K/V are reused without + recomputing. + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. + height (`int`, *optional*): + The height in pixels of the generated image. + width (`int`, *optional*): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 4): + The number of denoising steps. + sigmas (`List[float]`, *optional*): + Custom sigmas for the denoising schedule. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + Generator(s) for deterministic generation. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + Output format: `"pil"` or `"np"`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a `Flux2PipelineOutput` or a plain tuple. + attention_kwargs (`dict`, *optional*): + Extra kwargs passed to attention processors. + callback_on_step_end (`Callable`, *optional*): + Callback function called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`List`, *optional*): + Tensor inputs for the callback function. + max_sequence_length (`int`, defaults to 512): + Maximum sequence length for the prompt. + text_encoder_out_layers (`tuple[int]`): + Layer indices for text encoder hidden state extraction. + + Examples: + + Returns: + [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`. + """ + + # 1. Check inputs + self.check_inputs( + prompt=prompt, + height=height, + width=width, + prompt_embeds=prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. prepare text embeddings + prompt_embeds, text_ids = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + text_encoder_out_layers=text_encoder_out_layers, + ) + + # 4. process images + if image is not None and not isinstance(image, list): + image = [image] + + condition_images = None + if image is not None: + for img in image: + self.image_processor.check_image_input(img) + + condition_images = [] + for img in image: + image_width, image_height = img.size + if image_width * image_height > 1024 * 1024: + img = self.image_processor._resize_to_target_area(img, 1024 * 1024) + image_width, image_height = img.size + + multiple_of = self.vae_scale_factor * 2 + image_width = (image_width // multiple_of) * multiple_of + image_height = (image_height // multiple_of) * multiple_of + img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop") + condition_images.append(img) + height = height or image_height + width = width or image_width + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 5. prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_ids = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_latents_channels=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + image_latents = None + image_latent_ids = None + if condition_images is not None: + image_latents, image_latent_ids = self.prepare_image_latents( + images=condition_images, + batch_size=batch_size * num_images_per_prompt, + generator=generator, + device=device, + dtype=self.vae.dtype, + ) + + # 6. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas: + sigmas = None + image_seq_len = latents.shape[1] + mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 7. Denoising loop with KV caching + # Step 0 with ref images: forward_kv_extract (full pass, cache ref K/V) + # Steps 1+: forward_kv_cached (reuse cached ref K/V) + # No ref images: standard forward + self.scheduler.set_begin_index(0) + kv_cache = None + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + if i == 0 and image_latents is not None: + # Step 0: include ref tokens, extract KV cache + latent_model_input = torch.cat([image_latents, latents], dim=1).to(self.transformer.dtype) + latent_image_ids = torch.cat([image_latent_ids, latent_ids], dim=1) + + noise_pred, kv_cache = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=None, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.attention_kwargs, + return_dict=False, + kv_cache_mode="extract", + num_ref_tokens=image_latents.shape[1], + ) + + elif kv_cache is not None: + # Steps 1+: use cached ref KV, no ref tokens in input + noise_pred = self.transformer( + hidden_states=latents.to(self.transformer.dtype), + timestep=timestep / 1000, + guidance=None, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.attention_kwargs, + return_dict=False, + kv_cache=kv_cache, + kv_cache_mode="cached", + )[0] + + else: + # No reference images: standard forward + noise_pred = self.transformer( + hidden_states=latents.to(self.transformer.dtype), + timestep=timestep / 1000, + guidance=None, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # Clean up KV cache + if kv_cache is not None: + kv_cache.clear() + + self._current_timestep = None + + latents = self._unpack_latents_with_ids(latents, latent_ids) + + latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype) + latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to( + latents.device, latents.dtype + ) + latents = latents * latents_bn_std + latents_bn_mean + latents = self._unpatchify_latents(latents) + if output_type == "latent": + image = latents + else: + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return Flux2PipelineOutput(images=image) diff --git a/diffusers/pipelines/flux2/pipeline_output.py b/diffusers/pipelines/flux2/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..34ae9b574453349fdf42c2e98680f1cee315fa0a --- /dev/null +++ b/diffusers/pipelines/flux2/pipeline_output.py @@ -0,0 +1,22 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class Flux2PipelineOutput(BaseOutput): + """ + Output class for Flux2 image generation pipelines. + + Args: + images (`list[PIL.Image.Image]` or `torch.Tensor` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array or torch tensor of shape `(batch_size, + height, width, num_channels)`. PIL images or numpy array present the denoised images of the diffusion + pipeline. Torch tensors can represent either the denoised images or the intermediate latents ready to be + passed to the decoder. + """ + + images: list[PIL.Image.Image, np.ndarray] diff --git a/diffusers/pipelines/flux2/system_messages.py b/diffusers/pipelines/flux2/system_messages.py new file mode 100644 index 0000000000000000000000000000000000000000..ecdb1371f0d49afa94788e0f02f89837469a706d --- /dev/null +++ b/diffusers/pipelines/flux2/system_messages.py @@ -0,0 +1,33 @@ +# docstyle-ignore +""" +These system prompts come from: +https://github.com/black-forest-labs/flux2/blob/5a5d316b1b42f6b59a8c9194b77c8256be848432/src/flux2/system_messages.py#L54 +""" + +# docstyle-ignore +SYSTEM_MESSAGE = """You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object +attribution and actions without speculation.""" + +# docstyle-ignore +SYSTEM_MESSAGE_UPSAMPLING_T2I = """You are an expert prompt engineer for FLUX.2 by Black Forest Labs. Rewrite user prompts to be more descriptive while strictly preserving their core subject and intent. + +Guidelines: +1. Structure: Keep structured inputs structured (enhance within fields). Convert natural language to detailed paragraphs. +2. Details: Add concrete visual specifics - form, scale, textures, materials, lighting (quality, direction, color), shadows, spatial relationships, and environmental context. +3. Text in Images: Put ALL text in quotation marks, matching the prompt's language. Always provide explicit quoted text for objects that would contain text in reality (signs, labels, screens, etc.) - without it, the model generates gibberish. + +Output only the revised prompt and nothing else.""" + +# docstyle-ignore +SYSTEM_MESSAGE_UPSAMPLING_I2I = """You are FLUX.2 by Black Forest Labs, an image-editing expert. You convert editing requests into one concise instruction (50-80 words, ~30 for brief requests). + +Rules: +- Single instruction only, no commentary +- Use clear, analytical language (avoid "whimsical," "cascading," etc.) +- Specify what changes AND what stays the same (face, lighting, composition) +- Reference actual image elements +- Turn negatives into positives ("don't change X" → "keep X") +- Make abstractions concrete ("futuristic" → "glowing cyan neon, metallic panels") +- Keep content PG-13 + +Output only the final instruction in plain text and nothing else.""" diff --git a/diffusers/pipelines/free_init_utils.py b/diffusers/pipelines/free_init_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..5f00d76274f703c7b25260ecd33832d277696d89 --- /dev/null +++ b/diffusers/pipelines/free_init_utils.py @@ -0,0 +1,186 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math + +import torch +import torch.fft as fft + +from ..utils.torch_utils import randn_tensor + + +class FreeInitMixin: + r"""Mixin class for FreeInit.""" + + def enable_free_init( + self, + num_iters: int = 3, + use_fast_sampling: bool = False, + method: str = "butterworth", + order: int = 4, + spatial_stop_frequency: float = 0.25, + temporal_stop_frequency: float = 0.25, + ): + """Enables the FreeInit mechanism as in https://huggingface.co/papers/2312.07537. + + This implementation has been adapted from the [official repository](https://github.com/TianxingWu/FreeInit). + + Args: + num_iters (`int`, *optional*, defaults to `3`): + Number of FreeInit noise re-initialization iterations. + use_fast_sampling (`bool`, *optional*, defaults to `False`): + Whether or not to speedup sampling procedure at the cost of probably lower quality results. Enables the + "Coarse-to-Fine Sampling" strategy, as mentioned in the paper, if set to `True`. + method (`str`, *optional*, defaults to `butterworth`): + Must be one of `butterworth`, `ideal` or `gaussian` to use as the filtering method for the FreeInit low + pass filter. + order (`int`, *optional*, defaults to `4`): + Order of the filter used in `butterworth` method. Larger values lead to `ideal` method behaviour + whereas lower values lead to `gaussian` method behaviour. + spatial_stop_frequency (`float`, *optional*, defaults to `0.25`): + Normalized stop frequency for spatial dimensions. Must be between 0 to 1. Referred to as `d_s` in the + original implementation. + temporal_stop_frequency (`float`, *optional*, defaults to `0.25`): + Normalized stop frequency for temporal dimensions. Must be between 0 to 1. Referred to as `d_t` in the + original implementation. + """ + self._free_init_num_iters = num_iters + self._free_init_use_fast_sampling = use_fast_sampling + self._free_init_method = method + self._free_init_order = order + self._free_init_spatial_stop_frequency = spatial_stop_frequency + self._free_init_temporal_stop_frequency = temporal_stop_frequency + + def disable_free_init(self): + """Disables the FreeInit mechanism if enabled.""" + self._free_init_num_iters = None + + @property + def free_init_enabled(self): + return hasattr(self, "_free_init_num_iters") and self._free_init_num_iters is not None + + def _get_free_init_freq_filter( + self, + shape: tuple[int, ...], + device: str | torch.dtype, + filter_type: str, + order: float, + spatial_stop_frequency: float, + temporal_stop_frequency: float, + ) -> torch.Tensor: + r"""Returns the FreeInit filter based on filter type and other input conditions.""" + + time, height, width = shape[-3], shape[-2], shape[-1] + mask = torch.zeros(shape) + + if spatial_stop_frequency == 0 or temporal_stop_frequency == 0: + return mask + + if filter_type == "butterworth": + + def retrieve_mask(x): + return 1 / (1 + (x / spatial_stop_frequency**2) ** order) + elif filter_type == "gaussian": + + def retrieve_mask(x): + return math.exp(-1 / (2 * spatial_stop_frequency**2) * x) + elif filter_type == "ideal": + + def retrieve_mask(x): + return 1 if x <= spatial_stop_frequency * 2 else 0 + else: + raise NotImplementedError("`filter_type` must be one of gaussian, butterworth or ideal") + + for t in range(time): + for h in range(height): + for w in range(width): + d_square = ( + ((spatial_stop_frequency / temporal_stop_frequency) * (2 * t / time - 1)) ** 2 + + (2 * h / height - 1) ** 2 + + (2 * w / width - 1) ** 2 + ) + mask[..., t, h, w] = retrieve_mask(d_square) + + return mask.to(device) + + def _apply_freq_filter(self, x: torch.Tensor, noise: torch.Tensor, low_pass_filter: torch.Tensor) -> torch.Tensor: + r"""Noise reinitialization.""" + # FFT + x_freq = fft.fftn(x, dim=(-3, -2, -1)) + x_freq = fft.fftshift(x_freq, dim=(-3, -2, -1)) + noise_freq = fft.fftn(noise, dim=(-3, -2, -1)) + noise_freq = fft.fftshift(noise_freq, dim=(-3, -2, -1)) + + # frequency mix + high_pass_filter = 1 - low_pass_filter + x_freq_low = x_freq * low_pass_filter + noise_freq_high = noise_freq * high_pass_filter + x_freq_mixed = x_freq_low + noise_freq_high # mix in freq domain + + # IFFT + x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-3, -2, -1)) + x_mixed = fft.ifftn(x_freq_mixed, dim=(-3, -2, -1)).real + + return x_mixed + + def _apply_free_init( + self, + latents: torch.Tensor, + free_init_iteration: int, + num_inference_steps: int, + device: torch.device, + dtype: torch.dtype, + generator: torch.Generator, + ): + if free_init_iteration == 0: + self._free_init_initial_noise = latents.detach().clone() + else: + latent_shape = latents.shape + + free_init_filter_shape = (1, *latent_shape[1:]) + free_init_freq_filter = self._get_free_init_freq_filter( + shape=free_init_filter_shape, + device=device, + filter_type=self._free_init_method, + order=self._free_init_order, + spatial_stop_frequency=self._free_init_spatial_stop_frequency, + temporal_stop_frequency=self._free_init_temporal_stop_frequency, + ) + + current_diffuse_timestep = self.scheduler.config.num_train_timesteps - 1 + diffuse_timesteps = torch.full((latent_shape[0],), current_diffuse_timestep).long() + + z_t = self.scheduler.add_noise( + original_samples=latents, noise=self._free_init_initial_noise, timesteps=diffuse_timesteps.to(device) + ).to(dtype=torch.float32) + + z_rand = randn_tensor( + shape=latent_shape, + generator=generator, + device=device, + dtype=torch.float32, + ) + latents = self._apply_freq_filter(z_t, z_rand, low_pass_filter=free_init_freq_filter) + latents = latents.to(dtype) + + # Coarse-to-Fine Sampling for faster inference (can lead to lower quality) + if self._free_init_use_fast_sampling: + num_inference_steps = max( + 1, int(num_inference_steps / self._free_init_num_iters * (free_init_iteration + 1)) + ) + + if num_inference_steps > 0: + self.scheduler.set_timesteps(num_inference_steps, device=device) + + return latents, self.scheduler.timesteps diff --git a/diffusers/pipelines/free_noise_utils.py b/diffusers/pipelines/free_noise_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f52f9928c4a452e64fc0c070a965b49d61ba6069 --- /dev/null +++ b/diffusers/pipelines/free_noise_utils.py @@ -0,0 +1,597 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import torch +import torch.nn as nn + +from ..models.attention import BasicTransformerBlock, FreeNoiseTransformerBlock +from ..models.resnet import Downsample2D, ResnetBlock2D, Upsample2D +from ..models.transformers.transformer_2d import Transformer2DModel +from ..models.unets.unet_motion_model import ( + AnimateDiffTransformer3D, + CrossAttnDownBlockMotion, + DownBlockMotion, + UpBlockMotion, +) +from ..pipelines.pipeline_utils import DiffusionPipeline +from ..utils import logging +from ..utils.torch_utils import randn_tensor + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class SplitInferenceModule(nn.Module): + r""" + A wrapper module class that splits inputs along a specified dimension before performing a forward pass. + + This module is useful when you need to perform inference on large tensors in a memory-efficient way by breaking + them into smaller chunks, processing each chunk separately, and then reassembling the results. + + Args: + module (`nn.Module`): + The underlying PyTorch module that will be applied to each chunk of split inputs. + split_size (`int`, defaults to `1`): + The size of each chunk after splitting the input tensor. + split_dim (`int`, defaults to `0`): + The dimension along which the input tensors are split. + input_kwargs_to_split (`list[str]`, defaults to `["hidden_states"]`): + A list of keyword arguments (strings) that represent the input tensors to be split. + + Workflow: + 1. The keyword arguments specified in `input_kwargs_to_split` are split into smaller chunks using + `torch.split()` along the dimension `split_dim` and with a chunk size of `split_size`. + 2. The `module` is invoked once for each split with both the split inputs and any unchanged arguments + that were passed. + 3. The output tensors from each split are concatenated back together along `split_dim` before returning. + + Example: + ```python + >>> import torch + >>> import torch.nn as nn + + >>> model = nn.Linear(1000, 1000) + >>> split_module = SplitInferenceModule(model, split_size=2, split_dim=0, input_kwargs_to_split=["input"]) + + >>> input_tensor = torch.randn(42, 1000) + >>> # Will split the tensor into 21 slices of shape [2, 1000]. + >>> output = split_module(input=input_tensor) + ``` + + It is also possible to nest `SplitInferenceModule` across different split dimensions for more complex + multi-dimensional splitting. + """ + + def __init__( + self, + module: nn.Module, + split_size: int = 1, + split_dim: int = 0, + input_kwargs_to_split: list[str] = ["hidden_states"], + ) -> None: + super().__init__() + + self.module = module + self.split_size = split_size + self.split_dim = split_dim + self.input_kwargs_to_split = set(input_kwargs_to_split) + + def forward(self, *args, **kwargs) -> torch.Tensor | tuple[torch.Tensor]: + r"""Forward method for the `SplitInferenceModule`. + + This method processes the input by splitting specified keyword arguments along a given dimension, running the + underlying module on each split, and then concatenating the results. The splitting is controlled by the + `split_size` and `split_dim` parameters specified during initialization. + + Args: + *args (`Any`): + Positional arguments that are passed directly to the `module` without modification. + **kwargs (`dict[str, torch.Tensor]`): + Keyword arguments passed to the underlying `module`. Only keyword arguments whose names match the + entries in `input_kwargs_to_split` and are of type `torch.Tensor` will be split. The remaining keyword + arguments are passed unchanged. + + Returns: + `torch.Tensor | tuple[torch.Tensor]`: + The outputs obtained from `SplitInferenceModule` are the same as if the underlying module was inferred + without it. + - If the underlying module returns a single tensor, the result will be a single concatenated tensor + along the same `split_dim` after processing all splits. + - If the underlying module returns a tuple of tensors, each element of the tuple will be concatenated + along the `split_dim` across all splits, and the final result will be a tuple of concatenated tensors. + """ + split_inputs = {} + + # 1. Split inputs that were specified during initialization and also present in passed kwargs + for key in list(kwargs.keys()): + if key not in self.input_kwargs_to_split or not torch.is_tensor(kwargs[key]): + continue + split_inputs[key] = torch.split(kwargs[key], self.split_size, self.split_dim) + kwargs.pop(key) + + # 2. Invoke forward pass across each split + results = [] + for split_input in zip(*split_inputs.values()): + inputs = dict(zip(split_inputs.keys(), split_input)) + inputs.update(kwargs) + + intermediate_tensor_or_tensor_tuple = self.module(*args, **inputs) + results.append(intermediate_tensor_or_tensor_tuple) + + # 3. Concatenate split restuls to obtain final outputs + if isinstance(results[0], torch.Tensor): + return torch.cat(results, dim=self.split_dim) + elif isinstance(results[0], tuple): + return tuple([torch.cat(x, dim=self.split_dim) for x in zip(*results)]) + else: + raise ValueError( + "In order to use the SplitInferenceModule, it is necessary for the underlying `module` to either return a torch.Tensor or a tuple of torch.Tensor's." + ) + + +class AnimateDiffFreeNoiseMixin: + r"""Mixin class for [FreeNoise](https://huggingface.co/papers/2310.15169).""" + + def _enable_free_noise_in_block(self, block: CrossAttnDownBlockMotion | DownBlockMotion | UpBlockMotion): + r"""Helper function to enable FreeNoise in transformer blocks.""" + + for motion_module in block.motion_modules: + num_transformer_blocks = len(motion_module.transformer_blocks) + + for i in range(num_transformer_blocks): + if isinstance(motion_module.transformer_blocks[i], FreeNoiseTransformerBlock): + motion_module.transformer_blocks[i].set_free_noise_properties( + self._free_noise_context_length, + self._free_noise_context_stride, + self._free_noise_weighting_scheme, + ) + else: + assert isinstance(motion_module.transformer_blocks[i], BasicTransformerBlock) + basic_transfomer_block = motion_module.transformer_blocks[i] + + motion_module.transformer_blocks[i] = FreeNoiseTransformerBlock( + dim=basic_transfomer_block.dim, + num_attention_heads=basic_transfomer_block.num_attention_heads, + attention_head_dim=basic_transfomer_block.attention_head_dim, + dropout=basic_transfomer_block.dropout, + cross_attention_dim=basic_transfomer_block.cross_attention_dim, + activation_fn=basic_transfomer_block.activation_fn, + attention_bias=basic_transfomer_block.attention_bias, + only_cross_attention=basic_transfomer_block.only_cross_attention, + double_self_attention=basic_transfomer_block.double_self_attention, + positional_embeddings=basic_transfomer_block.positional_embeddings, + num_positional_embeddings=basic_transfomer_block.num_positional_embeddings, + context_length=self._free_noise_context_length, + context_stride=self._free_noise_context_stride, + weighting_scheme=self._free_noise_weighting_scheme, + ).to(device=self.device, dtype=self.dtype) + + motion_module.transformer_blocks[i].load_state_dict( + basic_transfomer_block.state_dict(), strict=True + ) + motion_module.transformer_blocks[i].set_chunk_feed_forward( + basic_transfomer_block._chunk_size, basic_transfomer_block._chunk_dim + ) + + def _disable_free_noise_in_block(self, block: CrossAttnDownBlockMotion | DownBlockMotion | UpBlockMotion): + r"""Helper function to disable FreeNoise in transformer blocks.""" + + for motion_module in block.motion_modules: + num_transformer_blocks = len(motion_module.transformer_blocks) + + for i in range(num_transformer_blocks): + if isinstance(motion_module.transformer_blocks[i], FreeNoiseTransformerBlock): + free_noise_transfomer_block = motion_module.transformer_blocks[i] + + motion_module.transformer_blocks[i] = BasicTransformerBlock( + dim=free_noise_transfomer_block.dim, + num_attention_heads=free_noise_transfomer_block.num_attention_heads, + attention_head_dim=free_noise_transfomer_block.attention_head_dim, + dropout=free_noise_transfomer_block.dropout, + cross_attention_dim=free_noise_transfomer_block.cross_attention_dim, + activation_fn=free_noise_transfomer_block.activation_fn, + attention_bias=free_noise_transfomer_block.attention_bias, + only_cross_attention=free_noise_transfomer_block.only_cross_attention, + double_self_attention=free_noise_transfomer_block.double_self_attention, + positional_embeddings=free_noise_transfomer_block.positional_embeddings, + num_positional_embeddings=free_noise_transfomer_block.num_positional_embeddings, + ).to(device=self.device, dtype=self.dtype) + + motion_module.transformer_blocks[i].load_state_dict( + free_noise_transfomer_block.state_dict(), strict=True + ) + motion_module.transformer_blocks[i].set_chunk_feed_forward( + free_noise_transfomer_block._chunk_size, free_noise_transfomer_block._chunk_dim + ) + + def _check_inputs_free_noise( + self, + prompt, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + num_frames, + ) -> None: + if not isinstance(prompt, (str, dict)): + raise ValueError(f"Expected `prompt` to have type `str` or `dict` but found {type(prompt)=}") + + if negative_prompt is not None: + if not isinstance(negative_prompt, (str, dict)): + raise ValueError( + f"Expected `negative_prompt` to have type `str` or `dict` but found {type(negative_prompt)=}" + ) + + if prompt_embeds is not None or negative_prompt_embeds is not None: + raise ValueError("`prompt_embeds` and `negative_prompt_embeds` is not supported in FreeNoise yet.") + + frame_indices = [isinstance(x, int) for x in prompt.keys()] + frame_prompts = [isinstance(x, str) for x in prompt.values()] + min_frame = min(list(prompt.keys())) + max_frame = max(list(prompt.keys())) + + if not all(frame_indices): + raise ValueError("Expected integer keys in `prompt` dict for FreeNoise.") + if not all(frame_prompts): + raise ValueError("Expected str values in `prompt` dict for FreeNoise.") + if min_frame != 0: + raise ValueError("The minimum frame index in `prompt` dict must be 0 as a starting prompt is necessary.") + if max_frame >= num_frames: + raise ValueError( + f"The maximum frame index in `prompt` dict must be lesser than {num_frames=} and follow 0-based indexing." + ) + + def _encode_prompt_free_noise( + self, + prompt: str | dict[int, str], + num_frames: int, + device: torch.device, + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + negative_prompt: str | dict[int, str] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ) -> torch.Tensor: + if negative_prompt is None: + negative_prompt = "" + + # Ensure that we have a dictionary of prompts + if isinstance(prompt, str): + prompt = {0: prompt} + if isinstance(negative_prompt, str): + negative_prompt = {0: negative_prompt} + + self._check_inputs_free_noise(prompt, negative_prompt, prompt_embeds, negative_prompt_embeds, num_frames) + + # Sort the prompts based on frame indices + prompt = dict(sorted(prompt.items())) + negative_prompt = dict(sorted(negative_prompt.items())) + + # Ensure that we have a prompt for the last frame index + prompt[num_frames - 1] = prompt[list(prompt.keys())[-1]] + negative_prompt[num_frames - 1] = negative_prompt[list(negative_prompt.keys())[-1]] + + frame_indices = list(prompt.keys()) + frame_prompts = list(prompt.values()) + frame_negative_indices = list(negative_prompt.keys()) + frame_negative_prompts = list(negative_prompt.values()) + + # Generate and interpolate positive prompts + prompt_embeds, _ = self.encode_prompt( + prompt=frame_prompts, + device=device, + num_images_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=False, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + lora_scale=lora_scale, + clip_skip=clip_skip, + ) + + shape = (num_frames, *prompt_embeds.shape[1:]) + prompt_interpolation_embeds = prompt_embeds.new_zeros(shape) + + for i in range(len(frame_indices) - 1): + start_frame = frame_indices[i] + end_frame = frame_indices[i + 1] + start_tensor = prompt_embeds[i].unsqueeze(0) + end_tensor = prompt_embeds[i + 1].unsqueeze(0) + + prompt_interpolation_embeds[start_frame : end_frame + 1] = self._free_noise_prompt_interpolation_callback( + start_frame, end_frame, start_tensor, end_tensor + ) + + # Generate and interpolate negative prompts + negative_prompt_embeds = None + negative_prompt_interpolation_embeds = None + + if do_classifier_free_guidance: + _, negative_prompt_embeds = self.encode_prompt( + prompt=[""] * len(frame_negative_prompts), + device=device, + num_images_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=True, + negative_prompt=frame_negative_prompts, + prompt_embeds=None, + negative_prompt_embeds=None, + lora_scale=lora_scale, + clip_skip=clip_skip, + ) + + negative_prompt_interpolation_embeds = negative_prompt_embeds.new_zeros(shape) + + for i in range(len(frame_negative_indices) - 1): + start_frame = frame_negative_indices[i] + end_frame = frame_negative_indices[i + 1] + start_tensor = negative_prompt_embeds[i].unsqueeze(0) + end_tensor = negative_prompt_embeds[i + 1].unsqueeze(0) + + negative_prompt_interpolation_embeds[start_frame : end_frame + 1] = ( + self._free_noise_prompt_interpolation_callback(start_frame, end_frame, start_tensor, end_tensor) + ) + + prompt_embeds = prompt_interpolation_embeds + negative_prompt_embeds = negative_prompt_interpolation_embeds + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds, negative_prompt_embeds + + def _prepare_latents_free_noise( + self, + batch_size: int, + num_channels_latents: int, + num_frames: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + context_num_frames = ( + self._free_noise_context_length if self._free_noise_context_length == "repeat_context" else num_frames + ) + + shape = ( + batch_size, + num_channels_latents, + context_num_frames, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + if self._free_noise_noise_type == "random": + return latents + else: + if latents.size(2) == num_frames: + return latents + elif latents.size(2) != self._free_noise_context_length: + raise ValueError( + f"You have passed `latents` as a parameter to FreeNoise. The expected number of frames is either {num_frames} or {self._free_noise_context_length}, but found {latents.size(2)}" + ) + latents = latents.to(device) + + if self._free_noise_noise_type == "shuffle_context": + for i in range(self._free_noise_context_length, num_frames, self._free_noise_context_stride): + # ensure window is within bounds + window_start = max(0, i - self._free_noise_context_length) + window_end = min(num_frames, window_start + self._free_noise_context_stride) + window_length = window_end - window_start + + if window_length == 0: + break + + indices = torch.LongTensor(list(range(window_start, window_end))) + shuffled_indices = indices[torch.randperm(window_length, generator=generator)] + + current_start = i + current_end = min(num_frames, current_start + window_length) + if current_end == current_start + window_length: + # batch of frames perfectly fits the window + latents[:, :, current_start:current_end] = latents[:, :, shuffled_indices] + else: + # handle the case where the last batch of frames does not fit perfectly with the window + prefix_length = current_end - current_start + shuffled_indices = shuffled_indices[:prefix_length] + latents[:, :, current_start:current_end] = latents[:, :, shuffled_indices] + + elif self._free_noise_noise_type == "repeat_context": + num_repeats = (num_frames + self._free_noise_context_length - 1) // self._free_noise_context_length + latents = torch.cat([latents] * num_repeats, dim=2) + + latents = latents[:, :, :num_frames] + return latents + + def _lerp( + self, start_index: int, end_index: int, start_tensor: torch.Tensor, end_tensor: torch.Tensor + ) -> torch.Tensor: + num_indices = end_index - start_index + 1 + interpolated_tensors = [] + + for i in range(num_indices): + alpha = i / (num_indices - 1) + interpolated_tensor = (1 - alpha) * start_tensor + alpha * end_tensor + interpolated_tensors.append(interpolated_tensor) + + interpolated_tensors = torch.cat(interpolated_tensors) + return interpolated_tensors + + def enable_free_noise( + self, + context_length: int | None = 16, + context_stride: int = 4, + weighting_scheme: str = "pyramid", + noise_type: str = "shuffle_context", + prompt_interpolation_callback: Callable[ + [DiffusionPipeline, int, int, torch.Tensor, torch.Tensor], torch.Tensor + ] + | None = None, + ) -> None: + r""" + Enable long video generation using FreeNoise. + + Args: + context_length (`int`, defaults to `16`, *optional*): + The number of video frames to process at once. It's recommended to set this to the maximum frames the + Motion Adapter was trained with (usually 16/24/32). If `None`, the default value from the motion + adapter config is used. + context_stride (`int`, *optional*): + Long videos are generated by processing many frames. FreeNoise processes these frames in sliding + windows of size `context_length`. Context stride allows you to specify how many frames to skip between + each window. For example, a context length of 16 and context stride of 4 would process 24 frames as: + [0, 15], [4, 19], [8, 23] (0-based indexing) + weighting_scheme (`str`, defaults to `pyramid`): + Weighting scheme for averaging latents after accumulation in FreeNoise blocks. The following weighting + schemes are supported currently: + - "flat" + Performs weighting averaging with a flat weight pattern: [1, 1, 1, 1, 1]. + - "pyramid" + Performs weighted averaging with a pyramid like weight pattern: [1, 2, 3, 2, 1]. + - "delayed_reverse_sawtooth" + Performs weighted averaging with low weights for earlier frames and high-to-low weights for + later frames: [0.01, 0.01, 3, 2, 1]. + noise_type (`str`, defaults to "shuffle_context"): + Must be one of ["shuffle_context", "repeat_context", "random"]. + - "shuffle_context" + Shuffles a fixed batch of `context_length` latents to create a final latent of size + `num_frames`. This is usually the best setting for most generation scenarios. However, there + might be visible repetition noticeable in the kinds of motion/animation generated. + - "repeated_context" + Repeats a fixed batch of `context_length` latents to create a final latent of size + `num_frames`. + - "random" + The final latents are random without any repetition. + """ + + allowed_weighting_scheme = ["flat", "pyramid", "delayed_reverse_sawtooth"] + allowed_noise_type = ["shuffle_context", "repeat_context", "random"] + + if context_length > self.motion_adapter.config.motion_max_seq_length: + logger.warning( + f"You have set {context_length=} which is greater than {self.motion_adapter.config.motion_max_seq_length=}. This can lead to bad generation results." + ) + if weighting_scheme not in allowed_weighting_scheme: + raise ValueError( + f"The parameter `weighting_scheme` must be one of {allowed_weighting_scheme}, but got {weighting_scheme=}" + ) + if noise_type not in allowed_noise_type: + raise ValueError(f"The parameter `noise_type` must be one of {allowed_noise_type}, but got {noise_type=}") + + self._free_noise_context_length = context_length or self.motion_adapter.config.motion_max_seq_length + self._free_noise_context_stride = context_stride + self._free_noise_weighting_scheme = weighting_scheme + self._free_noise_noise_type = noise_type + self._free_noise_prompt_interpolation_callback = prompt_interpolation_callback or self._lerp + + if hasattr(self.unet.mid_block, "motion_modules"): + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + else: + blocks = [*self.unet.down_blocks, *self.unet.up_blocks] + + for block in blocks: + self._enable_free_noise_in_block(block) + + def disable_free_noise(self) -> None: + r"""Disable the FreeNoise sampling mechanism.""" + self._free_noise_context_length = None + + if hasattr(self.unet.mid_block, "motion_modules"): + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + else: + blocks = [*self.unet.down_blocks, *self.unet.up_blocks] + + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + for block in blocks: + self._disable_free_noise_in_block(block) + + def _enable_split_inference_motion_modules_( + self, motion_modules: list[AnimateDiffTransformer3D], spatial_split_size: int + ) -> None: + for motion_module in motion_modules: + motion_module.proj_in = SplitInferenceModule(motion_module.proj_in, spatial_split_size, 0, ["input"]) + + for i in range(len(motion_module.transformer_blocks)): + motion_module.transformer_blocks[i] = SplitInferenceModule( + motion_module.transformer_blocks[i], + spatial_split_size, + 0, + ["hidden_states", "encoder_hidden_states"], + ) + + motion_module.proj_out = SplitInferenceModule(motion_module.proj_out, spatial_split_size, 0, ["input"]) + + def _enable_split_inference_attentions_( + self, attentions: list[Transformer2DModel], temporal_split_size: int + ) -> None: + for i in range(len(attentions)): + attentions[i] = SplitInferenceModule( + attentions[i], temporal_split_size, 0, ["hidden_states", "encoder_hidden_states"] + ) + + def _enable_split_inference_resnets_(self, resnets: list[ResnetBlock2D], temporal_split_size: int) -> None: + for i in range(len(resnets)): + resnets[i] = SplitInferenceModule(resnets[i], temporal_split_size, 0, ["input_tensor", "temb"]) + + def _enable_split_inference_samplers_( + self, samplers: list[Downsample2D] | list[Upsample2D], temporal_split_size: int + ) -> None: + for i in range(len(samplers)): + samplers[i] = SplitInferenceModule(samplers[i], temporal_split_size, 0, ["hidden_states"]) + + def enable_free_noise_split_inference(self, spatial_split_size: int = 256, temporal_split_size: int = 16) -> None: + r""" + Enable FreeNoise memory optimizations by utilizing + [`~diffusers.pipelines.free_noise_utils.SplitInferenceModule`] across different intermediate modeling blocks. + + Args: + spatial_split_size (`int`, defaults to `256`): + The split size across spatial dimensions for internal blocks. This is used in facilitating split + inference across the effective batch dimension (`[B x H x W, F, C]`) of intermediate tensors in motion + modeling blocks. + temporal_split_size (`int`, defaults to `16`): + The split size across temporal dimensions for internal blocks. This is used in facilitating split + inference across the effective batch dimension (`[B x F, H x W, C]`) of intermediate tensors in spatial + attention, resnets, downsampling and upsampling blocks. + """ + # TODO(aryan): Discuss on what's the best way to provide more control to users + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + for block in blocks: + if getattr(block, "motion_modules", None) is not None: + self._enable_split_inference_motion_modules_(block.motion_modules, spatial_split_size) + if getattr(block, "attentions", None) is not None: + self._enable_split_inference_attentions_(block.attentions, temporal_split_size) + if getattr(block, "resnets", None) is not None: + self._enable_split_inference_resnets_(block.resnets, temporal_split_size) + if getattr(block, "downsamplers", None) is not None: + self._enable_split_inference_samplers_(block.downsamplers, temporal_split_size) + if getattr(block, "upsamplers", None) is not None: + self._enable_split_inference_samplers_(block.upsamplers, temporal_split_size) + + @property + def free_noise_enabled(self): + return hasattr(self, "_free_noise_context_length") and self._free_noise_context_length is not None diff --git a/diffusers/pipelines/glm_image/__init__.py b/diffusers/pipelines/glm_image/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..140b9cc760cc1b163dbb7027e29d4a03f3c787d8 --- /dev/null +++ b/diffusers/pipelines/glm_image/__init__.py @@ -0,0 +1,59 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, + is_transformers_version, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["GlmImagePipelineOutput"]} + +# Import transformers components so they can be resolved during pipeline loading + +if is_transformers_available() and is_transformers_version(">=", "4.57.4"): + try: + from transformers import GlmImageForConditionalGeneration, GlmImageProcessor + + _additional_imports["GlmImageForConditionalGeneration"] = GlmImageForConditionalGeneration + _additional_imports["GlmImageProcessor"] = GlmImageProcessor + except ImportError: + pass + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_glm_image"] = ["GlmImagePipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_glm_image import GlmImagePipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/glm_image/pipeline_glm_image.py b/diffusers/pipelines/glm_image/pipeline_glm_image.py new file mode 100644 index 0000000000000000000000000000000000000000..8794e8195771552a7c79d9a67e41befb7dffbbad --- /dev/null +++ b/diffusers/pipelines/glm_image/pipeline_glm_image.py @@ -0,0 +1,1079 @@ +# Copyright 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import re +from typing import Any, Callable + +import numpy as np +import PIL +import torch +from transformers import ByT5Tokenizer, PreTrainedModel, ProcessorMixin, T5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKL, GlmImageTransformer2DModel +from ...models.transformers.transformer_glm_image import GlmImageKVCache +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, is_transformers_version, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import GlmImagePipelineOutput + + +# Because it's not released in stable as of 13/01/2026. So this is just a proxy. +GlmImageProcessor = ProcessorMixin +GlmImageForConditionalGeneration = PreTrainedModel +if is_transformers_version(">=", "5.0.0.dev0"): + from transformers import GlmImageForConditionalGeneration, GlmImageProcessor + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import GlmImagePipeline + + >>> pipe = GlmImagePipeline.from_pretrained("zai-org/GLM-Image", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "A photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + >>> image.save("output.png") + ``` +""" + + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + base_shift: float = 0.25, + max_shift: float = 0.75, +) -> float: + m = (image_seq_len / base_seq_len) ** 0.5 + mu = m * max_shift + base_shift + return mu + + +# Copied from diffusers.pipelines.cogview4.pipeline_cogview4.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + accepts_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + + if timesteps is not None and sigmas is not None: + if not accepts_timesteps and not accepts_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep or sigma schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif timesteps is not None and sigmas is None: + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif timesteps is None and sigmas is not None: + if not accepts_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class GlmImagePipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using GLM-Image. + + This pipeline integrates both the AR (autoregressive) model for token generation and the DiT (diffusion + transformer) model for image decoding. + + Args: + tokenizer (`PreTrainedTokenizer`): + Tokenizer for the text encoder. + processor (`AutoProcessor`): + Processor for the AR model to handle chat templates and tokenization. + text_encoder ([`T5EncoderModel`]): + Frozen text-encoder for glyph embeddings. + vision_language_encoder ([`GlmImageForConditionalGeneration`]): + The AR model that generates image tokens from text prompts. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + transformer ([`GlmImageTransformer2DModel`]): + A text conditioned transformer to denoise the encoded image latents (DiT). + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + """ + + _optional_components = [] + model_cpu_offload_seq = "vision_language_encoder->text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + tokenizer: ByT5Tokenizer, + processor: GlmImageProcessor, + text_encoder: T5EncoderModel, + vision_language_encoder: GlmImageForConditionalGeneration, + vae: AutoencoderKL, + transformer: GlmImageTransformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, + processor=processor, + text_encoder=text_encoder, + vision_language_encoder=vision_language_encoder, + vae=vae, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") + and self.transformer is not None + and hasattr(self.transformer.config, "sample_size") + else 128 + ) + + @staticmethod + def _compute_generation_params( + image_grid_thw, + is_text_to_image: bool, + ): + grid_sizes = [] + grid_hw = [] + + for i in range(image_grid_thw.shape[0]): + t, h, w = image_grid_thw[i].tolist() + grid_sizes.append(int(h * w)) + grid_hw.append((int(h), int(w))) + + if not is_text_to_image: + max_new_tokens = grid_sizes[-1] + 1 + large_image_start_offset = 0 + target_grid_h, target_grid_w = grid_hw[-1] + else: + total_tokens = sum(grid_sizes) + max_new_tokens = total_tokens + 1 + large_image_start_offset = sum(grid_sizes[1:]) + target_grid_h, target_grid_w = grid_hw[0] + return max_new_tokens, large_image_start_offset, target_grid_h, target_grid_w + + @staticmethod + def _extract_large_image_tokens( + outputs: torch.Tensor, input_length: int, large_image_start_offset: int, large_image_tokens: int + ) -> torch.Tensor: + generated_tokens = outputs[0][input_length:] + large_image_start = large_image_start_offset + large_image_end = large_image_start + large_image_tokens + return generated_tokens[large_image_start:large_image_end] + + @staticmethod + def _upsample_token_ids(token_ids: torch.Tensor, token_h: int, token_w: int) -> torch.Tensor: + token_ids = token_ids.view(1, 1, token_h, token_w) + token_ids = torch.nn.functional.interpolate(token_ids.float(), scale_factor=2, mode="nearest").to( + dtype=torch.long + ) + token_ids = token_ids.view(1, -1) + return token_ids + + @staticmethod + def _validate_and_normalize_images( + image: list[PIL.Image.Image] | list[list[PIL.Image.Image]], + batch_size: int, + ) -> list[list[PIL.Image.Image]]: + """ + Validate and normalize image inputs to List[List[PIL.Image]]. + + Rules: + - batch_size > 1: Only accepts List[List[PIL.Image]], each sublist must have equal length + - batch_size == 1: Accepts List[PIL.Image] for legacy compatibility (converted to [[img1, img2, ...]]) + - Other formats raise ValueError + + Args: + image: Input images in various formats + batch_size: Number of prompts in the batch + + Returns: + Normalized images as List[List[PIL.Image]], or None if no images provided + """ + if image is None or len(image) == 0: + return None + + first_element = image[0] + + if batch_size == 1: + # Legacy format: List[PIL.Image] -> [[img1, img2, ...]] + if not isinstance(first_element, (list, tuple)): + return [list(image)] + # Already in List[List[PIL.Image]] format + if len(image) != 1: + raise ValueError( + f"For batch_size=1 with List[List[PIL.Image]] format, expected 1 image list, got {len(image)}." + ) + return [list(image[0])] + + # batch_size > 1: must be List[List[PIL.Image]] + if not isinstance(first_element, (list, tuple)): + raise ValueError( + f"For batch_size > 1, images must be List[List[PIL.Image]] format. " + f"Got List[{type(first_element).__name__}] instead. " + f"Each prompt requires its own list of condition images." + ) + + if len(image) != batch_size: + raise ValueError(f"Number of image lists ({len(image)}) must match batch size ({batch_size}).") + + # Validate homogeneous: all sublists must have same length + num_input_images_per_prompt = len(image[0]) + for idx, imgs in enumerate(image): + if len(imgs) != num_input_images_per_prompt: + raise ValueError( + f"All prompts must have the same number of condition images. " + f"Prompt 0 has {num_input_images_per_prompt} images, but prompt {idx} has {len(imgs)} images." + ) + + return [list(imgs) for imgs in image] + + def generate_prior_tokens( + self, + prompt: str | list[str], + height: int, + width: int, + image: list[list[PIL.Image.Image]] | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + ): + """ + Generate prior tokens for the DiT model using the AR model. + + Args: + prompt: Single prompt or list of prompts + height: Target image height + width: Target image width + image: Normalized image input as List[List[PIL.Image]]. Should be pre-validated + using _validate_and_normalize_images() before calling this method. + device: Target device + generator: Random generator for reproducibility + + Returns: + Tuple of: + - prior_token_ids: Tensor of shape (batch_size, num_tokens) with upsampled prior tokens + - prior_token_image_ids_per_sample: List of tensors, one per sample. Each tensor contains + the upsampled prior token ids for all condition images in that sample. None for t2i. + - source_image_grid_thw_per_sample: List of tensors, one per sample. Each tensor has shape + (num_condition_images, 3) with upsampled grid info. None for t2i. + """ + device = device or self._execution_device + + # Normalize prompt to list format + prompt_list = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt_list) + + # Image is already normalized by _validate_and_normalize_images(): None or List[List[PIL.Image]] + is_text_to_image = image is None + # Build messages for each sample in the batch + all_messages = [] + for idx, p in enumerate(prompt_list): + content = [] + if not is_text_to_image: + for img in image[idx]: + content.append({"type": "image", "image": img}) + content.append({"type": "text", "text": p}) + all_messages.append([{"role": "user", "content": content}]) + # Process with the processor (supports batch with left padding) + inputs = self.processor.apply_chat_template( + all_messages, + tokenize=True, + padding=True if batch_size > 1 else False, + target_h=height, + target_w=width, + return_dict=True, + return_tensors="pt", + ).to(device) + + image_grid_thw = inputs.get("image_grid_thw") + images_per_sample = inputs.get("images_per_sample") + + # Determine number of condition images and grids per sample + num_condition_images = 0 if is_text_to_image else len(image[0]) + if images_per_sample is not None: + num_grids_per_sample = images_per_sample[0].item() + else: + # Fallback for batch_size=1: total grids is for single sample + num_grids_per_sample = image_grid_thw.shape[0] + + # Compute generation params (same for all samples in homogeneous batch) + first_sample_grids = image_grid_thw[:num_grids_per_sample] + max_new_tokens, large_image_offset, token_h, token_w = self._compute_generation_params( + image_grid_thw=first_sample_grids, is_text_to_image=is_text_to_image + ) + + # Generate source image tokens (prior_token_image_ids) for i2i mode + prior_token_image_ids = None + source_image_grid_thw = None + if not is_text_to_image: + # Extract source grids by selecting condition image indices (skip target grids) + # Grid order from processor: [s0_cond1, s0_cond2, ..., s0_target, s1_cond1, s1_cond2, ..., s1_target, ...] + # We need indices: [0, 1, ..., num_condition_images-1, num_grids_per_sample, num_grids_per_sample+1, ...] + source_indices = [] + for sample_idx in range(batch_size): + base = sample_idx * num_grids_per_sample + source_indices.extend(range(base, base + num_condition_images)) + source_grids = image_grid_thw[source_indices] + + if len(source_grids) > 0: + prior_token_image_embed = self.vision_language_encoder.get_image_features( + inputs["pixel_values"], source_grids + ).pooler_output + prior_token_image_embed = torch.cat(prior_token_image_embed, dim=0) + prior_token_image_ids_d32 = self.vision_language_encoder.get_image_tokens( + prior_token_image_embed, source_grids + ) + # Upsample each source image's prior tokens to match VAE/DiT resolution + split_sizes = source_grids.prod(dim=-1).tolist() + prior_ids_per_source = torch.split(prior_token_image_ids_d32, split_sizes) + upsampled_prior_ids = [] + for i, prior_ids in enumerate(prior_ids_per_source): + t, h, w = source_grids[i].tolist() + upsampled = self._upsample_token_ids(prior_ids, int(h), int(w)) + upsampled_prior_ids.append(upsampled.squeeze(0)) + prior_token_image_ids = torch.cat(upsampled_prior_ids, dim=0) + # Upsample grid dimensions for later splitting + upsampled_grids = source_grids.clone() + upsampled_grids[:, 1] = upsampled_grids[:, 1] * 2 + upsampled_grids[:, 2] = upsampled_grids[:, 2] * 2 + source_image_grid_thw = upsampled_grids + + # Generate with AR model + # Set torch random seed from generator for reproducibility + # (transformers generate() doesn't accept generator parameter) + if generator is not None: + seed = generator.initial_seed() + torch.manual_seed(seed) + if device is not None and device.type == "cuda": + torch.cuda.manual_seed(seed) + outputs = self.vision_language_encoder.generate( + **inputs, + max_new_tokens=max_new_tokens, + do_sample=True, + ) + + # Extract and upsample prior tokens for each sample + # For left-padded inputs, generated tokens start after the padded input sequence + all_prior_token_ids = [] + max_input_length = inputs["input_ids"].shape[-1] + for idx in range(batch_size): + # For left-padded sequences, generated tokens start at max_input_length + # (padding is on the left, so all sequences end at the same position) + prior_token_ids_d32 = self._extract_large_image_tokens( + outputs[idx : idx + 1], max_input_length, large_image_offset, token_h * token_w + ) + prior_token_ids = self._upsample_token_ids(prior_token_ids_d32, token_h, token_w) + all_prior_token_ids.append(prior_token_ids) + prior_token_ids = torch.cat(all_prior_token_ids, dim=0) + + # Split prior_token_image_ids and source_image_grid_thw into per-sample lists for easier consumption + prior_token_image_ids_per_sample = None + source_image_grid_thw_per_sample = None + if prior_token_image_ids is not None and source_image_grid_thw is not None: + # Split grids: each sample has num_condition_images grids + source_image_grid_thw_per_sample = list(torch.split(source_image_grid_thw, num_condition_images)) + # Split prior_token_image_ids: tokens per sample may vary due to different image sizes + tokens_per_image = source_image_grid_thw.prod(dim=-1).tolist() + tokens_per_sample = [] + for i in range(batch_size): + start_idx = i * num_condition_images + end_idx = start_idx + num_condition_images + tokens_per_sample.append(sum(tokens_per_image[start_idx:end_idx])) + prior_token_image_ids_per_sample = list(torch.split(prior_token_image_ids, tokens_per_sample)) + + return prior_token_ids, prior_token_image_ids_per_sample, source_image_grid_thw_per_sample + + def get_glyph_texts(self, prompt): + """Extract glyph texts from prompt(s). Returns a list of lists for batch processing.""" + if isinstance(prompt, str): + prompt = [prompt] + all_ocr_texts = [] + for p in prompt: + ocr_texts = ( + re.findall(r"'([^']*)'", p) + + re.findall(r"\u201c([^\u201c\u201d]*)\u201d", p) + + re.findall(r'"([^"]*)"', p) + + re.findall(r"「([^「」]*)」", p) + ) + all_ocr_texts.append(ocr_texts) + return all_ocr_texts + + def _get_glyph_embeds( + self, + prompt: str | list[str] = None, + max_sequence_length: int = 2048, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + """Get glyph embeddings for each prompt in the batch.""" + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + # get_glyph_texts now returns a list of lists (one per prompt) + all_glyph_texts = self.get_glyph_texts(prompt) + + all_glyph_embeds = [] + for glyph_texts in all_glyph_texts: + if len(glyph_texts) == 0: + glyph_texts = [""] + input_ids = self.tokenizer( + glyph_texts, + max_length=max_sequence_length, + truncation=True, + ).input_ids + input_ids = [ + [self.tokenizer.pad_token_id] * ((len(input_ids) + 1) % 2) + input_ids_ for input_ids_ in input_ids + ] + max_length = max(len(input_ids_) for input_ids_ in input_ids) + attention_mask = torch.tensor( + [[1] * len(input_ids_) + [0] * (max_length - len(input_ids_)) for input_ids_ in input_ids], + device=device, + ) + input_ids = torch.tensor( + [ + input_ids_ + [self.tokenizer.pad_token_id] * (max_length - len(input_ids_)) + for input_ids_ in input_ids + ], + device=device, + ) + outputs = self.text_encoder(input_ids, attention_mask=attention_mask) + glyph_embeds = outputs.last_hidden_state[attention_mask.bool()].unsqueeze(0) + all_glyph_embeds.append(glyph_embeds) + + # Pad to same sequence length and stack (use left padding to match transformers) + max_seq_len = max(emb.size(1) for emb in all_glyph_embeds) + padded_embeds = [] + for emb in all_glyph_embeds: + if emb.size(1) < max_seq_len: + pad = torch.zeros(emb.size(0), max_seq_len - emb.size(1), emb.size(2), device=device, dtype=emb.dtype) + emb = torch.cat([pad, emb], dim=1) # left padding + padded_embeds.append(emb) + + glyph_embeds = torch.cat(padded_embeds, dim=0) + return glyph_embeds.to(device=device, dtype=dtype) + + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 2048, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_images_per_prompt (`int`, *optional*, defaults to 1): + Number of images that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + max_sequence_length (`int`, defaults to `2048`): + Maximum sequence length in encoded prompt. Can be set to other values but may lead to poorer results. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_glyph_embeds(prompt, max_sequence_length, device, dtype) + + # Repeat embeddings for num_images_per_prompt + if num_images_per_prompt > 1: + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + # For GLM-Image, negative_prompt must be "" instead of None + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_embeds = self._get_glyph_embeds(negative_prompt, max_sequence_length, device, dtype) + + if num_images_per_prompt > 1: + negative_prompt_embeds = negative_prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + return prompt_embeds, negative_prompt_embeds + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + if latents is not None: + return latents.to(device) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs, + prompt_embeds=None, + negative_prompt_embeds=None, + prior_token_ids=None, + prior_token_image_ids=None, + source_image_grid_thw=None, + image=None, + ): + if ( + height is not None + and height % (self.vae_scale_factor * self.transformer.config.patch_size * 2) != 0 + or width is not None + and width % (self.transformer.config.patch_size * 2) != 0 + ): + # GLM-Image uses 32× downsampling, so the image dimensions must be multiples of 32. + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 4} but are {height} and {width}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + if prompt is None and prior_token_ids is None: + raise ValueError( + "Provide either `prompt` or `prior_token_ids`. Cannot leave both `prompt` and `prior_token_ids` undefined." + ) + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + # Validate prior token inputs: for i2i mode, all three must be provided together + # For t2i mode, only prior_token_ids is needed (prior_token_image_ids and source_image_grid_thw should be None) + prior_image_inputs = [prior_token_image_ids, source_image_grid_thw] + num_prior_image_inputs = sum(x is not None for x in prior_image_inputs) + if num_prior_image_inputs > 0 and num_prior_image_inputs < len(prior_image_inputs): + raise ValueError( + "`prior_token_image_ids` and `source_image_grid_thw` must be provided together for i2i mode. " + f"Got prior_token_image_ids={prior_token_image_ids is not None}, " + f"source_image_grid_thw={source_image_grid_thw is not None}." + ) + if num_prior_image_inputs > 0 and prior_token_ids is None: + raise ValueError( + "`prior_token_ids` must be provided when `prior_token_image_ids` and `source_image_grid_thw` are provided." + ) + if num_prior_image_inputs > 0 and image is None: + raise ValueError( + "`image` must be provided when `prior_token_image_ids` and `source_image_grid_thw` are provided " + "for i2i mode, as the images are needed for VAE encoding to build the KV cache." + ) + + if prior_token_ids is not None and prompt_embeds is None and prompt is None: + raise ValueError("`prompt_embeds` or `prompt` must also be provided with `prior_token_ids`.") + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + image: torch.Tensor + | PIL.Image.Image + | np.ndarray + | list[torch.Tensor] + | list[PIL.Image.Image] + | list[np.ndarray] + | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + guidance_scale: float = 1.5, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prior_token_ids: torch.Tensor | None = None, + prior_token_image_ids: list[torch.Tensor] | None = None, + source_image_grid_thw: list[torch.Tensor] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int, dict], None] + | PipelineCallback + | MultiPipelineCallbacks + | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 2048, + ) -> GlmImagePipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. Must contain shape info in the format 'H + W' where H and W are token dimensions (d32). Example: "A beautiful sunset36 24" + generates a 1152x768 image. + image: Optional condition images for image-to-image generation. + height (`int`, *optional*): + The height in pixels. If not provided, derived from prompt shape info. + width (`int`, *optional*): + The width in pixels. If not provided, derived from prompt shape info. + num_inference_steps (`int`, *optional*, defaults to `50`): + The number of denoising steps for DiT. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, the scheduler's default schedule for + `num_inference_steps` is used. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process. If not defined, the scheduler's default schedule is + used. + guidance_scale (`float`, *optional*, defaults to `1.5`): + Guidance scale for classifier-free guidance. + num_images_per_prompt (`int`, *optional*, defaults to `1`): + The number of images to generate per prompt. + generator (`torch.Generator`, *optional*): + Random generator for reproducibility. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents to be used as inputs for image generation. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. If not provided, embeddings are generated from `prompt`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Used when classifier-free guidance is enabled. + prior_token_ids (`torch.Tensor`, *optional*): + Pre-generated prior token ids from `generate_prior_tokens`. If supplied, prior generation is skipped. + prior_token_image_ids (`list[torch.Tensor]`, *optional*): + Image token ids associated with `prior_token_ids`. + source_image_grid_thw (`list[torch.Tensor]`, *optional*): + Per-sample THW grid information for the source image tokens. + crops_coords_top_left (`tuple[int, int]`, *optional*, defaults to `(0, 0)`): + The top-left coordinates of the crop used for conditioning embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + Output format: "pil", "np", or "latent". + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`GlmImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor`. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + Tensor inputs passed to `callback_on_step_end`. + max_sequence_length (`int`, *optional*, defaults to `2048`): + Maximum sequence length for the text encoder. + + Examples: + + Returns: + [`GlmImagePipelineOutput`] or `tuple`: Generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs + self.check_inputs( + prompt, + height, + width, + callback_on_step_end_tensor_inputs, + prompt_embeds, + negative_prompt_embeds, + prior_token_ids, + prior_token_image_ids, + source_image_grid_thw, + image, + ) + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 2. Validate and normalize image format + normalized_image = self._validate_and_normalize_images(image, batch_size) + + # 3. Generate prior tokens (batch mode) + # Get a single generator for AR model (use first if list provided) + ar_generator = generator[0] if isinstance(generator, list) else generator + if prior_token_ids is None: + prior_token_ids, prior_token_image_ids_per_sample, source_image_grid_thw_per_sample = ( + self.generate_prior_tokens( + prompt=prompt, + image=normalized_image, + height=height, + width=width, + device=device, + generator=ar_generator, + ) + ) + else: + # User provided prior_token_ids directly (from generate_prior_tokens) + prior_token_image_ids_per_sample = prior_token_image_ids + source_image_grid_thw_per_sample = source_image_grid_thw + + # 4. Preprocess images for VAE encoding + preprocessed_images = None + if normalized_image is not None: + preprocessed_images = [] + for prompt_images in normalized_image: + prompt_preprocessed = [] + for img in prompt_images: + image_height, image_width = img.size[::-1] if isinstance(img, PIL.Image.Image) else img.shape[:2] + multiple_of = self.vae_scale_factor * self.transformer.config.patch_size + image_height = (image_height // multiple_of) * multiple_of + image_width = (image_width // multiple_of) * multiple_of + img = self.image_processor.preprocess(img, height=image_height, width=image_width) + prompt_preprocessed.append(img) + height = height or image_height + width = width or image_width + preprocessed_images.append(prompt_preprocessed) + + # 5. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + num_images_per_prompt=num_images_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + dtype=self.dtype, + ) + + # 6. Prepare latents and (optional) image kv cache + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_channels_latents=latent_channels, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + kv_caches = GlmImageKVCache(num_layers=self.transformer.config.num_layers) + + if normalized_image is not None: + kv_caches.set_mode("write") + latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.latent_channels, 1, 1) + latents_std = torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.latent_channels, 1, 1) + + latents_mean = latents_mean.to(device=device, dtype=prompt_embeds.dtype) + latents_std = latents_std.to(device=device, dtype=prompt_embeds.dtype) + + # Process each sample's condition images + for prompt_idx in range(batch_size): + prompt_images = preprocessed_images[prompt_idx] + prompt_prior_ids = prior_token_image_ids_per_sample[prompt_idx] + prompt_grid_thw = source_image_grid_thw_per_sample[prompt_idx] + + # Split this sample's prior_token_image_ids by each image's token count + split_sizes = prompt_grid_thw.prod(dim=-1).tolist() + prior_ids_per_image = torch.split(prompt_prior_ids, split_sizes) + # Process each condition image for this sample + for condition_image, condition_image_prior_token_id in zip(prompt_images, prior_ids_per_image): + condition_image = condition_image.to(device=device, dtype=prompt_embeds.dtype) + condition_latent = retrieve_latents( + self.vae.encode(condition_image), generator=generator, sample_mode="argmax" + ) + condition_latent = (condition_latent - latents_mean) / latents_std + + _ = self.transformer( + hidden_states=condition_latent, + encoder_hidden_states=torch.zeros_like(prompt_embeds)[:1, :0, ...], + prior_token_id=condition_image_prior_token_id, + prior_token_drop=torch.full_like(condition_image_prior_token_id, False, dtype=torch.bool), + timestep=torch.zeros((1,), device=device), + target_size=torch.tensor([condition_image.shape[-2:]], device=device), + crop_coords=torch.zeros((1, 2), device=device), + attention_kwargs=attention_kwargs, + kv_caches=kv_caches, + ) + # Move to next sample's cache slot + kv_caches.next_sample() + + # 7. Prepare additional timestep conditions + target_size = (height, width) + target_size = torch.tensor([target_size], dtype=prompt_embeds.dtype, device=device) + crops_coords_top_left = torch.tensor([crops_coords_top_left], dtype=prompt_embeds.dtype, device=device) + + target_size = target_size.repeat(batch_size * num_images_per_prompt, 1) + crops_coords_top_left = crops_coords_top_left.repeat(batch_size * num_images_per_prompt, 1) + + # Prepare timesteps + image_seq_len = ((height // self.vae_scale_factor) * (width // self.vae_scale_factor)) // ( + self.transformer.config.patch_size**2 + ) + timesteps = ( + np.linspace(self.scheduler.config.num_train_timesteps, 1.0, num_inference_steps + 1)[:-1] + if timesteps is None + else np.array(timesteps) + ) + timesteps = timesteps.astype(np.int64).astype(np.float32) + sigmas = timesteps / self.scheduler.config.num_train_timesteps if sigmas is None else sigmas + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("base_shift", 0.25), + self.scheduler.config.get("max_shift", 0.75), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas, mu=mu + ) + self._num_timesteps = len(timesteps) + + # 8. Denoising loop + transformer_dtype = self.transformer.dtype + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # Repeat prior_token_ids for num_images_per_prompt + if num_images_per_prompt > 1: + prior_token_ids = prior_token_ids.repeat_interleave(num_images_per_prompt, dim=0) + prior_token_drop_cond = torch.full_like(prior_token_ids, False, dtype=torch.bool) + prior_token_drop_uncond = torch.full_like(prior_token_ids, True, dtype=torch.bool) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + + timestep = t.expand(latents.shape[0]) - 1 + + if prior_token_image_ids_per_sample is not None: + kv_caches.set_mode("read") + + noise_pred_cond = self.transformer( + hidden_states=latent_model_input, + encoder_hidden_states=prompt_embeds, + prior_token_id=prior_token_ids, + prior_token_drop=prior_token_drop_cond, + timestep=timestep, + target_size=target_size, + crop_coords=crops_coords_top_left, + attention_kwargs=attention_kwargs, + return_dict=False, + kv_caches=kv_caches, + )[0].float() + + # perform guidance + if self.do_classifier_free_guidance: + if prior_token_image_ids_per_sample is not None: + kv_caches.set_mode("skip") + noise_pred_uncond = self.transformer( + hidden_states=latent_model_input, + encoder_hidden_states=negative_prompt_embeds, + prior_token_id=prior_token_ids, + prior_token_drop=prior_token_drop_uncond, + timestep=timestep, + target_size=target_size, + crop_coords=crops_coords_top_left, + attention_kwargs=attention_kwargs, + return_dict=False, + kv_caches=kv_caches, + )[0].float() + + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond) + else: + noise_pred = noise_pred_cond + + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, self.scheduler.sigmas[i], callback_kwargs) + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + kv_caches.clear() + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.latent_channels, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, self.vae.config.latent_channels, 1, 1) + .to(latents.device, latents.dtype) + ) + latents = latents * latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False, generator=generator)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + else: + image = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return GlmImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/glm_image/pipeline_output.py b/diffusers/pipelines/glm_image/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..d4fd061335d4fc7cc4f0a03c25acccb78393b166 --- /dev/null +++ b/diffusers/pipelines/glm_image/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class GlmImagePipelineOutput(BaseOutput): + """ + Output class for CogView3 pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/helios/__init__.py b/diffusers/pipelines/helios/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ae08f5997279ff7417300e94f44787b34e4faa0b --- /dev/null +++ b/diffusers/pipelines/helios/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_helios"] = ["HeliosPipeline"] + _import_structure["pipeline_helios_pyramid"] = ["HeliosPyramidPipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_helios import HeliosPipeline + from .pipeline_helios_pyramid import HeliosPyramidPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/helios/pipeline_helios.py b/diffusers/pipelines/helios/pipeline_helios.py new file mode 100644 index 0000000000000000000000000000000000000000..90ac654bc77c4a94d9688e1cfdfbb654cbe6220b --- /dev/null +++ b/diffusers/pipelines/helios/pipeline_helios.py @@ -0,0 +1,951 @@ +# Copyright 2025 The Helios Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Any, Callable + +import numpy as np +import regex as re +import torch +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import HeliosLoraLoaderMixin +from ...models import AutoencoderKLWan, HeliosTransformer3DModel +from ...schedulers import HeliosScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HeliosPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers.utils import export_to_video + >>> from diffusers import AutoencoderKLWan, HeliosPipeline + + >>> # Available models: BestWishYsh/Helios-Base, BestWishYsh/Helios-Mid, BestWishYsh/Helios-Distilled + >>> model_id = "BestWishYsh/Helios-Base" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = HeliosPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=384, + ... width=640, + ... num_frames=132, + ... guidance_scale=5.0, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=24) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +class HeliosPipeline(DiffusionPipeline, HeliosLoraLoaderMixin): + r""" + Pipeline for text-to-video / image-to-video / video-to-video generation using Helios. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`HeliosTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`HeliosScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + _optional_components = ["transformer"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + vae: AutoencoderKLWan, + scheduler: HeliosScheduler, + transformer: HeliosTransformer3DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds, text_inputs.attention_mask.bool() + + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, _ = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, _ = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + image=None, + video=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if image is not None and video is not None: + raise ValueError("image and video cannot be provided simultaneously") + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 384, + width: int = 640, + num_frames: int = 33, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + def prepare_image_latents( + self, + image: torch.Tensor, + latents_mean: torch.Tensor, + latents_std: torch.Tensor, + num_latent_frames_per_chunk: int, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + fake_latents: torch.Tensor | None = None, + ) -> torch.Tensor: + device = device or self._execution_device + if latents is None: + image = image.unsqueeze(2).to(device=device, dtype=self.vae.dtype) + latents = self.vae.encode(image).latent_dist.sample(generator=generator) + latents = (latents - latents_mean) * latents_std + if fake_latents is None: + min_frames = (num_latent_frames_per_chunk - 1) * self.vae_scale_factor_temporal + 1 + fake_video = image.repeat(1, 1, min_frames, 1, 1).to(device=device, dtype=self.vae.dtype) + fake_latents_full = self.vae.encode(fake_video).latent_dist.sample(generator=generator) + fake_latents_full = (fake_latents_full - latents_mean) * latents_std + fake_latents = fake_latents_full[:, :, -1:, :, :] + return latents.to(device=device, dtype=dtype), fake_latents.to(device=device, dtype=dtype) + + def prepare_video_latents( + self, + video: torch.Tensor, + latents_mean: torch.Tensor, + latents_std: torch.Tensor, + num_latent_frames_per_chunk: int, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + device = device or self._execution_device + video = video.to(device=device, dtype=self.vae.dtype) + if latents is None: + num_frames = video.shape[2] + min_frames = (num_latent_frames_per_chunk - 1) * self.vae_scale_factor_temporal + 1 + num_chunks = num_frames // min_frames + if num_chunks == 0: + raise ValueError( + f"Video must have at least {min_frames} frames " + f"(got {num_frames} frames). " + f"Required: (num_latent_frames_per_chunk - 1) * {self.vae_scale_factor_temporal} + 1 = ({num_latent_frames_per_chunk} - 1) * {self.vae_scale_factor_temporal} + 1 = {min_frames}" + ) + total_valid_frames = num_chunks * min_frames + start_frame = num_frames - total_valid_frames + + first_frame = video[:, :, 0:1, :, :] + first_frame_latent = self.vae.encode(first_frame).latent_dist.sample(generator=generator) + first_frame_latent = (first_frame_latent - latents_mean) * latents_std + + latents_chunks = [] + for i in range(num_chunks): + chunk_start = start_frame + i * min_frames + chunk_end = chunk_start + min_frames + video_chunk = video[:, :, chunk_start:chunk_end, :, :] + chunk_latents = self.vae.encode(video_chunk).latent_dist.sample(generator=generator) + chunk_latents = (chunk_latents - latents_mean) * latents_std + latents_chunks.append(chunk_latents) + latents = torch.cat(latents_chunks, dim=2) + return first_frame_latent.to(device=device, dtype=dtype), latents.to(device=device, dtype=dtype) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 384, + width: int = 640, + num_frames: int = 132, + num_inference_steps: int = 50, + sigmas: list[float] = None, + guidance_scale: float = 5.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + # ------------ I2V ------------ + image: PipelineImageInput | None = None, + image_latents: torch.Tensor | None = None, + fake_image_latents: torch.Tensor | None = None, + add_noise_to_image_latents: bool = True, + image_noise_sigma_min: float = 0.111, + image_noise_sigma_max: float = 0.135, + # ------------ V2V ------------ + video: PipelineImageInput | None = None, + video_latents: torch.Tensor | None = None, + add_noise_to_video_latents: bool = True, + video_noise_sigma_min: float = 0.111, + video_noise_sigma_max: float = 0.135, + # ------------ Stage 1 ------------ + history_sizes: list = [16, 2, 1], + num_latent_frames_per_chunk: int = 9, + keep_first_frame: bool = True, + is_skip_first_chunk: bool = False, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to avoid during image generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`, defaults to `384`): + The height in pixels of the generated image. + width (`int`, defaults to `640`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `132`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process. If not defined, the scheduler's default schedule is + used. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. If not provided, they are generated from `negative_prompt`. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HeliosPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + image (`PipelineImageInput`, *optional*): + Input image used for image-to-video conditioning. + image_latents (`torch.Tensor`, *optional*): + Pre-encoded image latents to use instead of `image`. + fake_image_latents (`torch.Tensor`, *optional*): + Optional fake image latents used during conditioning. + add_noise_to_image_latents (`bool`, *optional*, defaults to `True`): + Whether to add noise to the image latents prior to denoising. + image_noise_sigma_min (`float`, *optional*, defaults to `0.111`): + Minimum sigma value for noise added to image latents. + image_noise_sigma_max (`float`, *optional*, defaults to `0.135`): + Maximum sigma value for noise added to image latents. + video (`PipelineImageInput`, *optional*): + Input video used for video-to-video conditioning. + video_latents (`torch.Tensor`, *optional*): + Pre-encoded video latents to use instead of `video`. + add_noise_to_video_latents (`bool`, *optional*, defaults to `True`): + Whether to add noise to the video latents prior to denoising. + video_noise_sigma_min (`float`, *optional*, defaults to `0.111`): + Minimum sigma value for noise added to video latents. + video_noise_sigma_max (`float`, *optional*, defaults to `0.135`): + Maximum sigma value for noise added to video latents. + history_sizes (`list`, *optional*, defaults to `[16, 2, 1]`): + History window sizes used for autoregressive chunked generation. + num_latent_frames_per_chunk (`int`, *optional*, defaults to `9`): + Number of latent frames produced per chunk during autoregressive generation. + keep_first_frame (`bool`, *optional*, defaults to `True`): + Whether to retain the first frame across chunks. + is_skip_first_chunk (`bool`, *optional*, defaults to `False`): + Whether to skip generation of the first chunk. + + Examples: + + Returns: + [`~HeliosPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HeliosPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + history_sizes = sorted(history_sizes, reverse=True) # From big to small + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + image, + video, + ) + + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + vae_dtype = self.vae.dtype + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(device, self.vae.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + device, self.vae.dtype + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare image or video + if image is not None: + image = self.video_processor.preprocess(image, height=height, width=width) + image_latents, fake_image_latents = self.prepare_image_latents( + image, + latents_mean=latents_mean, + latents_std=latents_std, + num_latent_frames_per_chunk=num_latent_frames_per_chunk, + dtype=torch.float32, + device=device, + generator=generator, + latents=image_latents, + fake_latents=fake_image_latents, + ) + + if image_latents is not None and add_noise_to_image_latents: + image_noise_sigma = ( + torch.rand(1, device=device, generator=generator) * (image_noise_sigma_max - image_noise_sigma_min) + + image_noise_sigma_min + ) + image_latents = ( + image_noise_sigma * randn_tensor(image_latents.shape, generator=generator, device=device) + + (1 - image_noise_sigma) * image_latents + ) + fake_image_noise_sigma = ( + torch.rand(1, device=device, generator=generator) * (video_noise_sigma_max - video_noise_sigma_min) + + video_noise_sigma_min + ) + fake_image_latents = ( + fake_image_noise_sigma * randn_tensor(fake_image_latents.shape, generator=generator, device=device) + + (1 - fake_image_noise_sigma) * fake_image_latents + ) + + if video is not None: + video = self.video_processor.preprocess_video(video, height=height, width=width) + image_latents, video_latents = self.prepare_video_latents( + video, + latents_mean=latents_mean, + latents_std=latents_std, + num_latent_frames_per_chunk=num_latent_frames_per_chunk, + dtype=torch.float32, + device=device, + generator=generator, + latents=video_latents, + ) + + if video_latents is not None and add_noise_to_video_latents: + image_noise_sigma = ( + torch.rand(1, device=device, generator=generator) * (image_noise_sigma_max - image_noise_sigma_min) + + image_noise_sigma_min + ) + image_latents = ( + image_noise_sigma * randn_tensor(image_latents.shape, generator=generator, device=device) + + (1 - image_noise_sigma) * image_latents + ) + + noisy_latents_chunks = [] + num_latent_chunks = video_latents.shape[2] // num_latent_frames_per_chunk + for i in range(num_latent_chunks): + chunk_start = i * num_latent_frames_per_chunk + chunk_end = chunk_start + num_latent_frames_per_chunk + latent_chunk = video_latents[:, :, chunk_start:chunk_end, :, :] + + chunk_frames = latent_chunk.shape[2] + frame_sigmas = ( + torch.rand(chunk_frames, device=device, generator=generator) + * (video_noise_sigma_max - video_noise_sigma_min) + + video_noise_sigma_min + ) + frame_sigmas = frame_sigmas.view(1, 1, chunk_frames, 1, 1) + + noisy_chunk = ( + frame_sigmas * randn_tensor(latent_chunk.shape, generator=generator, device=device) + + (1 - frame_sigmas) * latent_chunk + ) + noisy_latents_chunks.append(noisy_chunk) + video_latents = torch.cat(noisy_latents_chunks, dim=2) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + window_num_frames = (num_latent_frames_per_chunk - 1) * self.vae_scale_factor_temporal + 1 + num_latent_chunk = max(1, (num_frames + window_num_frames - 1) // window_num_frames) + num_history_latent_frames = sum(history_sizes) + history_video = None + total_generated_latent_frames = 0 + + if not keep_first_frame: + history_sizes[-1] = history_sizes[-1] + 1 + history_latents = torch.zeros( + batch_size, + num_channels_latents, + num_history_latent_frames, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + device=device, + dtype=torch.float32, + ) + if fake_image_latents is not None: + history_latents = torch.cat([history_latents[:, :, :-1, :, :], fake_image_latents], dim=2) + total_generated_latent_frames += 1 + if video_latents is not None: + history_frames = history_latents.shape[2] + video_frames = video_latents.shape[2] + if video_frames < history_frames: + keep_frames = history_frames - video_frames + history_latents = torch.cat([history_latents[:, :, :keep_frames, :, :], video_latents], dim=2) + else: + history_latents = video_latents + total_generated_latent_frames += video_latents.shape[2] + + if keep_first_frame: + indices = torch.arange(0, sum([1, *history_sizes, num_latent_frames_per_chunk])) + ( + indices_prefix, + indices_latents_history_long, + indices_latents_history_mid, + indices_latents_history_1x, + indices_hidden_states, + ) = indices.split([1, *history_sizes, num_latent_frames_per_chunk], dim=0) + indices_latents_history_short = torch.cat([indices_prefix, indices_latents_history_1x], dim=0) + else: + indices = torch.arange(0, sum([*history_sizes, num_latent_frames_per_chunk])) + ( + indices_latents_history_long, + indices_latents_history_mid, + indices_latents_history_short, + indices_hidden_states, + ) = indices.split([*history_sizes, num_latent_frames_per_chunk], dim=0) + indices_hidden_states = indices_hidden_states.unsqueeze(0) + indices_latents_history_short = indices_latents_history_short.unsqueeze(0) + indices_latents_history_mid = indices_latents_history_mid.unsqueeze(0) + indices_latents_history_long = indices_latents_history_long.unsqueeze(0) + + # 6. Denoising loop + patch_size = self.transformer.config.patch_size + image_seq_len = ( + num_latent_frames_per_chunk + * (height // self.vae_scale_factor_spatial) + * (width // self.vae_scale_factor_spatial) + // (patch_size[0] * patch_size[1] * patch_size[2]) + ) + sigmas = np.linspace(0.999, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + + for k in range(num_latent_chunk): + is_first_chunk = k == 0 + is_second_chunk = k == 1 + if keep_first_frame: + latents_history_long, latents_history_mid, latents_history_1x = history_latents[ + :, :, -num_history_latent_frames: + ].split(history_sizes, dim=2) + if image_latents is None and is_first_chunk: + latents_prefix = torch.zeros( + ( + batch_size, + num_channels_latents, + 1, + latents_history_1x.shape[-2], + latents_history_1x.shape[-1], + ), + device=device, + dtype=latents_history_1x.dtype, + ) + else: + latents_prefix = image_latents + latents_history_short = torch.cat([latents_prefix, latents_history_1x], dim=2) + else: + latents_history_long, latents_history_mid, latents_history_short = history_latents[ + :, :, -num_history_latent_frames: + ].split(history_sizes, dim=2) + + latents = self.prepare_latents( + batch_size, + num_channels_latents, + height, + width, + window_num_frames, + dtype=torch.float32, + device=device, + generator=generator, + latents=None, + ) + + self.scheduler.set_timesteps(num_inference_steps, device=device, sigmas=sigmas, mu=mu) + timesteps = self.scheduler.timesteps + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + timestep = t.expand(latents.shape[0]) + + latent_model_input = latents.to(transformer_dtype) + latents_history_short = latents_history_short.to(transformer_dtype) + latents_history_mid = latents_history_mid.to(transformer_dtype) + latents_history_long = latents_history_long.to(transformer_dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + indices_hidden_states=indices_hidden_states, + indices_latents_history_short=indices_latents_history_short, + indices_latents_history_mid=indices_latents_history_mid, + indices_latents_history_long=indices_latents_history_long, + latents_history_short=latents_history_short, + latents_history_mid=latents_history_mid, + latents_history_long=latents_history_long, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + indices_hidden_states=indices_hidden_states, + indices_latents_history_short=indices_latents_history_short, + indices_latents_history_mid=indices_latents_history_mid, + indices_latents_history_long=indices_latents_history_long, + latents_history_short=latents_history_short, + latents_history_mid=latents_history_mid, + latents_history_long=latents_history_long, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + return_dict=False, + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if keep_first_frame and ( + (is_first_chunk and image_latents is None) or (is_skip_first_chunk and is_second_chunk) + ): + image_latents = latents[:, :, 0:1, :, :] + + total_generated_latent_frames += latents.shape[2] + history_latents = torch.cat([history_latents, latents], dim=2) + real_history_latents = history_latents[:, :, -total_generated_latent_frames:] + current_latents = ( + real_history_latents[:, :, -num_latent_frames_per_chunk:].to(vae_dtype) / latents_std + + latents_mean + ) + current_video = self.vae.decode(current_latents, return_dict=False)[0] + + if history_video is None: + history_video = current_video + else: + history_video = torch.cat([history_video, current_video], dim=2) + + self._current_timestep = None + + if output_type != "latent": + generated_frames = history_video.size(2) + generated_frames = ( + generated_frames - 1 + ) // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + history_video = history_video[:, :, :generated_frames] + video = self.video_processor.postprocess_video(history_video, output_type=output_type) + else: + video = real_history_latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HeliosPipelineOutput(frames=video) diff --git a/diffusers/pipelines/helios/pipeline_helios_pyramid.py b/diffusers/pipelines/helios/pipeline_helios_pyramid.py new file mode 100644 index 0000000000000000000000000000000000000000..c187e436a857268a7df4a83bb80427beb9170c9b --- /dev/null +++ b/diffusers/pipelines/helios/pipeline_helios_pyramid.py @@ -0,0 +1,1131 @@ +# Copyright 2025 The Helios Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import math +from typing import Any, Callable + +import regex as re +import torch +import torch.nn.functional as F +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import HeliosLoraLoaderMixin +from ...models import AutoencoderKLWan, HeliosTransformer3DModel +from ...schedulers import HeliosDMDScheduler, HeliosScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HeliosPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers.utils import export_to_video + >>> from diffusers import AutoencoderKLWan, HeliosPyramidPipeline + + >>> # Available models: BestWishYsh/Helios-Base, BestWishYsh/Helios-Mid, BestWishYsh/Helios-Distilled + >>> model_id = "BestWishYsh/Helios-Base" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = HeliosPyramidPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=384, + ... width=640, + ... num_frames=132, + ... guidance_scale=5.0, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=24) + ``` +""" + + +def optimized_scale(positive_flat, negative_flat): + positive_flat = positive_flat.float() + negative_flat = negative_flat.float() + # Calculate dot production + dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True) + # Squared norm of uncondition + squared_norm = torch.sum(negative_flat**2, dim=1, keepdim=True) + 1e-8 + # st_star = v_cond^T * v_uncond / ||v_uncond||^2 + st_star = dot_product / squared_norm + return st_star + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +class HeliosPyramidPipeline(DiffusionPipeline, HeliosLoraLoaderMixin): + r""" + Pipeline for text-to-video / image-to-video / video-to-video generation using Helios. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`HeliosTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`HeliosScheduler`, `HeliosDMDScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + _optional_components = ["transformer"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + vae: AutoencoderKLWan, + scheduler: HeliosScheduler | HeliosDMDScheduler, + transformer: HeliosTransformer3DModel, + is_cfg_zero_star: bool = False, + is_distilled: bool = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.register_to_config(is_cfg_zero_star=is_cfg_zero_star) + self.register_to_config(is_distilled=is_distilled) + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.helios.pipeline_helios.HeliosPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds, text_inputs.attention_mask.bool() + + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, _ = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, _ = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + image=None, + video=None, + guidance_scale=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if image is not None and video is not None: + raise ValueError("image and video cannot be provided simultaneously") + + if guidance_scale > 1.0 and self.config.is_distilled: + logger.warning(f"Guidance scale {guidance_scale} is ignored for step-wise distilled models.") + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 384, + width: int = 640, + num_frames: int = 33, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + def prepare_image_latents( + self, + image: torch.Tensor, + latents_mean: torch.Tensor, + latents_std: torch.Tensor, + num_latent_frames_per_chunk: int, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + fake_latents: torch.Tensor | None = None, + ) -> torch.Tensor: + device = device or self._execution_device + if latents is None: + image = image.unsqueeze(2).to(device=device, dtype=self.vae.dtype) + latents = self.vae.encode(image).latent_dist.sample(generator=generator) + latents = (latents - latents_mean) * latents_std + if fake_latents is None: + min_frames = (num_latent_frames_per_chunk - 1) * self.vae_scale_factor_temporal + 1 + fake_video = image.repeat(1, 1, min_frames, 1, 1).to(device=device, dtype=self.vae.dtype) + fake_latents_full = self.vae.encode(fake_video).latent_dist.sample(generator=generator) + fake_latents_full = (fake_latents_full - latents_mean) * latents_std + fake_latents = fake_latents_full[:, :, -1:, :, :] + return latents.to(device=device, dtype=dtype), fake_latents.to(device=device, dtype=dtype) + + def prepare_video_latents( + self, + video: torch.Tensor, + latents_mean: torch.Tensor, + latents_std: torch.Tensor, + num_latent_frames_per_chunk: int, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + device = device or self._execution_device + video = video.to(device=device, dtype=self.vae.dtype) + if latents is None: + num_frames = video.shape[2] + min_frames = (num_latent_frames_per_chunk - 1) * self.vae_scale_factor_temporal + 1 + num_chunks = num_frames // min_frames + if num_chunks == 0: + raise ValueError( + f"Video must have at least {min_frames} frames " + f"(got {num_frames} frames). " + f"Required: (num_latent_frames_per_chunk - 1) * {self.vae_scale_factor_temporal} + 1 = ({num_latent_frames_per_chunk} - 1) * {self.vae_scale_factor_temporal} + 1 = {min_frames}" + ) + total_valid_frames = num_chunks * min_frames + start_frame = num_frames - total_valid_frames + + first_frame = video[:, :, 0:1, :, :] + first_frame_latent = self.vae.encode(first_frame).latent_dist.sample(generator=generator) + first_frame_latent = (first_frame_latent - latents_mean) * latents_std + + latents_chunks = [] + for i in range(num_chunks): + chunk_start = start_frame + i * min_frames + chunk_end = chunk_start + min_frames + video_chunk = video[:, :, chunk_start:chunk_end, :, :] + chunk_latents = self.vae.encode(video_chunk).latent_dist.sample(generator=generator) + chunk_latents = (chunk_latents - latents_mean) * latents_std + latents_chunks.append(chunk_latents) + latents = torch.cat(latents_chunks, dim=2) + return first_frame_latent.to(device=device, dtype=dtype), latents.to(device=device, dtype=dtype) + + def sample_block_noise( + self, + batch_size, + channel, + num_frames, + height, + width, + patch_size: tuple[int, ...] = (1, 2, 2), + device: torch.device | None = None, + generator: torch.Generator | None = None, + ): + # NOTE: A generator must be provided to ensure correct and reproducible results. + # Creating a default generator here is a fallback only — without a fixed seed, + # the output will be non-deterministic and may produce incorrect results in CP context. + if generator is None: + generator = torch.Generator(device=device) + elif isinstance(generator, list): + generator = generator[0] + + gamma = self.scheduler.config.gamma + _, ph, pw = patch_size + block_size = ph * pw + + cov = ( + torch.eye(block_size, device=device) * (1 + gamma) + - torch.ones(block_size, block_size, device=device) * gamma + ) + cov += torch.eye(block_size, device=device) * 1e-8 + cov = cov.float() # Upcast to fp32 for numerical stability — cholesky is unreliable in fp16/bf16. + + L = torch.linalg.cholesky(cov) + block_number = batch_size * channel * num_frames * (height // ph) * (width // pw) + z = torch.randn(block_number, block_size, generator=generator, device=generator.device) + z = z.to(device=device) + noise = z @ L.T + + noise = noise.view(batch_size, channel, num_frames, height // ph, width // pw, ph, pw) + noise = noise.permute(0, 1, 2, 3, 5, 4, 6).reshape(batch_size, channel, num_frames, height, width) + + return noise + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 384, + width: int = 640, + num_frames: int = 132, + sigmas: list[float] = None, + guidance_scale: float = 5.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + # ------------ I2V ------------ + image: PipelineImageInput | None = None, + image_latents: torch.Tensor | None = None, + fake_image_latents: torch.Tensor | None = None, + add_noise_to_image_latents: bool = True, + image_noise_sigma_min: float = 0.111, + image_noise_sigma_max: float = 0.135, + # ------------ V2V ------------ + video: PipelineImageInput | None = None, + video_latents: torch.Tensor | None = None, + add_noise_to_video_latents: bool = True, + video_noise_sigma_min: float = 0.111, + video_noise_sigma_max: float = 0.135, + # ------------ Stage 1 ------------ + history_sizes: list = [16, 2, 1], + num_latent_frames_per_chunk: int = 9, + keep_first_frame: bool = True, + is_skip_first_chunk: bool = False, + # ------------ Stage 2 ------------ + pyramid_num_inference_steps_list: list = [10, 10, 10], + # ------------ CFG Zero ------------ + use_zero_init: bool | None = True, + zero_steps: int | None = 1, + # ------------ DMD ------------ + is_amplify_first_chunk: bool = False, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to avoid during image generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`, defaults to `384`): + The height in pixels of the generated image. + width (`int`, defaults to `640`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `132`): + The number of frames in the generated video. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process. If not defined, the scheduler's default schedule is + used. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. If not provided, they are generated from `negative_prompt`. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HeliosPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + image (`PipelineImageInput`, *optional*): + Input image used for image-to-video conditioning. + image_latents (`torch.Tensor`, *optional*): + Pre-encoded image latents to use instead of `image`. + fake_image_latents (`torch.Tensor`, *optional*): + Optional fake image latents used during conditioning. + add_noise_to_image_latents (`bool`, *optional*, defaults to `True`): + Whether to add noise to the image latents prior to denoising. + image_noise_sigma_min (`float`, *optional*, defaults to `0.111`): + Minimum sigma value for noise added to image latents. + image_noise_sigma_max (`float`, *optional*, defaults to `0.135`): + Maximum sigma value for noise added to image latents. + video (`PipelineImageInput`, *optional*): + Input video used for video-to-video conditioning. + video_latents (`torch.Tensor`, *optional*): + Pre-encoded video latents to use instead of `video`. + add_noise_to_video_latents (`bool`, *optional*, defaults to `True`): + Whether to add noise to the video latents prior to denoising. + video_noise_sigma_min (`float`, *optional*, defaults to `0.111`): + Minimum sigma value for noise added to video latents. + video_noise_sigma_max (`float`, *optional*, defaults to `0.135`): + Maximum sigma value for noise added to video latents. + history_sizes (`list`, *optional*, defaults to `[16, 2, 1]`): + History window sizes used for autoregressive chunked generation. + num_latent_frames_per_chunk (`int`, *optional*, defaults to `9`): + Number of latent frames produced per chunk during autoregressive generation. + keep_first_frame (`bool`, *optional*, defaults to `True`): + Whether to retain the first frame across chunks. + is_skip_first_chunk (`bool`, *optional*, defaults to `False`): + Whether to skip generation of the first chunk. + pyramid_num_inference_steps_list (`list`, *optional*, defaults to `[10, 10, 10]`): + Number of inference steps for each pyramid stage during Stage 2 generation. + use_zero_init (`bool`, *optional*, defaults to `True`): + Whether to apply CFG zero-init at the start of denoising. + zero_steps (`int`, *optional*, defaults to `1`): + Number of initial steps that use CFG zero-init. + is_amplify_first_chunk (`bool`, *optional*, defaults to `False`): + Whether to amplify guidance on the first chunk (DMD-related). + + Examples: + + Returns: + [`~HeliosPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HeliosPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + history_sizes = sorted(history_sizes, reverse=True) # From big to small + pyramid_num_stages = len(pyramid_num_inference_steps_list) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + image, + video, + guidance_scale, + ) + + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + vae_dtype = self.vae.dtype + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(device, self.vae.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + device, self.vae.dtype + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare image or video + if image is not None: + image = self.video_processor.preprocess(image, height=height, width=width) + image_latents, fake_image_latents = self.prepare_image_latents( + image, + latents_mean=latents_mean, + latents_std=latents_std, + num_latent_frames_per_chunk=num_latent_frames_per_chunk, + dtype=torch.float32, + device=device, + generator=generator, + latents=image_latents, + fake_latents=fake_image_latents, + ) + + if image_latents is not None and add_noise_to_image_latents: + image_noise_sigma = ( + torch.rand(1, device=device, generator=generator) * (image_noise_sigma_max - image_noise_sigma_min) + + image_noise_sigma_min + ) + image_latents = ( + image_noise_sigma * randn_tensor(image_latents.shape, generator=generator, device=device) + + (1 - image_noise_sigma) * image_latents + ) + fake_image_noise_sigma = ( + torch.rand(1, device=device, generator=generator) * (video_noise_sigma_max - video_noise_sigma_min) + + video_noise_sigma_min + ) + fake_image_latents = ( + fake_image_noise_sigma * randn_tensor(fake_image_latents.shape, generator=generator, device=device) + + (1 - fake_image_noise_sigma) * fake_image_latents + ) + + if video is not None: + video = self.video_processor.preprocess_video(video, height=height, width=width) + image_latents, video_latents = self.prepare_video_latents( + video, + latents_mean=latents_mean, + latents_std=latents_std, + num_latent_frames_per_chunk=num_latent_frames_per_chunk, + dtype=torch.float32, + device=device, + generator=generator, + latents=video_latents, + ) + + if video_latents is not None and add_noise_to_video_latents: + image_noise_sigma = ( + torch.rand(1, device=device, generator=generator) * (image_noise_sigma_max - image_noise_sigma_min) + + image_noise_sigma_min + ) + image_latents = ( + image_noise_sigma * randn_tensor(image_latents.shape, generator=generator, device=device) + + (1 - image_noise_sigma) * image_latents + ) + + noisy_latents_chunks = [] + num_latent_chunks = video_latents.shape[2] // num_latent_frames_per_chunk + for i in range(num_latent_chunks): + chunk_start = i * num_latent_frames_per_chunk + chunk_end = chunk_start + num_latent_frames_per_chunk + latent_chunk = video_latents[:, :, chunk_start:chunk_end, :, :] + + chunk_frames = latent_chunk.shape[2] + frame_sigmas = ( + torch.rand(chunk_frames, device=device, generator=generator) + * (video_noise_sigma_max - video_noise_sigma_min) + + video_noise_sigma_min + ) + frame_sigmas = frame_sigmas.view(1, 1, chunk_frames, 1, 1) + + noisy_chunk = ( + frame_sigmas * randn_tensor(latent_chunk.shape, generator=generator, device=device) + + (1 - frame_sigmas) * latent_chunk + ) + noisy_latents_chunks.append(noisy_chunk) + video_latents = torch.cat(noisy_latents_chunks, dim=2) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + window_num_frames = (num_latent_frames_per_chunk - 1) * self.vae_scale_factor_temporal + 1 + num_latent_chunk = max(1, (num_frames + window_num_frames - 1) // window_num_frames) + num_history_latent_frames = sum(history_sizes) + history_video = None + total_generated_latent_frames = 0 + + if not keep_first_frame: + history_sizes[-1] = history_sizes[-1] + 1 + history_latents = torch.zeros( + batch_size, + num_channels_latents, + num_history_latent_frames, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + device=device, + dtype=torch.float32, + ) + if fake_image_latents is not None: + history_latents = torch.cat([history_latents[:, :, :-1, :, :], fake_image_latents], dim=2) + total_generated_latent_frames += 1 + if video_latents is not None: + history_frames = history_latents.shape[2] + video_frames = video_latents.shape[2] + if video_frames < history_frames: + keep_frames = history_frames - video_frames + history_latents = torch.cat([history_latents[:, :, :keep_frames, :, :], video_latents], dim=2) + else: + history_latents = video_latents + total_generated_latent_frames += video_latents.shape[2] + + if keep_first_frame: + indices = torch.arange(0, sum([1, *history_sizes, num_latent_frames_per_chunk])) + ( + indices_prefix, + indices_latents_history_long, + indices_latents_history_mid, + indices_latents_history_1x, + indices_hidden_states, + ) = indices.split([1, *history_sizes, num_latent_frames_per_chunk], dim=0) + indices_latents_history_short = torch.cat([indices_prefix, indices_latents_history_1x], dim=0) + else: + indices = torch.arange(0, sum([*history_sizes, num_latent_frames_per_chunk])) + ( + indices_latents_history_long, + indices_latents_history_mid, + indices_latents_history_short, + indices_hidden_states, + ) = indices.split([*history_sizes, num_latent_frames_per_chunk], dim=0) + indices_hidden_states = indices_hidden_states.unsqueeze(0) + indices_latents_history_short = indices_latents_history_short.unsqueeze(0) + indices_latents_history_mid = indices_latents_history_mid.unsqueeze(0) + indices_latents_history_long = indices_latents_history_long.unsqueeze(0) + + # 6. Denoising loop + for k in range(num_latent_chunk): + is_first_chunk = k == 0 + is_second_chunk = k == 1 + if keep_first_frame: + latents_history_long, latents_history_mid, latents_history_1x = history_latents[ + :, :, -num_history_latent_frames: + ].split(history_sizes, dim=2) + if image_latents is None and is_first_chunk: + latents_prefix = torch.zeros( + ( + batch_size, + num_channels_latents, + 1, + latents_history_1x.shape[-2], + latents_history_1x.shape[-1], + ), + device=device, + dtype=latents_history_1x.dtype, + ) + else: + latents_prefix = image_latents + latents_history_short = torch.cat([latents_prefix, latents_history_1x], dim=2) + else: + latents_history_long, latents_history_mid, latents_history_short = history_latents[ + :, :, -num_history_latent_frames: + ].split(history_sizes, dim=2) + + latents = self.prepare_latents( + batch_size, + num_channels_latents, + height, + width, + window_num_frames, + dtype=torch.float32, + device=device, + generator=generator, + latents=None, + ) + + num_inference_steps = ( + sum(pyramid_num_inference_steps_list) * 2 + if is_amplify_first_chunk and self.config.is_distilled and is_first_chunk + else sum(pyramid_num_inference_steps_list) + ) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + _, _, _, pyramid_height, pyramid_width = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape( + batch_size * num_latent_frames_per_chunk, num_channels_latents, pyramid_height, pyramid_width + ) + for _ in range(pyramid_num_stages - 1): + pyramid_height //= 2 + pyramid_width //= 2 + latents = ( + F.interpolate( + latents, + size=(pyramid_height, pyramid_width), + mode="bilinear", + ) + * 2 + ) + latents = latents.reshape( + batch_size, num_latent_frames_per_chunk, num_channels_latents, pyramid_height, pyramid_width + ).permute(0, 2, 1, 3, 4) + + start_point_list = None + if self.config.is_distilled: + start_point_list = [latents] + + for stage_idx in range(pyramid_num_stages): + patch_size = self.transformer.config.patch_size + image_seq_len = (latents.shape[-1] * latents.shape[-2] * latents.shape[-3]) // ( + patch_size[0] * patch_size[1] * patch_size[2] + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + self.scheduler.set_timesteps( + pyramid_num_inference_steps_list[stage_idx], + stage_idx, + device=device, + mu=mu, + is_amplify_first_chunk=is_amplify_first_chunk and is_first_chunk, + ) + timesteps = self.scheduler.timesteps + num_warmup_steps = 0 + self._num_timesteps = len(timesteps) + + if stage_idx > 0: + pyramid_height *= 2 + pyramid_width *= 2 + num_frames = latents.shape[2] + latents = latents.permute(0, 2, 1, 3, 4).reshape( + batch_size * num_latent_frames_per_chunk, + num_channels_latents, + pyramid_height // 2, + pyramid_width // 2, + ) + latents = F.interpolate(latents, size=(pyramid_height, pyramid_width), mode="nearest") + latents = latents.reshape( + batch_size, + num_latent_frames_per_chunk, + num_channels_latents, + pyramid_height, + pyramid_width, + ).permute(0, 2, 1, 3, 4) + # Fix the stage + ori_sigma = 1 - self.scheduler.ori_start_sigmas[stage_idx] # the original coeff of signal + gamma = self.scheduler.config.gamma + alpha = 1 / (math.sqrt(1 + (1 / gamma)) * (1 - ori_sigma) + ori_sigma) + beta = alpha * (1 - ori_sigma) / math.sqrt(gamma) + + batch_size, channel, num_frames, pyramid_height, pyramid_width = latents.shape + noise = self.sample_block_noise( + batch_size, + channel, + num_frames, + pyramid_height, + pyramid_width, + patch_size, + device, + generator, + ) + noise = noise.to(device=device, dtype=transformer_dtype) + latents = alpha * latents + beta * noise # To fix the block artifact + + if self.config.is_distilled: + start_point_list.append(latents) + + for i, t in enumerate(timesteps): + timestep = t.expand(latents.shape[0]).to(torch.int64) + + latent_model_input = latents.to(transformer_dtype) + latents_history_short = latents_history_short.to(transformer_dtype) + latents_history_mid = latents_history_mid.to(transformer_dtype) + latents_history_long = latents_history_long.to(transformer_dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + indices_hidden_states=indices_hidden_states, + indices_latents_history_short=indices_latents_history_short, + indices_latents_history_mid=indices_latents_history_mid, + indices_latents_history_long=indices_latents_history_long, + latents_history_short=latents_history_short, + latents_history_mid=latents_history_mid, + latents_history_long=latents_history_long, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + indices_hidden_states=indices_hidden_states, + indices_latents_history_short=indices_latents_history_short, + indices_latents_history_mid=indices_latents_history_mid, + indices_latents_history_long=indices_latents_history_long, + latents_history_short=latents_history_short, + latents_history_mid=latents_history_mid, + latents_history_long=latents_history_long, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.config.is_cfg_zero_star: + noise_pred_text = noise_pred + positive_flat = noise_pred_text.view(batch_size, -1) + negative_flat = noise_uncond.view(batch_size, -1) + + alpha = optimized_scale(positive_flat, negative_flat) + alpha = alpha.view(batch_size, *([1] * (len(noise_pred_text.shape) - 1))) + alpha = alpha.to(noise_pred_text.dtype) + + if (stage_idx == 0 and i <= zero_steps) and use_zero_init: + noise_pred = noise_pred_text * 0.0 + else: + noise_pred = noise_uncond * alpha + guidance_scale * ( + noise_pred_text - noise_uncond * alpha + ) + else: + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + extra_kwargs = ( + { + "cur_sampling_step": i, + "dmd_noisy_tensor": start_point_list[stage_idx] + if start_point_list is not None + else None, + "dmd_sigmas": self.scheduler.sigmas, + "dmd_timesteps": self.scheduler.timesteps, + "all_timesteps": timesteps, + } + if self.config.is_distilled + else {} + ) + + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + return_dict=False, + **extra_kwargs, + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop( + "negative_prompt_embeds", negative_prompt_embeds + ) + + if i == len(timesteps) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0 + ): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if keep_first_frame and ( + (is_first_chunk and image_latents is None) or (is_skip_first_chunk and is_second_chunk) + ): + image_latents = latents[:, :, 0:1, :, :] + + total_generated_latent_frames += latents.shape[2] + history_latents = torch.cat([history_latents, latents], dim=2) + real_history_latents = history_latents[:, :, -total_generated_latent_frames:] + current_latents = ( + real_history_latents[:, :, -num_latent_frames_per_chunk:].to(vae_dtype) / latents_std + + latents_mean + ) + current_video = self.vae.decode(current_latents, return_dict=False)[0] + + if history_video is None: + history_video = current_video + else: + history_video = torch.cat([history_video, current_video], dim=2) + + self._current_timestep = None + + if output_type != "latent": + generated_frames = history_video.size(2) + generated_frames = ( + generated_frames - 1 + ) // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + history_video = history_video[:, :, :generated_frames] + video = self.video_processor.postprocess_video(history_video, output_type=output_type) + else: + video = real_history_latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HeliosPipelineOutput(frames=video) diff --git a/diffusers/pipelines/helios/pipeline_output.py b/diffusers/pipelines/helios/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..08546289ef4c0739916c3106b8d9e6a93120d64a --- /dev/null +++ b/diffusers/pipelines/helios/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class HeliosPipelineOutput(BaseOutput): + r""" + Output class for Helios pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/diffusers/pipelines/hidream_image/__init__.py b/diffusers/pipelines/hidream_image/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..498df900e68b70920e5fc49764eb6bdcd2dc3354 --- /dev/null +++ b/diffusers/pipelines/hidream_image/__init__.py @@ -0,0 +1,47 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["HiDreamImagePipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_hidream_image"] = ["HiDreamImagePipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_hidream_image import HiDreamImagePipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/hidream_image/pipeline_hidream_image.py b/diffusers/pipelines/hidream_image/pipeline_hidream_image.py new file mode 100644 index 0000000000000000000000000000000000000000..d723c7941ae96dd2fe2c1200b8eeb3f7c84d3556 --- /dev/null +++ b/diffusers/pipelines/hidream_image/pipeline_hidream_image.py @@ -0,0 +1,1005 @@ +# Copyright 2025 HiDream-ai Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + LlamaForCausalLM, + PreTrainedTokenizerFast, + T5EncoderModel, + T5Tokenizer, +) + +from ...image_processor import VaeImageProcessor +from ...loaders import HiDreamImageLoraLoaderMixin +from ...models import AutoencoderKL, HiDreamImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler, UniPCMultistepScheduler +from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HiDreamImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from transformers import AutoTokenizer, LlamaForCausalLM + >>> from diffusers import HiDreamImagePipeline + + >>> tokenizer_4 = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct") + >>> text_encoder_4 = LlamaForCausalLM.from_pretrained( + ... "meta-llama/Meta-Llama-3.1-8B-Instruct", + ... output_hidden_states=True, + ... output_attentions=True, + ... torch_dtype=torch.bfloat16, + ... ) + + >>> pipe = HiDreamImagePipeline.from_pretrained( + ... "HiDream-ai/HiDream-I1-Full", + ... tokenizer_4=tokenizer_4, + ... text_encoder_4=text_encoder_4, + ... torch_dtype=torch.bfloat16, + ... ) + >>> pipe.enable_model_cpu_offload() + + >>> image = pipe( + ... 'A cat holding a sign that says "Hi-Dreams.ai".', + ... height=1024, + ... width=1024, + ... guidance_scale=5.0, + ... num_inference_steps=50, + ... generator=torch.Generator("cuda").manual_seed(0), + ... ).images[0] + >>> image.save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class HiDreamImagePipeline(DiffusionPipeline, HiDreamImageLoraLoaderMixin): + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->text_encoder_4->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds_t5", "prompt_embeds_llama3", "pooled_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5Tokenizer, + text_encoder_4: LlamaForCausalLM, + tokenizer_4: PreTrainedTokenizerFast, + transformer: HiDreamImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + text_encoder_4=text_encoder_4, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + tokenizer_4=tokenizer_4, + scheduler=scheduler, + transformer=transformer, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + # HiDreamImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.default_sample_size = 128 + if getattr(self, "tokenizer_4", None) is not None: + self.tokenizer_4.pad_token = self.tokenizer_4.eos_token + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder_3.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=min(max_sequence_length, self.tokenizer_3.model_max_length), + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode( + untruncated_ids[:, min(max_sequence_length, self.tokenizer_3.model_max_length) - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {min(max_sequence_length, self.tokenizer_3.model_max_length)} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device), attention_mask=attention_mask.to(device))[0] + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + return prompt_embeds + + def _get_clip_prompt_embeds( + self, + tokenizer, + text_encoder, + prompt: str | list[str], + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=min(max_sequence_length, 218), + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, 218 - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {218} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds[0] + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + return prompt_embeds + + def _get_llama3_prompt_embeds( + self, + prompt: str | list[str] = None, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder_4.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + text_inputs = self.tokenizer_4( + prompt, + padding="max_length", + max_length=min(max_sequence_length, self.tokenizer_4.model_max_length), + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + untruncated_ids = self.tokenizer_4(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_4.batch_decode( + untruncated_ids[:, min(max_sequence_length, self.tokenizer_4.model_max_length) - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {min(max_sequence_length, self.tokenizer_4.model_max_length)} tokens: {removed_text}" + ) + + outputs = self.text_encoder_4( + text_input_ids.to(device), + attention_mask=attention_mask.to(device), + output_hidden_states=True, + output_attentions=True, + ) + + prompt_embeds = outputs.hidden_states[1:] + prompt_embeds = torch.stack(prompt_embeds, dim=0) + return prompt_embeds + + def encode_prompt( + self, + prompt: str | list[str] | None = None, + prompt_2: str | list[str] | None = None, + prompt_3: str | list[str] | None = None, + prompt_4: str | list[str] | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + negative_prompt_4: str | list[str] | None = None, + prompt_embeds_t5: list[torch.FloatTensor] | None = None, + prompt_embeds_llama3: list[torch.FloatTensor] | None = None, + negative_prompt_embeds_t5: list[torch.FloatTensor] | None = None, + negative_prompt_embeds_llama3: list[torch.FloatTensor] | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 128, + lora_scale: float | None = None, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = pooled_prompt_embeds.shape[0] + + device = device or self._execution_device + + if pooled_prompt_embeds is None: + pooled_prompt_embeds_1 = self._get_clip_prompt_embeds( + self.tokenizer, self.text_encoder, prompt, max_sequence_length, device, dtype + ) + + if do_classifier_free_guidance and negative_pooled_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if len(negative_prompt) > 1 and len(negative_prompt) != batch_size: + raise ValueError(f"negative_prompt must be of length 1 or {batch_size}") + + negative_pooled_prompt_embeds_1 = self._get_clip_prompt_embeds( + self.tokenizer, self.text_encoder, negative_prompt, max_sequence_length, device, dtype + ) + + if negative_pooled_prompt_embeds_1.shape[0] == 1 and batch_size > 1: + negative_pooled_prompt_embeds_1 = negative_pooled_prompt_embeds_1.repeat(batch_size, 1) + + if pooled_prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + if len(prompt_2) > 1 and len(prompt_2) != batch_size: + raise ValueError(f"prompt_2 must be of length 1 or {batch_size}") + + pooled_prompt_embeds_2 = self._get_clip_prompt_embeds( + self.tokenizer_2, self.text_encoder_2, prompt_2, max_sequence_length, device, dtype + ) + + if pooled_prompt_embeds_2.shape[0] == 1 and batch_size > 1: + pooled_prompt_embeds_2 = pooled_prompt_embeds_2.repeat(batch_size, 1) + + if do_classifier_free_guidance and negative_pooled_prompt_embeds is None: + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_2 = [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + + if len(negative_prompt_2) > 1 and len(negative_prompt_2) != batch_size: + raise ValueError(f"negative_prompt_2 must be of length 1 or {batch_size}") + + negative_pooled_prompt_embeds_2 = self._get_clip_prompt_embeds( + self.tokenizer_2, self.text_encoder_2, negative_prompt_2, max_sequence_length, device, dtype + ) + + if negative_pooled_prompt_embeds_2.shape[0] == 1 and batch_size > 1: + negative_pooled_prompt_embeds_2 = negative_pooled_prompt_embeds_2.repeat(batch_size, 1) + + if pooled_prompt_embeds is None: + pooled_prompt_embeds = torch.cat([pooled_prompt_embeds_1, pooled_prompt_embeds_2], dim=-1) + + if do_classifier_free_guidance and negative_pooled_prompt_embeds is None: + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embeds_1, negative_pooled_prompt_embeds_2], dim=-1 + ) + + if prompt_embeds_t5 is None: + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + if len(prompt_3) > 1 and len(prompt_3) != batch_size: + raise ValueError(f"prompt_3 must be of length 1 or {batch_size}") + + prompt_embeds_t5 = self._get_t5_prompt_embeds(prompt_3, max_sequence_length, device, dtype) + + if prompt_embeds_t5.shape[0] == 1 and batch_size > 1: + prompt_embeds_t5 = prompt_embeds_t5.repeat(batch_size, 1, 1) + + if do_classifier_free_guidance and negative_prompt_embeds_t5 is None: + negative_prompt_3 = negative_prompt_3 or negative_prompt + negative_prompt_3 = [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + + if len(negative_prompt_3) > 1 and len(negative_prompt_3) != batch_size: + raise ValueError(f"negative_prompt_3 must be of length 1 or {batch_size}") + + negative_prompt_embeds_t5 = self._get_t5_prompt_embeds( + negative_prompt_3, max_sequence_length, device, dtype + ) + + if negative_prompt_embeds_t5.shape[0] == 1 and batch_size > 1: + negative_prompt_embeds_t5 = negative_prompt_embeds_t5.repeat(batch_size, 1, 1) + + if prompt_embeds_llama3 is None: + prompt_4 = prompt_4 or prompt + prompt_4 = [prompt_4] if isinstance(prompt_4, str) else prompt_4 + + if len(prompt_4) > 1 and len(prompt_4) != batch_size: + raise ValueError(f"prompt_4 must be of length 1 or {batch_size}") + + prompt_embeds_llama3 = self._get_llama3_prompt_embeds(prompt_4, max_sequence_length, device, dtype) + + if prompt_embeds_llama3.shape[0] == 1 and batch_size > 1: + prompt_embeds_llama3 = prompt_embeds_llama3.repeat(1, batch_size, 1, 1) + + if do_classifier_free_guidance and negative_prompt_embeds_llama3 is None: + negative_prompt_4 = negative_prompt_4 or negative_prompt + negative_prompt_4 = [negative_prompt_4] if isinstance(negative_prompt_4, str) else negative_prompt_4 + + if len(negative_prompt_4) > 1 and len(negative_prompt_4) != batch_size: + raise ValueError(f"negative_prompt_4 must be of length 1 or {batch_size}") + + negative_prompt_embeds_llama3 = self._get_llama3_prompt_embeds( + negative_prompt_4, max_sequence_length, device, dtype + ) + + if negative_prompt_embeds_llama3.shape[0] == 1 and batch_size > 1: + negative_prompt_embeds_llama3 = negative_prompt_embeds_llama3.repeat(1, batch_size, 1, 1) + + # duplicate pooled_prompt_embeds for each generation per prompt + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + # duplicate t5_prompt_embeds for batch_size and num_images_per_prompt + bs_embed, seq_len, _ = prompt_embeds_t5.shape + if bs_embed == 1 and batch_size > 1: + prompt_embeds_t5 = prompt_embeds_t5.repeat(batch_size, 1, 1) + elif bs_embed > 1 and bs_embed != batch_size: + raise ValueError(f"cannot duplicate prompt_embeds_t5 of batch size {bs_embed}") + prompt_embeds_t5 = prompt_embeds_t5.repeat(1, num_images_per_prompt, 1) + prompt_embeds_t5 = prompt_embeds_t5.view(batch_size * num_images_per_prompt, seq_len, -1) + + # duplicate llama3_prompt_embeds for batch_size and num_images_per_prompt + _, bs_embed, seq_len, dim = prompt_embeds_llama3.shape + if bs_embed == 1 and batch_size > 1: + prompt_embeds_llama3 = prompt_embeds_llama3.repeat(1, batch_size, 1, 1) + elif bs_embed > 1 and bs_embed != batch_size: + raise ValueError(f"cannot duplicate prompt_embeds_llama3 of batch size {bs_embed}") + prompt_embeds_llama3 = prompt_embeds_llama3.repeat(1, 1, num_images_per_prompt, 1) + prompt_embeds_llama3 = prompt_embeds_llama3.view(-1, batch_size * num_images_per_prompt, seq_len, dim) + + if do_classifier_free_guidance: + # duplicate negative_pooled_prompt_embeds for batch_size and num_images_per_prompt + bs_embed, seq_len = negative_pooled_prompt_embeds.shape + if bs_embed == 1 and batch_size > 1: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(batch_size, 1) + elif bs_embed > 1 and bs_embed != batch_size: + raise ValueError(f"cannot duplicate negative_pooled_prompt_embeds of batch size {bs_embed}") + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt) + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + # duplicate negative_t5_prompt_embeds for batch_size and num_images_per_prompt + bs_embed, seq_len, _ = negative_prompt_embeds_t5.shape + if bs_embed == 1 and batch_size > 1: + negative_prompt_embeds_t5 = negative_prompt_embeds_t5.repeat(batch_size, 1, 1) + elif bs_embed > 1 and bs_embed != batch_size: + raise ValueError(f"cannot duplicate negative_prompt_embeds_t5 of batch size {bs_embed}") + negative_prompt_embeds_t5 = negative_prompt_embeds_t5.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds_t5 = negative_prompt_embeds_t5.view(batch_size * num_images_per_prompt, seq_len, -1) + + # duplicate negative_prompt_embeds_llama3 for batch_size and num_images_per_prompt + _, bs_embed, seq_len, dim = negative_prompt_embeds_llama3.shape + if bs_embed == 1 and batch_size > 1: + negative_prompt_embeds_llama3 = negative_prompt_embeds_llama3.repeat(1, batch_size, 1, 1) + elif bs_embed > 1 and bs_embed != batch_size: + raise ValueError(f"cannot duplicate negative_prompt_embeds_llama3 of batch size {bs_embed}") + negative_prompt_embeds_llama3 = negative_prompt_embeds_llama3.repeat(1, 1, num_images_per_prompt, 1) + negative_prompt_embeds_llama3 = negative_prompt_embeds_llama3.view( + -1, batch_size * num_images_per_prompt, seq_len, dim + ) + + return ( + prompt_embeds_t5, + negative_prompt_embeds_t5, + prompt_embeds_llama3, + negative_prompt_embeds_llama3, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + prompt_4, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + negative_prompt_4=None, + prompt_embeds_t5=None, + prompt_embeds_llama3=None, + negative_prompt_embeds_t5=None, + negative_prompt_embeds_llama3=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and pooled_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `pooled_prompt_embeds`: {pooled_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and pooled_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `pooled_prompt_embeds`: {pooled_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds_t5 is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_3} and `prompt_embeds_t5`: {prompt_embeds_t5}. Please make sure to" + " only forward one of the two." + ) + elif prompt_4 is not None and prompt_embeds_llama3 is not None: + raise ValueError( + f"Cannot forward both `prompt_4`: {prompt_4} and `prompt_embeds_llama3`: {prompt_embeds_llama3}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and pooled_prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `pooled_prompt_embeds`. Cannot leave both `prompt` and `pooled_prompt_embeds` undefined." + ) + elif prompt is None and prompt_embeds_t5 is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_t5`. Cannot leave both `prompt` and `prompt_embeds_t5` undefined." + ) + elif prompt is None and prompt_embeds_llama3 is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_llama3`. Cannot leave both `prompt` and `prompt_embeds_llama3` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + elif prompt_4 is not None and (not isinstance(prompt_4, str) and not isinstance(prompt_4, list)): + raise ValueError(f"`prompt_4` has to be of type `str` or `list` but is {type(prompt_4)}") + + if negative_prompt is not None and negative_pooled_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_pooled_prompt_embeds`:" + f" {negative_pooled_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_pooled_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_pooled_prompt_embeds`:" + f" {negative_pooled_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds_t5 is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds_t5`:" + f" {negative_prompt_embeds_t5}. Please make sure to only forward one of the two." + ) + elif negative_prompt_4 is not None and negative_prompt_embeds_llama3 is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_4`: {negative_prompt_4} and `negative_prompt_embeds_llama3`:" + f" {negative_prompt_embeds_llama3}. Please make sure to only forward one of the two." + ) + + if pooled_prompt_embeds is not None and negative_pooled_prompt_embeds is not None: + if pooled_prompt_embeds.shape != negative_pooled_prompt_embeds.shape: + raise ValueError( + "`pooled_prompt_embeds` and `negative_pooled_prompt_embeds` must have the same shape when passed directly, but" + f" got: `pooled_prompt_embeds` {pooled_prompt_embeds.shape} != `negative_pooled_prompt_embeds`" + f" {negative_pooled_prompt_embeds.shape}." + ) + if prompt_embeds_t5 is not None and negative_prompt_embeds_t5 is not None: + if prompt_embeds_t5.shape != negative_prompt_embeds_t5.shape: + raise ValueError( + "`prompt_embeds_t5` and `negative_prompt_embeds_t5` must have the same shape when passed directly, but" + f" got: `prompt_embeds_t5` {prompt_embeds_t5.shape} != `negative_prompt_embeds_t5`" + f" {negative_prompt_embeds_t5.shape}." + ) + if prompt_embeds_llama3 is not None and negative_prompt_embeds_llama3 is not None: + if prompt_embeds_llama3.shape != negative_prompt_embeds_llama3.shape: + raise ValueError( + "`prompt_embeds_llama3` and `negative_prompt_embeds_llama3` must have the same shape when passed directly, but" + f" got: `prompt_embeds_llama3` {prompt_embeds_llama3.shape} != `negative_prompt_embeds_llama3`" + f" {negative_prompt_embeds_llama3.shape}." + ) + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + prompt_3: str | list[str] | None = None, + prompt_4: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + negative_prompt_4: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds_t5: torch.FloatTensor | None = None, + prompt_embeds_llama3: torch.FloatTensor | None = None, + negative_prompt_embeds_t5: torch.FloatTensor | None = None, + negative_prompt_embeds_llama3: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 128, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead. + prompt_4 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_4` and `text_encoder_4`. If not defined, `prompt` is + will be used instead. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 3.5): + Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with `prompt` at the expense of lower image quality. + + Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_4 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_4` and + `text_encoder_4`. If not defined, `negative_prompt` is used in all the text-encoders. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds_t5 (`torch.FloatTensor`, *optional*): + Pre-generated T5 text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If + not provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_llama3 (`torch.FloatTensor`, *optional*): + Pre-generated LLaMA3 text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds_t5 (`torch.FloatTensor`, *optional*): + Pre-generated negative T5 text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, embeddings will be generated from `negative_prompt` input argument. + negative_prompt_embeds_llama3 (`torch.FloatTensor`, *optional*): + Pre-generated negative LLaMA3 text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, embeddings will be generated from `negative_prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 128): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.hidream_image.HiDreamImagePipelineOutput`] or `tuple`: + [`~pipelines.hidream_image.HiDreamImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated. images. + """ + + prompt_embeds = kwargs.get("prompt_embeds", None) + negative_prompt_embeds = kwargs.get("negative_prompt_embeds", None) + + if prompt_embeds is not None: + deprecation_message = "The `prompt_embeds` argument is deprecated. Please use `prompt_embeds_t5` and `prompt_embeds_llama3` instead." + deprecate("prompt_embeds", "0.35.0", deprecation_message) + prompt_embeds_t5 = prompt_embeds[0] + prompt_embeds_llama3 = prompt_embeds[1] + + if negative_prompt_embeds is not None: + deprecation_message = "The `negative_prompt_embeds` argument is deprecated. Please use `negative_prompt_embeds_t5` and `negative_prompt_embeds_llama3` instead." + deprecate("negative_prompt_embeds", "0.35.0", deprecation_message) + negative_prompt_embeds_t5 = negative_prompt_embeds[0] + negative_prompt_embeds_llama3 = negative_prompt_embeds[1] + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + division = self.vae_scale_factor * 2 + S_max = (self.default_sample_size * self.vae_scale_factor) ** 2 + scale = S_max / (width * height) + scale = math.sqrt(scale) + width, height = int(width * scale // division * division), int(height * scale // division * division) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + prompt_4, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + negative_prompt_4=negative_prompt_4, + prompt_embeds_t5=prompt_embeds_t5, + prompt_embeds_llama3=prompt_embeds_llama3, + negative_prompt_embeds_t5=negative_prompt_embeds_t5, + negative_prompt_embeds_llama3=negative_prompt_embeds_llama3, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + elif pooled_prompt_embeds is not None: + batch_size = pooled_prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode prompt + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + ( + prompt_embeds_t5, + negative_prompt_embeds_t5, + prompt_embeds_llama3, + negative_prompt_embeds_llama3, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + prompt_4=prompt_4, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + negative_prompt_4=negative_prompt_4, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds_t5=prompt_embeds_t5, + prompt_embeds_llama3=prompt_embeds_llama3, + negative_prompt_embeds_t5=negative_prompt_embeds_t5, + negative_prompt_embeds_llama3=negative_prompt_embeds_llama3, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_classifier_free_guidance: + prompt_embeds_t5 = torch.cat([negative_prompt_embeds_t5, prompt_embeds_t5], dim=0) + prompt_embeds_llama3 = torch.cat([negative_prompt_embeds_llama3, prompt_embeds_llama3], dim=1) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + pooled_prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + mu = calculate_shift(self.transformer.max_seq) + scheduler_kwargs = {"mu": mu} + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + if isinstance(self.scheduler, UniPCMultistepScheduler): + self.scheduler.set_timesteps(num_inference_steps, device=timestep_device) # , shift=math.exp(mu)) + timesteps = self.scheduler.timesteps + else: + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + **scheduler_kwargs, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timesteps=timestep, + encoder_hidden_states_t5=prompt_embeds_t5, + encoder_hidden_states_llama3=prompt_embeds_llama3, + pooled_embeds=pooled_prompt_embeds, + return_dict=False, + )[0] + noise_pred = -noise_pred + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds_t5 = callback_outputs.pop("prompt_embeds_t5", prompt_embeds_t5) + prompt_embeds_llama3 = callback_outputs.pop("prompt_embeds_llama3", prompt_embeds_llama3) + pooled_prompt_embeds = callback_outputs.pop("pooled_prompt_embeds", pooled_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return HiDreamImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/hidream_image/pipeline_output.py b/diffusers/pipelines/hidream_image/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..1802c7220691d5b14c99580de3a6bae5846b838d --- /dev/null +++ b/diffusers/pipelines/hidream_image/pipeline_output.py @@ -0,0 +1,34 @@ +# Copyright 2025 HiDream-ai Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class HiDreamImagePipelineOutput(BaseOutput): + """ + Output class for HiDreamImage pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/hunyuan_image/__init__.py b/diffusers/pipelines/hunyuan_image/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7da72fa12b2c18c1b9ec7f9fb9b2170e4b516041 --- /dev/null +++ b/diffusers/pipelines/hunyuan_image/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_hunyuanimage"] = ["HunyuanImagePipeline"] + _import_structure["pipeline_hunyuanimage_refiner"] = ["HunyuanImageRefinerPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_hunyuanimage import HunyuanImagePipeline + from .pipeline_hunyuanimage_refiner import HunyuanImageRefinerPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/hunyuan_image/pipeline_hunyuanimage.py b/diffusers/pipelines/hunyuan_image/pipeline_hunyuanimage.py new file mode 100644 index 0000000000000000000000000000000000000000..50239e9afa22678bfe1532291fa75cda6e756988 --- /dev/null +++ b/diffusers/pipelines/hunyuan_image/pipeline_hunyuanimage.py @@ -0,0 +1,868 @@ +# Copyright 2025 Hunyuan-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import inspect +import re +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ByT5Tokenizer, Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, T5EncoderModel + +from ...guiders import AdaptiveProjectedMixGuidance +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKLHunyuanImage, HunyuanImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HunyuanImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import HunyuanImagePipeline + + >>> pipe = HunyuanImagePipeline.from_pretrained( + ... "hunyuanvideo-community/HunyuanImage-2.1-Diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, negative_prompt="", num_inference_steps=50).images[0] + >>> image.save("hunyuanimage.png") + ``` +""" + + +def extract_glyph_text(prompt: str): + """ + Extract text enclosed in quotes for glyph rendering. + + Finds text in single quotes, double quotes, and Chinese quotes, then formats it for byT5 processing. + + Args: + prompt: Input text prompt + + Returns: + Formatted glyph text string or None if no quoted text found + """ + text_prompt_texts = [] + pattern_quote_single = r"\'(.*?)\'" + pattern_quote_double = r"\"(.*?)\"" + pattern_quote_chinese_single = r"‘(.*?)’" + pattern_quote_chinese_double = r"“(.*?)”" + + matches_quote_single = re.findall(pattern_quote_single, prompt) + matches_quote_double = re.findall(pattern_quote_double, prompt) + matches_quote_chinese_single = re.findall(pattern_quote_chinese_single, prompt) + matches_quote_chinese_double = re.findall(pattern_quote_chinese_double, prompt) + + text_prompt_texts.extend(matches_quote_single) + text_prompt_texts.extend(matches_quote_double) + text_prompt_texts.extend(matches_quote_chinese_single) + text_prompt_texts.extend(matches_quote_chinese_double) + + if text_prompt_texts: + glyph_text_formatted = ". ".join([f'Text "{text}"' for text in text_prompt_texts]) + ". " + else: + glyph_text_formatted = None + + return glyph_text_formatted + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class HunyuanImagePipeline(DiffusionPipeline): + r""" + The HunyuanImage pipeline for text-to-image generation. + + Args: + transformer ([`HunyuanImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLHunyuanImage`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): Tokenizer of class [Qwen2Tokenizer]. + text_encoder_2 ([`T5EncoderModel`]): + [T5EncoderModel](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel) + variant. + tokenizer_2 (`ByT5Tokenizer`): Tokenizer of class [ByT5Tokenizer] + guider ([`AdaptiveProjectedMixGuidance`]): + [AdaptiveProjectedMixGuidance]to be used to guide the image generation. + ocr_guider ([`AdaptiveProjectedMixGuidance`], *optional*): + [AdaptiveProjectedMixGuidance] to be used to guide the image generation when text rendering is needed. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + _optional_components = ["ocr_guider", "guider"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLHunyuanImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: ByT5Tokenizer, + transformer: HunyuanImageTransformer2DModel, + guider: AdaptiveProjectedMixGuidance | None = None, + ocr_guider: AdaptiveProjectedMixGuidance | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + guider=guider, + ocr_guider=ocr_guider, + ) + + self.vae_scale_factor = self.vae.config.spatial_compression_ratio if getattr(self, "vae", None) else 32 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.tokenizer_max_length = 1000 + self.tokenizer_2_max_length = 128 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 64 + + def _get_qwen_prompt_embeds( + self, + tokenizer: Qwen2Tokenizer, + text_encoder: Qwen2_5_VLForConditionalGeneration, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + tokenizer_max_length: int = 1000, + template: str = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>", + drop_idx: int = 34, + hidden_state_skip_layer: int = 2, + ): + device = device or self._execution_device + dtype = dtype or text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + txt = [template.format(e) for e in prompt] + txt_tokens = tokenizer( + txt, max_length=tokenizer_max_length + drop_idx, padding="max_length", truncation=True, return_tensors="pt" + ).to(device) + + encoder_hidden_states = text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + prompt_embeds = encoder_hidden_states.hidden_states[-(hidden_state_skip_layer + 1)] + + prompt_embeds = prompt_embeds[:, drop_idx:] + encoder_attention_mask = txt_tokens.attention_mask[:, drop_idx:] + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + encoder_attention_mask = encoder_attention_mask.to(device=device) + + return prompt_embeds, encoder_attention_mask + + def _get_byt5_prompt_embeds( + self, + tokenizer: ByT5Tokenizer, + text_encoder: T5EncoderModel, + prompt: str, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + tokenizer_max_length: int = 128, + ): + device = device or self._execution_device + dtype = dtype or text_encoder.dtype + + if isinstance(prompt, list): + raise ValueError("byt5 prompt should be a string") + elif prompt is None: + raise ValueError("byt5 prompt should not be None") + + txt_tokens = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer_max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ).to(device) + + prompt_embeds = text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask.float(), + )[0] + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + encoder_attention_mask = txt_tokens.attention_mask.to(device=device) + + return prompt_embeds, encoder_attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + batch_size: int = 1, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + prompt_embeds_mask_2: torch.Tensor | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + batch_size (`int`): + batch size of prompts, defaults to 1 + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. If not provided, text embeddings will be generated from `prompt` input + argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated text mask. If not provided, text mask will be generated from `prompt` input argument. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text embeddings from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text mask from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + """ + device = device or self._execution_device + + if prompt is None: + prompt = [""] * batch_size + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds( + tokenizer=self.tokenizer, + text_encoder=self.text_encoder, + prompt=prompt, + device=device, + tokenizer_max_length=self.tokenizer_max_length, + template=self.prompt_template_encode, + drop_idx=self.prompt_template_encode_start_idx, + ) + + if prompt_embeds_2 is None: + prompt_embeds_2_list = [] + prompt_embeds_mask_2_list = [] + + glyph_texts = [extract_glyph_text(p) for p in prompt] + for glyph_text in glyph_texts: + if glyph_text is None: + glyph_text_embeds = torch.zeros( + (1, self.tokenizer_2_max_length, self.text_encoder_2.config.d_model), device=device + ) + glyph_text_embeds_mask = torch.zeros( + (1, self.tokenizer_2_max_length), device=device, dtype=torch.int64 + ) + else: + glyph_text_embeds, glyph_text_embeds_mask = self._get_byt5_prompt_embeds( + tokenizer=self.tokenizer_2, + text_encoder=self.text_encoder_2, + prompt=glyph_text, + device=device, + tokenizer_max_length=self.tokenizer_2_max_length, + ) + + prompt_embeds_2_list.append(glyph_text_embeds) + prompt_embeds_mask_2_list.append(glyph_text_embeds_mask) + + prompt_embeds_2 = torch.cat(prompt_embeds_2_list, dim=0) + prompt_embeds_mask_2 = torch.cat(prompt_embeds_mask_2_list, dim=0) + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + _, seq_len_2, _ = prompt_embeds_2.shape + prompt_embeds_2 = prompt_embeds_2.repeat(1, num_images_per_prompt, 1) + prompt_embeds_2 = prompt_embeds_2.view(batch_size * num_images_per_prompt, seq_len_2, -1) + prompt_embeds_mask_2 = prompt_embeds_mask_2.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask_2 = prompt_embeds_mask_2.view(batch_size * num_images_per_prompt, seq_len_2) + + return prompt_embeds, prompt_embeds_mask, prompt_embeds_2, prompt_embeds_mask_2 + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + prompt_embeds_2=None, + prompt_embeds_mask_2=None, + negative_prompt_embeds_2=None, + negative_prompt_embeds_mask_2=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if prompt is None and prompt_embeds_2 is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_2`. Cannot leave both `prompt` and `prompt_embeds_2` undefined." + ) + + if prompt_embeds_2 is not None and prompt_embeds_mask_2 is None: + raise ValueError( + "If `prompt_embeds_2` are provided, `prompt_embeds_mask_2` also have to be passed. Make sure to generate `prompt_embeds_mask_2` from the same text encoder that was used to generate `prompt_embeds_2`." + ) + if negative_prompt_embeds_2 is not None and negative_prompt_embeds_mask_2 is None: + raise ValueError( + "If `negative_prompt_embeds_2` are provided, `negative_prompt_embeds_mask_2` also have to be passed. Make sure to generate `negative_prompt_embeds_mask_2` from the same text encoder that was used to generate `negative_prompt_embeds_2`." + ) + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = int(height) // self.vae_scale_factor + width = int(width) // self.vae_scale_factor + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + return latents + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + distilled_guidance_scale: float | None = 3.25, + sigmas: list[float] | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + prompt_embeds_mask_2: torch.Tensor | None = None, + negative_prompt_embeds_2: torch.Tensor | None = None, + negative_prompt_embeds_mask_2: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined and negative_prompt_embeds is + not provided, will use an empty negative prompt. Ignored when not using guidance. ). + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + distilled_guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + is enabled by setting `distilled_guidance_scale > 1`. Higher guidance scale encourages to generate + images that are closely linked to the text `prompt`, usually at the expense of lower image quality. For + guidance distilled models, this parameter is required. For non-distilled models, this parameter will be + ignored. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated text embeddings mask. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, text embeddings mask will be generated from `prompt` input argument. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated text embeddings for ocr. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, text embeddings for ocr will be generated from `prompt` input argument. + prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated text embeddings mask for ocr. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, text embeddings mask for ocr will be generated from `prompt` input + argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings mask. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative text embeddings mask will be generated from `negative_prompt` + input argument. + negative_prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings for ocr. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative text embeddings for ocr will be generated from `negative_prompt` + input argument. + negative_prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings mask for ocr. Can be used to easily tweak text inputs, *e.g.* + prompt weighting. If not provided, negative text embeddings mask for ocr will be generated from + `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.hunyuan_image.HunyuanImagePipelineOutput`] or `tuple`: + [`~pipelines.hunyuan_image.HunyuanImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds_2=prompt_embeds_2, + prompt_embeds_mask_2=prompt_embeds_mask_2, + negative_prompt_embeds_2=negative_prompt_embeds_2, + negative_prompt_embeds_mask_2=negative_prompt_embeds_mask_2, + ) + + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. prepare prompt embeds + + prompt_embeds, prompt_embeds_mask, prompt_embeds_2, prompt_embeds_mask_2 = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + prompt_embeds_2=prompt_embeds_2, + prompt_embeds_mask_2=prompt_embeds_mask_2, + ) + + prompt_embeds = prompt_embeds.to(self.transformer.dtype) + prompt_embeds_2 = prompt_embeds_2.to(self.transformer.dtype) + + # select guider + if not torch.all(prompt_embeds_2 == 0) and self.ocr_guider is not None: + # prompt contains ocr and pipeline has a guider for ocr + guider = self.ocr_guider + elif self.guider is not None: + guider = self.guider + # distilled model does not use guidance method, use default guider with enabled=False + else: + guider = AdaptiveProjectedMixGuidance(enabled=False) + + if guider._enabled and guider.num_conditions > 1: + ( + negative_prompt_embeds, + negative_prompt_embeds_mask, + negative_prompt_embeds_2, + negative_prompt_embeds_mask_2, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + prompt_embeds_2=negative_prompt_embeds_2, + prompt_embeds_mask_2=negative_prompt_embeds_mask_2, + ) + + negative_prompt_embeds = negative_prompt_embeds.to(self.transformer.dtype) + negative_prompt_embeds_2 = negative_prompt_embeds_2.to(self.transformer.dtype) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance (for guidance-distilled model) + if self.transformer.config.guidance_embeds and distilled_guidance_scale is None: + raise ValueError("`distilled_guidance_scale` is required for guidance-distilled model.") + + if self.transformer.config.guidance_embeds: + guidance = ( + torch.tensor( + [distilled_guidance_scale] * latents.shape[0], dtype=self.transformer.dtype, device=device + ) + * 1000.0 + ) + + else: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + if self.transformer.config.use_meanflow: + if i == len(timesteps) - 1: + timestep_r = torch.tensor([0.0], device=device) + else: + timestep_r = timesteps[i + 1] + timestep_r = timestep_r.expand(latents.shape[0]).to(latents.dtype) + else: + timestep_r = None + + # Step 1: Collect model inputs needed for the guidance method + # conditional inputs should always be first element in the tuple + guider_inputs = { + "encoder_hidden_states": (prompt_embeds, negative_prompt_embeds), + "encoder_attention_mask": (prompt_embeds_mask, negative_prompt_embeds_mask), + "encoder_hidden_states_2": (prompt_embeds_2, negative_prompt_embeds_2), + "encoder_attention_mask_2": (prompt_embeds_mask_2, negative_prompt_embeds_mask_2), + } + + # Step 2: Update guider's internal state for this denoising step + guider.set_state(step=i, num_inference_steps=num_inference_steps, timestep=t) + + # Step 3: Prepare batched model inputs based on the guidance method + # The guider splits model inputs into separate batches for conditional/unconditional predictions. + # For CFG with guider_inputs = {"encoder_hidden_states": (prompt_embeds, negative_prompt_embeds)}: + # you will get a guider_state with two batches: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "__guidance_identifier__": "pred_cond"}, # conditional batch + # {"encoder_hidden_states": negative_prompt_embeds, "__guidance_identifier__": "pred_uncond"}, # unconditional batch + # ] + # Other guidance methods may return 1 batch (no guidance) or 3+ batches (e.g., PAG, APG). + guider_state = guider.prepare_inputs(guider_inputs) + # Step 4: Run the denoiser for each batch + # Each batch in guider_state represents a different conditioning (conditional, unconditional, etc.). + # We run the model once per batch and store the noise prediction in guider_state_batch.noise_pred. + for guider_state_batch in guider_state: + guider.prepare_models(self.transformer) + + # Extract conditioning kwargs for this batch (e.g., encoder_hidden_states) + cond_kwargs = { + input_name: getattr(guider_state_batch, input_name) for input_name in guider_inputs.keys() + } + + # e.g. "pred_cond"/"pred_uncond" + context_name = getattr(guider_state_batch, guider._identifier_key) + with self.transformer.cache_context(context_name): + # Run denoiser and store noise prediction in this batch + guider_state_batch.noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep, + timestep_r=timestep_r, + guidance=guidance, + attention_kwargs=self.attention_kwargs, + return_dict=False, + **cond_kwargs, + )[0] + + # Cleanup model (e.g., remove hooks) + guider.cleanup_models(self.transformer) + + # Step 5: Combine predictions using the guidance method + # The guider takes all noise predictions from guider_state and combines them according to the guidance algorithm. + # Continuing the CFG example, the guider receives: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "noise_pred": noise_pred_cond, "__guidance_identifier__": "pred_cond"}, # batch 0 + # {"encoder_hidden_states": negative_prompt_embeds, "noise_pred": noise_pred_uncond, "__guidance_identifier__": "pred_uncond"}, # batch 1 + # ] + # And extracts predictions using the __guidance_identifier__: + # pred_cond = guider_state[0]["noise_pred"] # extracts noise_pred_cond + # pred_uncond = guider_state[1]["noise_pred"] # extracts noise_pred_uncond + # Then applies CFG formula: + # noise_pred = pred_uncond + guidance_scale * (pred_cond - pred_uncond) + # Returns GuiderOutput(pred=noise_pred, pred_cond=pred_cond, pred_uncond=pred_uncond) + noise_pred = guider(guider_state)[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return HunyuanImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/hunyuan_image/pipeline_hunyuanimage_refiner.py b/diffusers/pipelines/hunyuan_image/pipeline_hunyuanimage_refiner.py new file mode 100644 index 0000000000000000000000000000000000000000..efdb5505e60411a2dd49f0fc3a4ab4f029b52eed --- /dev/null +++ b/diffusers/pipelines/hunyuan_image/pipeline_hunyuanimage_refiner.py @@ -0,0 +1,760 @@ +# Copyright 2025 Hunyuan-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...guiders import AdaptiveProjectedMixGuidance +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...models import AutoencoderKLHunyuanImageRefiner, HunyuanImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HunyuanImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import HunyuanImageRefinerPipeline + + >>> pipe = HunyuanImageRefinerPipeline.from_pretrained( + ... "hunyuanvideo-community/HunyuanImage-2.1-Refiner-Diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> image = load_image("path/to/image.png") + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, image=image, num_inference_steps=4).images[0] + >>> image.save("hunyuanimage.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class HunyuanImageRefinerPipeline(DiffusionPipeline): + r""" + The HunyuanImage pipeline for text-to-image generation. + + Args: + transformer ([`HunyuanImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLHunyuanImageRefiner`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): Tokenizer of class [Qwen2Tokenizer]. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + _optional_components = ["guider"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLHunyuanImageRefiner, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: HunyuanImageTransformer2DModel, + guider: AdaptiveProjectedMixGuidance | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + guider=guider, + ) + + self.vae_scale_factor = self.vae.config.spatial_compression_ratio if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.tokenizer_max_length = 256 + self.prompt_template_encode = "<|start_header_id|>system<|end_header_id|>\n\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>" + self.prompt_template_encode_start_idx = 36 + self.default_sample_size = 64 + self.latent_channels = self.transformer.config.in_channels // 2 if getattr(self, "transformer", None) else 64 + + # Copied from diffusers.pipelines.hunyuan_image.pipeline_hunyuanimage.HunyuanImagePipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + tokenizer: Qwen2Tokenizer, + text_encoder: Qwen2_5_VLForConditionalGeneration, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + tokenizer_max_length: int = 1000, + template: str = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>", + drop_idx: int = 34, + hidden_state_skip_layer: int = 2, + ): + device = device or self._execution_device + dtype = dtype or text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + txt = [template.format(e) for e in prompt] + txt_tokens = tokenizer( + txt, max_length=tokenizer_max_length + drop_idx, padding="max_length", truncation=True, return_tensors="pt" + ).to(device) + + encoder_hidden_states = text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + prompt_embeds = encoder_hidden_states.hidden_states[-(hidden_state_skip_layer + 1)] + + prompt_embeds = prompt_embeds[:, drop_idx:] + encoder_attention_mask = txt_tokens.attention_mask[:, drop_idx:] + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + encoder_attention_mask = encoder_attention_mask.to(device=device) + + return prompt_embeds, encoder_attention_mask + + def encode_prompt( + self, + prompt: str | list[str] | None = None, + device: torch.device | None = None, + batch_size: int = 1, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + batch_size (`int`): + batch size of prompts, defaults to 1 + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. If not provided, text embeddings will be generated from `prompt` input + argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated text mask. If not provided, text mask will be generated from `prompt` input argument. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text embeddings from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text mask from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + """ + device = device or self._execution_device + + if prompt is None: + prompt = [""] * batch_size + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds( + tokenizer=self.tokenizer, + text_encoder=self.text_encoder, + prompt=prompt, + device=device, + tokenizer_max_length=self.tokenizer_max_length, + template=self.prompt_template_encode, + drop_idx=self.prompt_template_encode_start_idx, + ) + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + def prepare_latents( + self, + image_latents, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + strength=0.25, + ): + height = int(height) // self.vae_scale_factor + width = int(width) // self.vae_scale_factor + + shape = (batch_size, num_channels_latents, 1, height, width) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + cond_latents = strength * noise + (1 - strength) * image_latents + + return latents, cond_latents + + @staticmethod + def _reorder_image_tokens(image_latents): + image_latents = torch.cat((image_latents[:, :, :1], image_latents), dim=2) + batch_size, num_latent_channels, num_latent_frames, latent_height, latent_width = image_latents.shape + image_latents = image_latents.permute(0, 2, 1, 3, 4) + image_latents = image_latents.reshape( + batch_size, num_latent_frames // 2, num_latent_channels * 2, latent_height, latent_width + ) + image_latents = image_latents.permute(0, 2, 1, 3, 4).contiguous() + + return image_latents + + @staticmethod + def _restore_image_tokens_order(latents): + """Restore image tokens order by splitting channels and removing first frame slice.""" + batch_size, num_latent_channels, num_latent_frames, latent_height, latent_width = latents.shape + + latents = latents.permute(0, 2, 1, 3, 4) # B, F, C, H, W + latents = latents.reshape( + batch_size, num_latent_frames * 2, num_latent_channels // 2, latent_height, latent_width + ) # B, F*2, C//2, H, W + + latents = latents.permute(0, 2, 1, 3, 4) # B, C//2, F*2, H, W + # Remove first frame slice + latents = latents[:, :, 1:] + + return latents + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="sample") + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="sample") + image_latents = self._reorder_image_tokens(image_latents) + + image_latents = image_latents * self.vae.config.scaling_factor + + return image_latents + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + distilled_guidance_scale: float | None = 3.25, + image: PipelineImageInput | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 4, + sigmas: list[float] | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, will use an empty negative + prompt. Ignored when not using guidance. + distilled_guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + is enabled by setting `distilled_guidance_scale > 1`. Higher guidance scale encourages to generate + images that are closely linked to the text `prompt`, usually at the expense of lower image quality. For + guidance distilled models, this parameter is required. For non-distilled models, this parameter will be + ignored. + image (`PipelineImageInput`, *optional*): + The input image to be refined. + num_images_per_prompt (`int`, *optional*, defaults to 1): + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.hunyuan_image.HunyuanImagePipelineOutput`] or `tuple`: + [`~pipelines.hunyuan_image.HunyuanImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. process image + if image is not None and isinstance(image, torch.Tensor) and image.shape[1] == self.latent_channels: + image_latents = image + else: + image = self.image_processor.preprocess(image, height, width) + image = image.unsqueeze(2).to(device, dtype=self.vae.dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + + # 3.prepare prompt embeds + + if self.guider is not None: + guider = self.guider + else: + # distilled model does not use guidance method, use default guider with enabled=False + guider = AdaptiveProjectedMixGuidance(enabled=False) + + requires_unconditional_embeds = guider._enabled and guider.num_conditions > 1 + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + ) + + prompt_embeds = prompt_embeds.to(self.transformer.dtype) + + if requires_unconditional_embeds: + ( + negative_prompt_embeds, + negative_prompt_embeds_mask, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + ) + + negative_prompt_embeds = negative_prompt_embeds.to(self.transformer.dtype) + + # 4. Prepare latent variables + latents, cond_latents = self.prepare_latents( + image_latents=image_latents, + batch_size=batch_size * num_images_per_prompt, + num_channels_latents=self.latent_channels, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance (this pipeline only supports guidance-distilled models) + if distilled_guidance_scale is None: + raise ValueError("`distilled_guidance_scale` is required for guidance-distilled model.") + guidance = ( + torch.tensor([distilled_guidance_scale] * latents.shape[0], dtype=self.transformer.dtype, device=device) + * 1000.0 + ) + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + latent_model_input = torch.cat([latents, cond_latents], dim=1).to(self.transformer.dtype) + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + # Step 1: Collect model inputs needed for the guidance method + # conditional inputs should always be first element in the tuple + guider_inputs = { + "encoder_hidden_states": (prompt_embeds, negative_prompt_embeds), + "encoder_attention_mask": (prompt_embeds_mask, negative_prompt_embeds_mask), + } + + # Step 2: Update guider's internal state for this denoising step + guider.set_state(step=i, num_inference_steps=num_inference_steps, timestep=t) + + # Step 3: Prepare batched model inputs based on the guidance method + # The guider splits model inputs into separate batches for conditional/unconditional predictions. + # For CFG with guider_inputs = {"encoder_hidden_states": (prompt_embeds, negative_prompt_embeds)}: + # you will get a guider_state with two batches: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "__guidance_identifier__": "pred_cond"}, # conditional batch + # {"encoder_hidden_states": negative_prompt_embeds, "__guidance_identifier__": "pred_uncond"}, # unconditional batch + # ] + # Other guidance methods may return 1 batch (no guidance) or 3+ batches (e.g., PAG, APG). + guider_state = guider.prepare_inputs(guider_inputs) + + # Step 4: Run the denoiser for each batch + # Each batch in guider_state represents a different conditioning (conditional, unconditional, etc.). + # We run the model once per batch and store the noise prediction in guider_state_batch.noise_pred. + for guider_state_batch in guider_state: + guider.prepare_models(self.transformer) + + # Extract conditioning kwargs for this batch (e.g., encoder_hidden_states) + cond_kwargs = { + input_name: getattr(guider_state_batch, input_name) for input_name in guider_inputs.keys() + } + + # e.g. "pred_cond"/"pred_uncond" + context_name = getattr(guider_state_batch, guider._identifier_key) + with self.transformer.cache_context(context_name): + # Run denoiser and store noise prediction in this batch + guider_state_batch.noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + guidance=guidance, + attention_kwargs=self.attention_kwargs, + return_dict=False, + **cond_kwargs, + )[0] + + # Cleanup model (e.g., remove hooks) + guider.cleanup_models(self.transformer) + + # Step 5: Combine predictions using the guidance method + # The guider takes all noise predictions from guider_state and combines them according to the guidance algorithm. + # Continuing the CFG example, the guider receives: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "noise_pred": noise_pred_cond, "__guidance_identifier__": "pred_cond"}, # batch 0 + # {"encoder_hidden_states": negative_prompt_embeds, "noise_pred": noise_pred_uncond, "__guidance_identifier__": "pred_uncond"}, # batch 1 + # ] + # And extracts predictions using the __guidance_identifier__: + # pred_cond = guider_state[0]["noise_pred"] # extracts noise_pred_cond + # pred_uncond = guider_state[1]["noise_pred"] # extracts noise_pred_uncond + # Then applies CFG formula: + # noise_pred = pred_uncond + guidance_scale * (pred_cond - pred_uncond) + # Returns GuiderOutput(pred=noise_pred, pred_cond=pred_cond, pred_uncond=pred_uncond) + noise_pred = guider(guider_state)[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor + latents = self._restore_image_tokens_order(latents) + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image.squeeze(2), output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return HunyuanImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/hunyuan_image/pipeline_output.py b/diffusers/pipelines/hunyuan_image/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..169436b7d86f78d4584396a42a2d3f1d0209d3a3 --- /dev/null +++ b/diffusers/pipelines/hunyuan_image/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class HunyuanImagePipelineOutput(BaseOutput): + """ + Output class for HunyuanImage pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image, np.ndarray] diff --git a/diffusers/pipelines/hunyuan_video/__init__.py b/diffusers/pipelines/hunyuan_video/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..d42d38fac9793db2c2fb94c4b41d3755be75d898 --- /dev/null +++ b/diffusers/pipelines/hunyuan_video/__init__.py @@ -0,0 +1,54 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_hunyuan_skyreels_image2video"] = ["HunyuanSkyreelsImageToVideoPipeline"] + _import_structure["pipeline_hunyuan_video"] = ["HunyuanVideoPipeline"] + _import_structure["pipeline_hunyuan_video_framepack"] = ["HunyuanVideoFramepackPipeline"] + _import_structure["pipeline_hunyuan_video_image2video"] = ["HunyuanVideoImageToVideoPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_hunyuan_skyreels_image2video import HunyuanSkyreelsImageToVideoPipeline + from .pipeline_hunyuan_video import HunyuanVideoPipeline + from .pipeline_hunyuan_video_framepack import HunyuanVideoFramepackPipeline + from .pipeline_hunyuan_video_image2video import HunyuanVideoImageToVideoPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_skyreels_image2video.py b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_skyreels_image2video.py new file mode 100644 index 0000000000000000000000000000000000000000..bd54f2563b524864c4c9704b30ec371da677fa98 --- /dev/null +++ b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_skyreels_image2video.py @@ -0,0 +1,786 @@ +# Copyright 2025 The HunyuanVideo Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import HunyuanVideoLoraLoaderMixin +from ...models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HunyuanVideoPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import HunyuanSkyreelsImageToVideoPipeline, HunyuanVideoTransformer3DModel + >>> from diffusers.utils import load_image, export_to_video + + >>> model_id = "hunyuanvideo-community/HunyuanVideo" + >>> transformer_model_id = "Skywork/SkyReels-V1-Hunyuan-I2V" + >>> transformer = HunyuanVideoTransformer3DModel.from_pretrained( + ... transformer_model_id, torch_dtype=torch.bfloat16 + ... ) + >>> pipe = HunyuanSkyreelsImageToVideoPipeline.from_pretrained( + ... model_id, transformer=transformer, torch_dtype=torch.float16 + ... ) + >>> pipe.vae.enable_tiling() + >>> pipe.to("cuda") + + >>> prompt = "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot." + >>> negative_prompt = "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion" + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + ... ) + + >>> output = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... num_inference_steps=30, + ... true_cfg_scale=6.0, + ... guidance_scale=1.0, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=15) + ``` +""" + + +DEFAULT_PROMPT_TEMPLATE = { + "template": ( + "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: " + "1. The main content and theme of the video." + "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects." + "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects." + "4. background environment, light, style and atmosphere." + "5. camera angles, movements, and transitions used in the video:<|eot_id|>" + "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>" + ), + "crop_start": 95, +} + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class HunyuanSkyreelsImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin): + r""" + Pipeline for image-to-video generation using HunyuanVideo. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`LlamaModel`]): + [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + tokenizer (`LlamaTokenizer`): + Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + transformer ([`HunyuanVideoTransformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLHunyuanVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder_2 ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + text_encoder: LlamaModel, + tokenizer: LlamaTokenizerFast, + transformer: HunyuanVideoTransformer3DModel, + vae: AutoencoderKLHunyuanVideo, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + ) + + self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8 + self.vae_scaling_factor = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.476986 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline._get_llama_prompt_embeds + def _get_llama_prompt_embeds( + self, + prompt: str | list[str], + prompt_template: dict[str, Any], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + num_hidden_layers_to_skip: int = 2, + ) -> tuple[torch.Tensor, torch.Tensor]: + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + prompt = [prompt_template["template"].format(p) for p in prompt] + + crop_start = prompt_template.get("crop_start", None) + if crop_start is None: + prompt_template_input = self.tokenizer( + prompt_template["template"], + padding="max_length", + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=False, + ) + crop_start = prompt_template_input["input_ids"].shape[-1] + # Remove <|eot_id|> token and placeholder {} + crop_start -= 2 + + max_sequence_length += crop_start + text_inputs = self.tokenizer( + prompt, + max_length=max_sequence_length, + padding="max_length", + truncation=True, + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=True, + ) + text_input_ids = text_inputs.input_ids.to(device=device) + prompt_attention_mask = text_inputs.attention_mask.to(device=device) + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-(num_hidden_layers_to_skip + 1)] + prompt_embeds = prompt_embeds.to(dtype=dtype) + + if crop_start is not None and crop_start > 0: + prompt_embeds = prompt_embeds[:, crop_start:] + prompt_attention_mask = prompt_attention_mask[:, crop_start:] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.repeat(1, num_videos_per_prompt) + prompt_attention_mask = prompt_attention_mask.view(batch_size * num_videos_per_prompt, seq_len) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 77, + ) -> torch.Tensor: + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] = None, + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + ): + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds( + prompt, + prompt_template, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=max_sequence_length, + ) + + if pooled_prompt_embeds is None: + if prompt_2 is None: + prompt_2 = prompt + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=77, + ) + + return prompt_embeds, pooled_prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + prompt_template=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_template is not None: + if not isinstance(prompt_template, dict): + raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}") + if "template" not in prompt_template: + raise ValueError( + f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}" + ) + + def prepare_latents( + self, + image: torch.Tensor, + batch_size: int, + num_channels_latents: int = 32, + height: int = 544, + width: int = 960, + num_frames: int = 97, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + image = image.unsqueeze(2) # [B, C, 1, H, W] + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i].unsqueeze(0)), generator[i]) for i in range(batch_size) + ] + else: + image_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator) for img in image] + + image_latents = torch.cat(image_latents, dim=0).to(dtype) * self.vae_scaling_factor + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height, latent_width = height // self.vae_scale_factor_spatial, width // self.vae_scale_factor_spatial + shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width) + padding_shape = (batch_size, num_channels_latents, num_latent_frames - 1, latent_height, latent_width) + + latents_padding = torch.zeros(padding_shape, dtype=dtype, device=device) + image_latents = torch.cat([image_latents, latents_padding], dim=2) + + if latents is None: + latents = randn_tensor(shape, generator=generator, dtype=dtype, device=device) + else: + latents = latents.to(dtype=dtype, device=device) + + return latents, image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + prompt_2: str | list[str] = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] = None, + height: int = 544, + width: int = 960, + num_frames: int = 97, + num_inference_steps: int = 50, + sigmas: list[float] = None, + true_cfg_scale: float = 6.0, + guidance_scale: float = 1.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + max_sequence_length: int = 256, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + height (`int`, defaults to `720`): + The height in pixels of the generated image. + width (`int`, defaults to `1280`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `129`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + guidance_scale (`float`, defaults to `6.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. Note that the only available + HunyuanVideo model is CFG-distilled, which means that traditional guidance between unconditional and + conditional latent is not applied. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. Required when `negative_prompt_embeds` is passed directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + prompt_template (`dict`, *optional*): + Template used to format the prompt before encoding. Defaults to the model's default template. + max_sequence_length (`int`, *optional*, defaults to 256): + Maximum sequence length to use for the prompt encoder. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~HunyuanVideoPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + prompt_embeds, + callback_on_step_end_tensor_inputs, + prompt_template, + ) + + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + transformer_dtype = self.transformer.dtype + prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + ) + prompt_embeds = prompt_embeds.to(transformer_dtype) + prompt_attention_mask = prompt_attention_mask.to(transformer_dtype) + pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype) + + if do_true_cfg: + negative_prompt_embeds, negative_pooled_prompt_embeds, negative_prompt_attention_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + prompt_attention_mask=negative_prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + ) + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + negative_prompt_attention_mask = negative_prompt_attention_mask.to(transformer_dtype) + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, sigmas=sigmas + ) + + # 5. Prepare latent variables + vae_dtype = self.vae.dtype + image = self.video_processor.preprocess(image, height=height, width=width).to(device, vae_dtype) + num_channels_latents = self.transformer.config.in_channels // 2 + latents, image_latents = self.prepare_latents( + image, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + ) + latent_image_input = image_latents.to(transformer_dtype) + + # 6. Prepare guidance condition + guidance = torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0 + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + latent_model_input = torch.cat([latent_model_input, latent_image_input], dim=1) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + pooled_projections=pooled_prompt_embeds, + guidance=guidance, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_attention_mask=negative_prompt_attention_mask, + pooled_projections=negative_pooled_prompt_embeds, + guidance=guidance, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HunyuanVideoPipelineOutput(frames=video) diff --git a/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py new file mode 100644 index 0000000000000000000000000000000000000000..5e785c38891c1e325271285d75d85800991486fc --- /dev/null +++ b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py @@ -0,0 +1,735 @@ +# Copyright 2025 The HunyuanVideo Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import HunyuanVideoLoraLoaderMixin +from ...models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HunyuanVideoPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel + >>> from diffusers.utils import export_to_video + + >>> model_id = "hunyuanvideo-community/HunyuanVideo" + >>> transformer = HunyuanVideoTransformer3DModel.from_pretrained( + ... model_id, subfolder="transformer", torch_dtype=torch.bfloat16 + ... ) + >>> pipe = HunyuanVideoPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.float16) + >>> pipe.vae.enable_tiling() + >>> pipe.to("cuda") + + >>> output = pipe( + ... prompt="A cat walks on the grass, realistic", + ... height=320, + ... width=512, + ... num_frames=61, + ... num_inference_steps=30, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=15) + ``` +""" + + +DEFAULT_PROMPT_TEMPLATE = { + "template": ( + "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: " + "1. The main content and theme of the video." + "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects." + "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects." + "4. background environment, light, style and atmosphere." + "5. camera angles, movements, and transitions used in the video:<|eot_id|>" + "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>" + ), + "crop_start": 95, +} + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class HunyuanVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin): + r""" + Pipeline for text-to-video generation using HunyuanVideo. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`LlamaModel`]): + [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + tokenizer (`LlamaTokenizer`): + Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + transformer ([`HunyuanVideoTransformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLHunyuanVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder_2 ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + text_encoder: LlamaModel, + tokenizer: LlamaTokenizerFast, + transformer: HunyuanVideoTransformer3DModel, + vae: AutoencoderKLHunyuanVideo, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + ) + + self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + def _get_llama_prompt_embeds( + self, + prompt: str | list[str], + prompt_template: dict[str, Any], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + num_hidden_layers_to_skip: int = 2, + ) -> tuple[torch.Tensor, torch.Tensor]: + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + prompt = [prompt_template["template"].format(p) for p in prompt] + + crop_start = prompt_template.get("crop_start", None) + if crop_start is None: + prompt_template_input = self.tokenizer( + prompt_template["template"], + padding="max_length", + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=False, + ) + crop_start = prompt_template_input["input_ids"].shape[-1] + # Remove <|eot_id|> token and placeholder {} + crop_start -= 2 + + max_sequence_length += crop_start + text_inputs = self.tokenizer( + prompt, + max_length=max_sequence_length, + padding="max_length", + truncation=True, + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=True, + ) + text_input_ids = text_inputs.input_ids.to(device=device) + prompt_attention_mask = text_inputs.attention_mask.to(device=device) + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-(num_hidden_layers_to_skip + 1)] + prompt_embeds = prompt_embeds.to(dtype=dtype) + + if crop_start is not None and crop_start > 0: + prompt_embeds = prompt_embeds[:, crop_start:] + prompt_attention_mask = prompt_attention_mask[:, crop_start:] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.repeat(1, num_videos_per_prompt) + prompt_attention_mask = prompt_attention_mask.view(batch_size * num_videos_per_prompt, seq_len) + + return prompt_embeds, prompt_attention_mask + + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 77, + ) -> torch.Tensor: + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, -1) + + return prompt_embeds + + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] = None, + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + ): + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds( + prompt, + prompt_template, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=max_sequence_length, + ) + + if pooled_prompt_embeds is None: + if prompt_2 is None: + prompt_2 = prompt + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=77, + ) + + return prompt_embeds, pooled_prompt_embeds, prompt_attention_mask + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + prompt_template=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_template is not None: + if not isinstance(prompt_template, dict): + raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}") + if "template" not in prompt_template: + raise ValueError( + f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}" + ) + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 32, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] = None, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + num_inference_steps: int = 50, + sigmas: list[float] = None, + true_cfg_scale: float = 1.0, + guidance_scale: float = 6.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + max_sequence_length: int = 256, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + height (`int`, defaults to `720`): + The height in pixels of the generated image. + width (`int`, defaults to `1280`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `129`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + True classifier-free guidance (guidance scale) is enabled when `true_cfg_scale` > 1 and + `negative_prompt` is provided. + guidance_scale (`float`, defaults to `6.0`): + Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with `prompt` at the expense of lower image quality. + + Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. Required when `negative_prompt_embeds` is passed directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + prompt_template (`dict`, *optional*): + Template used to format the prompt before encoding. Defaults to the model's default template. + max_sequence_length (`int`, *optional*, defaults to 256): + Maximum sequence length to use for the prompt encoder. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~HunyuanVideoPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + prompt_embeds, + callback_on_step_end_tensor_inputs, + prompt_template, + ) + + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + transformer_dtype = self.transformer.dtype + prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + ) + prompt_embeds = prompt_embeds.to(transformer_dtype) + prompt_attention_mask = prompt_attention_mask.to(transformer_dtype) + pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype) + + if do_true_cfg: + negative_prompt_embeds, negative_pooled_prompt_embeds, negative_prompt_attention_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + prompt_attention_mask=negative_prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + ) + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + negative_prompt_attention_mask = negative_prompt_attention_mask.to(transformer_dtype) + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, sigmas=sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + ) + + # 6. Prepare guidance condition + guidance = torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0 + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + pooled_projections=pooled_prompt_embeds, + guidance=guidance, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_attention_mask=negative_prompt_attention_mask, + pooled_projections=negative_pooled_prompt_embeds, + guidance=guidance, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HunyuanVideoPipelineOutput(frames=video) diff --git a/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_framepack.py b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_framepack.py new file mode 100644 index 0000000000000000000000000000000000000000..349481492ac05e0b21ece0b19697cfdbb98bc57c --- /dev/null +++ b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_framepack.py @@ -0,0 +1,1090 @@ +# Copyright 2025 The Framepack Team, The HunyuanVideo Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from enum import Enum +from typing import Any, Callable + +import numpy as np +import torch +from transformers import ( + CLIPTextModel, + CLIPTokenizer, + LlamaModel, + LlamaTokenizerFast, + SiglipImageProcessor, + SiglipVisionModel, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import HunyuanVideoLoraLoaderMixin +from ...models import AutoencoderKLHunyuanVideo, HunyuanVideoFramepackTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HunyuanVideoFramepackPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# TODO(yiyi): We can pack the checkpoints nicely with modular loader +EXAMPLE_DOC_STRING = """ + Examples: + ##### Image-to-Video + + ```python + >>> import torch + >>> from diffusers import HunyuanVideoFramepackPipeline, HunyuanVideoFramepackTransformer3DModel + >>> from diffusers.utils import export_to_video, load_image + >>> from transformers import SiglipImageProcessor, SiglipVisionModel + + >>> transformer = HunyuanVideoFramepackTransformer3DModel.from_pretrained( + ... "lllyasviel/FramePackI2V_HY", torch_dtype=torch.bfloat16 + ... ) + >>> feature_extractor = SiglipImageProcessor.from_pretrained( + ... "lllyasviel/flux_redux_bfl", subfolder="feature_extractor" + ... ) + >>> image_encoder = SiglipVisionModel.from_pretrained( + ... "lllyasviel/flux_redux_bfl", subfolder="image_encoder", torch_dtype=torch.float16 + ... ) + >>> pipe = HunyuanVideoFramepackPipeline.from_pretrained( + ... "hunyuanvideo-community/HunyuanVideo", + ... transformer=transformer, + ... feature_extractor=feature_extractor, + ... image_encoder=image_encoder, + ... torch_dtype=torch.float16, + ... ) + >>> pipe.vae.enable_tiling() + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/penguin.png" + ... ) + >>> output = pipe( + ... image=image, + ... prompt="A penguin dancing in the snow", + ... height=832, + ... width=480, + ... num_frames=91, + ... num_inference_steps=30, + ... guidance_scale=9.0, + ... generator=torch.Generator().manual_seed(0), + ... sampling_type="inverted_anti_drifting", + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=30) + ``` + + ##### First and Last Image-to-Video + + ```python + >>> import torch + >>> from diffusers import HunyuanVideoFramepackPipeline, HunyuanVideoFramepackTransformer3DModel + >>> from diffusers.utils import export_to_video, load_image + >>> from transformers import SiglipImageProcessor, SiglipVisionModel + + >>> transformer = HunyuanVideoFramepackTransformer3DModel.from_pretrained( + ... "lllyasviel/FramePackI2V_HY", torch_dtype=torch.bfloat16 + ... ) + >>> feature_extractor = SiglipImageProcessor.from_pretrained( + ... "lllyasviel/flux_redux_bfl", subfolder="feature_extractor" + ... ) + >>> image_encoder = SiglipVisionModel.from_pretrained( + ... "lllyasviel/flux_redux_bfl", subfolder="image_encoder", torch_dtype=torch.float16 + ... ) + >>> pipe = HunyuanVideoFramepackPipeline.from_pretrained( + ... "hunyuanvideo-community/HunyuanVideo", + ... transformer=transformer, + ... feature_extractor=feature_extractor, + ... image_encoder=image_encoder, + ... torch_dtype=torch.float16, + ... ) + >>> pipe.to("cuda") + + >>> prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective." + >>> first_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png" + ... ) + >>> last_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png" + ... ) + >>> output = pipe( + ... image=first_image, + ... last_image=last_image, + ... prompt=prompt, + ... height=512, + ... width=512, + ... num_frames=91, + ... num_inference_steps=30, + ... guidance_scale=9.0, + ... generator=torch.Generator().manual_seed(0), + ... sampling_type="inverted_anti_drifting", + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=30) + ``` +""" + + +DEFAULT_PROMPT_TEMPLATE = { + "template": ( + "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: " + "1. The main content and theme of the video." + "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects." + "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects." + "4. background environment, light, style and atmosphere." + "5. camera angles, movements, and transitions used in the video:<|eot_id|>" + "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>" + ), + "crop_start": 95, +} + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FramepackSamplingType(str, Enum): + VANILLA = "vanilla" + INVERTED_ANTI_DRIFTING = "inverted_anti_drifting" + + +class HunyuanVideoFramepackPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin): + r""" + Pipeline for text-to-video generation using HunyuanVideo. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`LlamaModel`]): + [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + tokenizer (`LlamaTokenizer`): + Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + transformer ([`HunyuanVideoTransformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLHunyuanVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder_2 ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + text_encoder: LlamaModel, + tokenizer: LlamaTokenizerFast, + transformer: HunyuanVideoFramepackTransformer3DModel, + vae: AutoencoderKLHunyuanVideo, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + image_encoder: SiglipVisionModel, + feature_extractor: SiglipImageProcessor, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + + self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline._get_llama_prompt_embeds + def _get_llama_prompt_embeds( + self, + prompt: str | list[str], + prompt_template: dict[str, Any], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + num_hidden_layers_to_skip: int = 2, + ) -> tuple[torch.Tensor, torch.Tensor]: + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + prompt = [prompt_template["template"].format(p) for p in prompt] + + crop_start = prompt_template.get("crop_start", None) + if crop_start is None: + prompt_template_input = self.tokenizer( + prompt_template["template"], + padding="max_length", + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=False, + ) + crop_start = prompt_template_input["input_ids"].shape[-1] + # Remove <|eot_id|> token and placeholder {} + crop_start -= 2 + + max_sequence_length += crop_start + text_inputs = self.tokenizer( + prompt, + max_length=max_sequence_length, + padding="max_length", + truncation=True, + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=True, + ) + text_input_ids = text_inputs.input_ids.to(device=device) + prompt_attention_mask = text_inputs.attention_mask.to(device=device) + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-(num_hidden_layers_to_skip + 1)] + prompt_embeds = prompt_embeds.to(dtype=dtype) + + if crop_start is not None and crop_start > 0: + prompt_embeds = prompt_embeds[:, crop_start:] + prompt_attention_mask = prompt_attention_mask[:, crop_start:] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.repeat(1, num_videos_per_prompt) + prompt_attention_mask = prompt_attention_mask.view(batch_size * num_videos_per_prompt, seq_len) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 77, + ) -> torch.Tensor: + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str] = None, + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + ): + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds( + prompt, + prompt_template, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=max_sequence_length, + ) + + if pooled_prompt_embeds is None: + if prompt_2 is None: + prompt_2 = prompt + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=77, + ) + + return prompt_embeds, pooled_prompt_embeds, prompt_attention_mask + + def encode_image(self, image: torch.Tensor, device: torch.device | None = None, dtype: torch.dtype | None = None): + device = device or self._execution_device + image = (image + 1) / 2.0 # [-1, 1] -> [0, 1] + image = self.feature_extractor(images=image, return_tensors="pt", do_rescale=False).to( + device=device, dtype=self.image_encoder.dtype + ) + image_embeds = self.image_encoder(**image).last_hidden_state + return image_embeds.to(dtype=dtype) + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + prompt_template=None, + image=None, + image_latents=None, + last_image=None, + last_image_latents=None, + sampling_type=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_template is not None: + if not isinstance(prompt_template, dict): + raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}") + if "template" not in prompt_template: + raise ValueError( + f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}" + ) + + sampling_types = [x.value for x in FramepackSamplingType.__members__.values()] + if sampling_type not in sampling_types: + raise ValueError(f"`sampling_type` has to be one of '{sampling_types}' but is '{sampling_type}'") + + if image is not None and image_latents is not None: + raise ValueError("Only one of `image` or `image_latents` can be passed.") + if last_image is not None and last_image_latents is not None: + raise ValueError("Only one of `last_image` or `last_image_latents` can be passed.") + if sampling_type != FramepackSamplingType.INVERTED_ANTI_DRIFTING and ( + last_image is not None or last_image_latents is not None + ): + raise ValueError( + 'Only `"inverted_anti_drifting"` inference type supports `last_image` or `last_image_latents`.' + ) + + def prepare_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 16, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + shape = ( + batch_size, + num_channels_latents, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + def prepare_image_latents( + self, + image: torch.Tensor, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + device = device or self._execution_device + if latents is None: + image = image.unsqueeze(2).to(device=device, dtype=self.vae.dtype) + latents = self.vae.encode(image).latent_dist.sample(generator=generator) + latents = latents * self.vae.config.scaling_factor + return latents.to(device=device, dtype=dtype) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + last_image: PipelineImageInput | None = None, + prompt: str | list[str] = None, + prompt_2: str | list[str] = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] = None, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + latent_window_size: int = 9, + num_inference_steps: int = 50, + sigmas: list[float] = None, + true_cfg_scale: float = 1.0, + guidance_scale: float = 6.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + image_latents: torch.Tensor | None = None, + last_image_latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + max_sequence_length: int = 256, + sampling_type: FramepackSamplingType = FramepackSamplingType.INVERTED_ANTI_DRIFTING, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`): + The image to be used as the starting point for the video generation. + last_image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`, *optional*): + The optional last image to be used as the ending point for the video generation. This is useful for + generating transitions between two images. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + height (`int`, defaults to `720`): + The height in pixels of the generated image. + width (`int`, defaults to `1280`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `129`): + The number of frames in the generated video. + latent_window_size (`int`, defaults to `9`): + Number of latent frames produced per Framepack sampling window. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + guidance_scale (`float`, defaults to `6.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. Note that the only available + HunyuanVideo model is CFG-distilled, which means that traditional guidance between unconditional and + conditional latent is not applied. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + image_latents (`torch.Tensor`, *optional*): + Pre-encoded image latents. If not provided, the image will be encoded using the VAE. + last_image_latents (`torch.Tensor`, *optional*): + Pre-encoded last image latents. If not provided, the last image will be encoded using the VAE. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. Required when `negative_prompt_embeds` is passed directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HunyuanVideoFramepackPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + prompt_template (`dict`, *optional*): + Template used to format the prompt before encoding. Defaults to the model's default template. + max_sequence_length (`int`, *optional*, defaults to 256): + Maximum sequence length to use for the prompt encoder. + sampling_type (`FramepackSamplingType`, *optional*): + The Framepack sampling strategy to use when iterating over latent windows. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~HunyuanVideoFramepackPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HunyuanVideoFramepackPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images and the second element is a list + of `bool`s indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) + content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + prompt_embeds, + callback_on_step_end_tensor_inputs, + prompt_template, + image, + image_latents, + last_image, + last_image_latents, + sampling_type, + ) + + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + transformer_dtype = self.transformer.dtype + vae_dtype = self.vae.dtype + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + transformer_dtype = self.transformer.dtype + prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + ) + prompt_embeds = prompt_embeds.to(transformer_dtype) + prompt_attention_mask = prompt_attention_mask.to(transformer_dtype) + pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype) + + if do_true_cfg: + negative_prompt_embeds, negative_pooled_prompt_embeds, negative_prompt_attention_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + prompt_attention_mask=negative_prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + ) + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + negative_prompt_attention_mask = negative_prompt_attention_mask.to(transformer_dtype) + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(transformer_dtype) + + # 4. Prepare image + image = self.video_processor.preprocess(image, height, width) + image_embeds = self.encode_image(image, device=device).to(transformer_dtype) + if last_image is not None: + # Credits: https://github.com/lllyasviel/FramePack/pull/167 + # Users can modify the weighting strategy applied here + last_image = self.video_processor.preprocess(last_image, height, width) + last_image_embeds = self.encode_image(last_image, device=device).to(transformer_dtype) + last_image_embeds = (image_embeds + last_image_embeds) / 2 + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + window_num_frames = (latent_window_size - 1) * self.vae_scale_factor_temporal + 1 + num_latent_sections = max(1, (num_frames + window_num_frames - 1) // window_num_frames) + history_video = None + total_generated_latent_frames = 0 + + image_latents = self.prepare_image_latents( + image, dtype=torch.float32, device=device, generator=generator, latents=image_latents + ) + if last_image is not None: + last_image_latents = self.prepare_image_latents( + last_image, dtype=torch.float32, device=device, generator=generator + ) + + # Specific to the released checkpoints: + # - https://huggingface.co/lllyasviel/FramePackI2V_HY + # - https://huggingface.co/lllyasviel/FramePack_F1_I2V_HY_20250503 + # TODO: find a more generic way in future if there are more checkpoints + if sampling_type == FramepackSamplingType.INVERTED_ANTI_DRIFTING: + history_sizes = [1, 2, 16] + history_latents = torch.zeros( + batch_size, + num_channels_latents, + sum(history_sizes), + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + device=device, + dtype=torch.float32, + ) + + elif sampling_type == FramepackSamplingType.VANILLA: + history_sizes = [16, 2, 1] + history_latents = torch.zeros( + batch_size, + num_channels_latents, + sum(history_sizes), + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + device=device, + dtype=torch.float32, + ) + history_latents = torch.cat([history_latents, image_latents], dim=2) + total_generated_latent_frames += 1 + + else: + assert False + + # 6. Prepare guidance condition + guidance = torch.tensor([guidance_scale] * batch_size, dtype=transformer_dtype, device=device) * 1000.0 + + # 7. Denoising loop + for k in range(num_latent_sections): + if sampling_type == FramepackSamplingType.INVERTED_ANTI_DRIFTING: + latent_paddings = list(reversed(range(num_latent_sections))) + if num_latent_sections > 4: + latent_paddings = [3] + [2] * (num_latent_sections - 3) + [1, 0] + + is_first_section = k == 0 + is_last_section = k == num_latent_sections - 1 + latent_padding_size = latent_paddings[k] * latent_window_size + + indices = torch.arange(0, sum([1, latent_padding_size, latent_window_size, *history_sizes])) + ( + indices_prefix, + indices_padding, + indices_latents, + indices_latents_history_1x, + indices_latents_history_2x, + indices_latents_history_4x, + ) = indices.split([1, latent_padding_size, latent_window_size, *history_sizes], dim=0) + # Inverted anti-drifting sampling: Figure 2(c) in the paper + indices_clean_latents = torch.cat([indices_prefix, indices_latents_history_1x], dim=0) + + latents_prefix = image_latents + latents_history_1x, latents_history_2x, latents_history_4x = history_latents[ + :, :, : sum(history_sizes) + ].split(history_sizes, dim=2) + if last_image is not None and is_first_section: + latents_history_1x = last_image_latents + latents_clean = torch.cat([latents_prefix, latents_history_1x], dim=2) + + elif sampling_type == FramepackSamplingType.VANILLA: + indices = torch.arange(0, sum([1, *history_sizes, latent_window_size])) + ( + indices_prefix, + indices_latents_history_4x, + indices_latents_history_2x, + indices_latents_history_1x, + indices_latents, + ) = indices.split([1, *history_sizes, latent_window_size], dim=0) + indices_clean_latents = torch.cat([indices_prefix, indices_latents_history_1x], dim=0) + + latents_prefix = image_latents + latents_history_4x, latents_history_2x, latents_history_1x = history_latents[ + :, :, -sum(history_sizes) : + ].split(history_sizes, dim=2) + latents_clean = torch.cat([latents_prefix, latents_history_1x], dim=2) + + else: + assert False + + latents = self.prepare_latents( + batch_size, + num_channels_latents, + height, + width, + window_num_frames, + dtype=torch.float32, + device=device, + generator=generator, + latents=None, + ) + + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + image_seq_len = ( + latents.shape[2] * latents.shape[3] * latents.shape[4] / self.transformer.config.patch_size**2 + ) + exp_max = 7.0 + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + mu = min(mu, math.log(exp_max)) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, sigmas=sigmas, mu=mu + ) + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + timestep = t.expand(latents.shape[0]) + + noise_pred = self.transformer( + hidden_states=latents.to(transformer_dtype), + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + pooled_projections=pooled_prompt_embeds, + image_embeds=image_embeds, + indices_latents=indices_latents, + guidance=guidance, + latents_clean=latents_clean.to(transformer_dtype), + indices_latents_clean=indices_clean_latents, + latents_history_2x=latents_history_2x.to(transformer_dtype), + indices_latents_history_2x=indices_latents_history_2x, + latents_history_4x=latents_history_4x.to(transformer_dtype), + indices_latents_history_4x=indices_latents_history_4x, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + neg_noise_pred = self.transformer( + hidden_states=latents.to(transformer_dtype), + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_attention_mask=negative_prompt_attention_mask, + pooled_projections=negative_pooled_prompt_embeds, + image_embeds=image_embeds, + indices_latents=indices_latents, + guidance=guidance, + latents_clean=latents_clean.to(transformer_dtype), + indices_latents_clean=indices_clean_latents, + latents_history_2x=latents_history_2x.to(transformer_dtype), + indices_latents_history_2x=indices_latents_history_2x, + latents_history_4x=latents_history_4x.to(transformer_dtype), + indices_latents_history_4x=indices_latents_history_4x, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred.float(), t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if sampling_type == FramepackSamplingType.INVERTED_ANTI_DRIFTING: + if is_last_section: + latents = torch.cat([image_latents, latents], dim=2) + total_generated_latent_frames += latents.shape[2] + history_latents = torch.cat([latents, history_latents], dim=2) + real_history_latents = history_latents[:, :, :total_generated_latent_frames] + section_latent_frames = ( + (latent_window_size * 2 + 1) if is_last_section else (latent_window_size * 2) + ) + index_slice = (slice(None), slice(None), slice(0, section_latent_frames)) + + elif sampling_type == FramepackSamplingType.VANILLA: + total_generated_latent_frames += latents.shape[2] + history_latents = torch.cat([history_latents, latents], dim=2) + real_history_latents = history_latents[:, :, -total_generated_latent_frames:] + section_latent_frames = latent_window_size * 2 + index_slice = (slice(None), slice(None), slice(-section_latent_frames, None)) + + else: + assert False + + if history_video is None: + if not output_type == "latent": + current_latents = real_history_latents.to(vae_dtype) / self.vae.config.scaling_factor + history_video = self.vae.decode(current_latents, return_dict=False)[0] + else: + history_video = [real_history_latents] + else: + if not output_type == "latent": + overlapped_frames = (latent_window_size - 1) * self.vae_scale_factor_temporal + 1 + current_latents = ( + real_history_latents[index_slice].to(vae_dtype) / self.vae.config.scaling_factor + ) + current_video = self.vae.decode(current_latents, return_dict=False)[0] + + if sampling_type == FramepackSamplingType.INVERTED_ANTI_DRIFTING: + history_video = self._soft_append(current_video, history_video, overlapped_frames) + elif sampling_type == FramepackSamplingType.VANILLA: + history_video = self._soft_append(history_video, current_video, overlapped_frames) + else: + assert False + else: + history_video.append(real_history_latents) + + self._current_timestep = None + + if not output_type == "latent": + generated_frames = history_video.size(2) + generated_frames = ( + generated_frames - 1 + ) // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + history_video = history_video[:, :, :generated_frames] + video = self.video_processor.postprocess_video(history_video, output_type=output_type) + else: + video = history_video + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HunyuanVideoFramepackPipelineOutput(frames=video) + + def _soft_append(self, history: torch.Tensor, current: torch.Tensor, overlap: int = 0): + if overlap <= 0: + return torch.cat([history, current], dim=2) + + assert history.shape[2] >= overlap, f"Current length ({history.shape[2]}) must be >= overlap ({overlap})" + assert current.shape[2] >= overlap, f"History length ({current.shape[2]}) must be >= overlap ({overlap})" + + weights = torch.linspace(1, 0, overlap, dtype=history.dtype, device=history.device).view(1, 1, -1, 1, 1) + blended = weights * history[:, :, -overlap:] + (1 - weights) * current[:, :, :overlap] + output = torch.cat([history[:, :, :-overlap], blended, current[:, :, overlap:]], dim=2) + + return output.to(history) diff --git a/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_image2video.py b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_image2video.py new file mode 100644 index 0000000000000000000000000000000000000000..13eb3538600182a1c0732bf46fd20a88a009751c --- /dev/null +++ b/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_image2video.py @@ -0,0 +1,964 @@ +# Copyright 2025 The HunyuanVideo Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + LlamaTokenizerFast, + LlavaForConditionalGeneration, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import HunyuanVideoLoraLoaderMixin +from ...models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import HunyuanVideoPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import HunyuanVideoImageToVideoPipeline, HunyuanVideoTransformer3DModel + >>> from diffusers.utils import load_image, export_to_video + + >>> # Available checkpoints: hunyuanvideo-community/HunyuanVideo-I2V, hunyuanvideo-community/HunyuanVideo-I2V-33ch + >>> model_id = "hunyuanvideo-community/HunyuanVideo-I2V" + >>> transformer = HunyuanVideoTransformer3DModel.from_pretrained( + ... model_id, subfolder="transformer", torch_dtype=torch.bfloat16 + ... ) + >>> pipe = HunyuanVideoImageToVideoPipeline.from_pretrained( + ... model_id, transformer=transformer, torch_dtype=torch.float16 + ... ) + >>> pipe.vae.enable_tiling() + >>> pipe.to("cuda") + + >>> prompt = "A man with short gray hair plays a red electric guitar." + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png" + ... ) + + >>> # If using hunyuanvideo-community/HunyuanVideo-I2V + >>> output = pipe(image=image, prompt=prompt, guidance_scale=6.0).frames[0] + + >>> # If using hunyuanvideo-community/HunyuanVideo-I2V-33ch + >>> output = pipe(image=image, prompt=prompt, guidance_scale=1.0, true_cfg_scale=1.0).frames[0] + + >>> export_to_video(output, "output.mp4", fps=15) + ``` +""" + + +DEFAULT_PROMPT_TEMPLATE = { + "template": ( + "<|start_header_id|>system<|end_header_id|>\n\n\nDescribe the video by detailing the following aspects according to the reference image: " + "1. The main content and theme of the video." + "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects." + "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects." + "4. background environment, light, style and atmosphere." + "5. camera angles, movements, and transitions used in the video:<|eot_id|>\n\n" + "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>" + "<|start_header_id|>assistant<|end_header_id|>\n\n" + ), + "crop_start": 103, + "image_emb_start": 5, + "image_emb_end": 581, + "image_emb_len": 576, +} + + +def _expand_input_ids_with_image_tokens( + text_input_ids, + prompt_attention_mask, + max_sequence_length, + image_token_index, + image_emb_len, + image_emb_start, + image_emb_end, + pad_token_id, +): + special_image_token_mask = text_input_ids == image_token_index + num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1) + batch_indices, non_image_indices = torch.where(text_input_ids != image_token_index) + + max_expanded_length = max_sequence_length + (num_special_image_tokens.max() * (image_emb_len - 1)) + new_token_positions = torch.cumsum((special_image_token_mask * (image_emb_len - 1) + 1), -1) - 1 + text_to_overwrite = new_token_positions[batch_indices, non_image_indices] + + expanded_input_ids = torch.full( + (text_input_ids.shape[0], max_expanded_length), + pad_token_id, + dtype=text_input_ids.dtype, + device=text_input_ids.device, + ) + expanded_input_ids[batch_indices, text_to_overwrite] = text_input_ids[batch_indices, non_image_indices] + expanded_input_ids[batch_indices, image_emb_start:image_emb_end] = image_token_index + + expanded_attention_mask = torch.zeros( + (text_input_ids.shape[0], max_expanded_length), + dtype=prompt_attention_mask.dtype, + device=prompt_attention_mask.device, + ) + attn_batch_indices, attention_indices = torch.where(expanded_input_ids != pad_token_id) + expanded_attention_mask[attn_batch_indices, attention_indices] = 1.0 + expanded_attention_mask = expanded_attention_mask.to(prompt_attention_mask.dtype) + position_ids = (expanded_attention_mask.cumsum(-1) - 1).masked_fill_((expanded_attention_mask == 0), 1) + + return { + "input_ids": expanded_input_ids, + "attention_mask": expanded_attention_mask, + "position_ids": position_ids, + } + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin): + r""" + Pipeline for image-to-video generation using HunyuanVideo. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`LlavaForConditionalGeneration`]): + [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + tokenizer (`LlamaTokenizer`): + Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). + transformer ([`HunyuanVideoTransformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLHunyuanVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder_2 ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + text_encoder: LlavaForConditionalGeneration, + tokenizer: LlamaTokenizerFast, + transformer: HunyuanVideoTransformer3DModel, + vae: AutoencoderKLHunyuanVideo, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + image_processor=image_processor, + ) + + self.vae_scaling_factor = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.476986 + self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + def _get_llama_prompt_embeds( + self, + image: torch.Tensor, + prompt: str | list[str], + prompt_template: dict[str, Any], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + num_hidden_layers_to_skip: int = 2, + image_embed_interleave: int = 2, + ) -> tuple[torch.Tensor, torch.Tensor]: + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_template["template"].format(p) for p in prompt] + + crop_start = prompt_template.get("crop_start", None) + + image_emb_len = prompt_template.get("image_emb_len", 576) + image_emb_start = prompt_template.get("image_emb_start", 5) + image_emb_end = prompt_template.get("image_emb_end", 581) + + if crop_start is None: + prompt_template_input = self.tokenizer( + prompt_template["template"], + padding="max_length", + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=False, + ) + crop_start = prompt_template_input["input_ids"].shape[-1] + # Remove <|start_header_id|>, <|end_header_id|>, assistant, <|eot_id|>, and placeholder {} + crop_start -= 5 + + max_sequence_length += crop_start + text_inputs = self.tokenizer( + prompt, + max_length=max_sequence_length, + padding="max_length", + truncation=True, + return_tensors="pt", + return_length=False, + return_overflowing_tokens=False, + return_attention_mask=True, + ) + text_input_ids = text_inputs.input_ids.to(device=device) + prompt_attention_mask = text_inputs.attention_mask.to(device=device) + + image_embeds = self.image_processor(image, return_tensors="pt").pixel_values.to(device) + + image_token_index = self.text_encoder.config.image_token_index + pad_token_id = self.text_encoder.config.pad_token_id + expanded_inputs = _expand_input_ids_with_image_tokens( + text_input_ids, + prompt_attention_mask, + max_sequence_length, + image_token_index, + image_emb_len, + image_emb_start, + image_emb_end, + pad_token_id, + ) + prompt_embeds = self.text_encoder( + **expanded_inputs, + pixel_values=image_embeds, + output_hidden_states=True, + ).hidden_states[-(num_hidden_layers_to_skip + 1)] + prompt_embeds = prompt_embeds.to(dtype=dtype) + + if crop_start is not None and crop_start > 0: + text_crop_start = crop_start - 1 + image_emb_len + + # Find assistant section marker using <|end_header_id|> token (works across all transformers versions) + end_header_token_id = self.tokenizer.convert_tokens_to_ids("<|end_header_id|>") + batch_indices, end_header_indices = torch.where(text_input_ids == end_header_token_id) + + # Expected: 3 <|end_header_id|> per prompt (system, user, assistant) + # If truncated (only 2 found for batch_size=1), add text length as fallback position + if end_header_indices.shape[0] == 2: + # in case the prompt is too long + end_header_indices = torch.cat( + ( + end_header_indices, + torch.tensor([text_input_ids.shape[-1] - 1], device=end_header_indices.device), + ) + ) + batch_indices = torch.cat((batch_indices, torch.tensor([0], device=batch_indices.device))) + + # Get the last <|end_header_id|> position per batch, then +1 to get the position after it + assistant_start_indices = end_header_indices.reshape(text_input_ids.shape[0], -1)[:, -1] + 1 + batch_indices = batch_indices.reshape(text_input_ids.shape[0], -1)[:, -1] + assistant_crop_start = assistant_start_indices - 1 + image_emb_len - 4 + assistant_crop_end = assistant_start_indices - 1 + image_emb_len + attention_mask_assistant_crop_start = assistant_start_indices - 4 + attention_mask_assistant_crop_end = assistant_start_indices + + prompt_embed_list = [] + prompt_attention_mask_list = [] + image_embed_list = [] + image_attention_mask_list = [] + + for i in range(text_input_ids.shape[0]): + prompt_embed_list.append( + torch.cat( + [ + prompt_embeds[i, text_crop_start : assistant_crop_start[i].item()], + prompt_embeds[i, assistant_crop_end[i].item() :], + ] + ) + ) + prompt_attention_mask_list.append( + torch.cat( + [ + prompt_attention_mask[i, crop_start : attention_mask_assistant_crop_start[i].item()], + prompt_attention_mask[i, attention_mask_assistant_crop_end[i].item() :], + ] + ) + ) + image_embed_list.append(prompt_embeds[i, image_emb_start:image_emb_end]) + image_attention_mask_list.append( + torch.ones(image_embed_list[-1].shape[0]).to(prompt_embeds.device).to(prompt_attention_mask.dtype) + ) + + prompt_embed_list = torch.stack(prompt_embed_list) + prompt_attention_mask_list = torch.stack(prompt_attention_mask_list) + image_embed_list = torch.stack(image_embed_list) + image_attention_mask_list = torch.stack(image_attention_mask_list) + + if 0 < image_embed_interleave < 6: + image_embed_list = image_embed_list[:, ::image_embed_interleave, :] + image_attention_mask_list = image_attention_mask_list[:, ::image_embed_interleave] + + assert ( + prompt_embed_list.shape[0] == prompt_attention_mask_list.shape[0] + and image_embed_list.shape[0] == image_attention_mask_list.shape[0] + ) + + prompt_embeds = torch.cat([image_embed_list, prompt_embed_list], dim=1) + prompt_attention_mask = torch.cat([image_attention_mask_list, prompt_attention_mask_list], dim=1) + + return prompt_embeds, prompt_attention_mask + + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 77, + ) -> torch.Tensor: + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output + return prompt_embeds + + def encode_prompt( + self, + image: torch.Tensor, + prompt: str | list[str], + prompt_2: str | list[str] = None, + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + max_sequence_length: int = 256, + image_embed_interleave: int = 2, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds( + image, + prompt, + prompt_template, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=max_sequence_length, + image_embed_interleave=image_embed_interleave, + ) + + if pooled_prompt_embeds is None: + if prompt_2 is None: + prompt_2 = prompt + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt, + num_videos_per_prompt, + device=device, + dtype=dtype, + max_sequence_length=77, + ) + + return prompt_embeds, pooled_prompt_embeds, prompt_attention_mask + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + prompt_template=None, + true_cfg_scale=1.0, + guidance_scale=1.0, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_template is not None: + if not isinstance(prompt_template, dict): + raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}") + if "template" not in prompt_template: + raise ValueError( + f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}" + ) + + if true_cfg_scale > 1.0 and guidance_scale > 1.0: + logger.warning( + "Both `true_cfg_scale` and `guidance_scale` are greater than 1.0. This will result in both " + "classifier-free guidance and embedded-guidance to be applied. This is not recommended " + "as it may lead to higher memory usage, slower inference and potentially worse results." + ) + + def prepare_latents( + self, + image: torch.Tensor, + batch_size: int, + num_channels_latents: int = 32, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + image_condition_type: str = "latent_concat", + ) -> torch.Tensor: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height, latent_width = height // self.vae_scale_factor_spatial, width // self.vae_scale_factor_spatial + shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width) + + image = image.unsqueeze(2) # [B, C, 1, H, W] + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i].unsqueeze(0)), generator[i], "argmax") + for i in range(batch_size) + ] + else: + image_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator, "argmax") for img in image] + + image_latents = torch.cat(image_latents, dim=0).to(dtype) * self.vae_scaling_factor + image_latents = image_latents.repeat(1, 1, num_latent_frames, 1, 1) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + t = torch.tensor([0.999]).to(device=device) + latents = latents * t + image_latents * (1 - t) + + if image_condition_type == "token_replace": + image_latents = image_latents[:, :, :1] + + return latents, image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PIL.Image.Image, + prompt: str | list[str] = None, + prompt_2: str | list[str] = None, + negative_prompt: str | list[str] = None, + negative_prompt_2: str | list[str] = None, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + num_inference_steps: int = 50, + sigmas: list[float] = None, + true_cfg_scale: float = 1.0, + guidance_scale: float = 1.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + prompt_template: dict[str, Any] = DEFAULT_PROMPT_TEMPLATE, + max_sequence_length: int = 256, + image_embed_interleave: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image`): + The input image to condition the video generation on. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + height (`int`, defaults to `720`): + The height in pixels of the generated image. + width (`int`, defaults to `1280`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `129`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + guidance_scale (`float`, defaults to `1.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. Note that the only available + HunyuanVideo model is CFG-distilled, which means that traditional guidance between unconditional and + conditional latent is not applied. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. Required when `negative_prompt_embeds` is passed directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + prompt_template (`dict`, *optional*): + Template used to format the prompt before encoding. Defaults to the model's default template. + max_sequence_length (`int`, *optional*, defaults to 256): + Maximum sequence length to use for the prompt encoder. + image_embed_interleave (`int`, *optional*): + Number of image embedding tokens to interleave with text tokens. If not provided, a sensible default is + chosen based on the transformer's `image_condition_type`. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~HunyuanVideoPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + prompt_embeds, + callback_on_step_end_tensor_inputs, + prompt_template, + true_cfg_scale, + guidance_scale, + ) + + image_condition_type = self.transformer.config.image_condition_type + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + image_embed_interleave = ( + image_embed_interleave + if image_embed_interleave is not None + else ( + 2 if image_condition_type == "latent_concat" else 4 if image_condition_type == "token_replace" else 1 + ) + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Prepare latent variables + vae_dtype = self.vae.dtype + image_tensor = self.video_processor.preprocess(image, height, width).to(device, vae_dtype) + + if image_condition_type == "latent_concat": + num_channels_latents = (self.transformer.config.in_channels - 1) // 2 + elif image_condition_type == "token_replace": + num_channels_latents = self.transformer.config.in_channels + + latents, image_latents = self.prepare_latents( + image_tensor, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + image_condition_type, + ) + if image_condition_type == "latent_concat": + image_latents[:, :, 1:] = 0 + mask = image_latents.new_ones(image_latents.shape[0], 1, *image_latents.shape[2:]) + mask[:, :, 1:] = 0 + + # 4. Encode input prompt + transformer_dtype = self.transformer.dtype + prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt( + image=image, + prompt=prompt, + prompt_2=prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + image_embed_interleave=image_embed_interleave, + ) + prompt_embeds = prompt_embeds.to(transformer_dtype) + prompt_attention_mask = prompt_attention_mask.to(transformer_dtype) + pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype) + + if do_true_cfg: + black_image = PIL.Image.new("RGB", (width, height), 0) + negative_prompt_embeds, negative_pooled_prompt_embeds, negative_prompt_attention_mask = self.encode_prompt( + image=black_image, + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_template=prompt_template, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + prompt_attention_mask=negative_prompt_attention_mask, + device=device, + max_sequence_length=max_sequence_length, + ) + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + negative_prompt_attention_mask = negative_prompt_attention_mask.to(transformer_dtype) + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(transformer_dtype) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas) + + # 6. Prepare guidance condition + guidance = None + if self.transformer.config.guidance_embeds: + guidance = ( + torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0 + ) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + if image_condition_type == "latent_concat": + latent_model_input = torch.cat([latents, image_latents, mask], dim=1).to(transformer_dtype) + elif image_condition_type == "token_replace": + latent_model_input = torch.cat([image_latents, latents[:, :, 1:]], dim=2).to(transformer_dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + pooled_projections=pooled_prompt_embeds, + guidance=guidance, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_attention_mask=negative_prompt_attention_mask, + pooled_projections=negative_pooled_prompt_embeds, + guidance=guidance, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + if image_condition_type == "latent_concat": + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + elif image_condition_type == "token_replace": + latents = latents = self.scheduler.step( + noise_pred[:, :, 1:], t, latents[:, :, 1:], return_dict=False + )[0] + latents = torch.cat([image_latents, latents], dim=2) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) / self.vae_scaling_factor + video = self.vae.decode(latents, return_dict=False)[0] + if image_condition_type == "latent_concat": + video = video[:, :, 4:, :, :] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + if image_condition_type == "latent_concat": + video = latents[:, :, 1:, :, :] + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HunyuanVideoPipelineOutput(frames=video) diff --git a/diffusers/pipelines/hunyuan_video/pipeline_output.py b/diffusers/pipelines/hunyuan_video/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..4cf96626277047d35d7d06e1481af6ee2fb56447 --- /dev/null +++ b/diffusers/pipelines/hunyuan_video/pipeline_output.py @@ -0,0 +1,38 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class HunyuanVideoPipelineOutput(BaseOutput): + r""" + Output class for HunyuanVideo pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor + + +@dataclass +class HunyuanVideoFramepackPipelineOutput(BaseOutput): + r""" + Output class for HunyuanVideo pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. Or, a list of torch tensors where each tensor + corresponds to a latent that decodes to multiple frames. + """ + + frames: torch.Tensor | np.ndarray | list[list[PIL.Image.Image]] | list[torch.Tensor] diff --git a/diffusers/pipelines/hunyuan_video1_5/__init__.py b/diffusers/pipelines/hunyuan_video1_5/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..846320f4ace07d502c0c8d067d63527a80664882 --- /dev/null +++ b/diffusers/pipelines/hunyuan_video1_5/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_hunyuan_video1_5"] = ["HunyuanVideo15Pipeline"] + _import_structure["pipeline_hunyuan_video1_5_image2video"] = ["HunyuanVideo15ImageToVideoPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_hunyuan_video1_5 import HunyuanVideo15Pipeline + from .pipeline_hunyuan_video1_5_image2video import HunyuanVideo15ImageToVideoPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/hunyuan_video1_5/image_processor.py b/diffusers/pipelines/hunyuan_video1_5/image_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..82817365b6a529c2307540af8e3fbb268ad0b05a --- /dev/null +++ b/diffusers/pipelines/hunyuan_video1_5/image_processor.py @@ -0,0 +1,103 @@ +# Copyright 2025 The HunyuanVideo Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np + +from ...configuration_utils import register_to_config +from ...video_processor import VideoProcessor + + +# copied from https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5/blob/main/hyvideo/utils/data_utils.py#L20 +def generate_crop_size_list(base_size=256, patch_size=16, max_ratio=4.0): + num_patches = round((base_size / patch_size) ** 2) + assert max_ratio >= 1.0 + crop_size_list = [] + wp, hp = num_patches, 1 + while wp > 0: + if max(wp, hp) / min(wp, hp) <= max_ratio: + crop_size_list.append((wp * patch_size, hp * patch_size)) + if (hp + 1) * wp <= num_patches: + hp += 1 + else: + wp -= 1 + return crop_size_list + + +# copied from https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5/blob/main/hyvideo/utils/data_utils.py#L38 +def get_closest_ratio(height: float, width: float, ratios: list, buckets: list): + """ + Get the closest ratio in the buckets. + + Args: + height (float): video height + width (float): video width + ratios (list): video aspect ratio + buckets (list): buckets generated by `generate_crop_size_list` + + Returns: + the closest size in the buckets and the corresponding ratio + """ + aspect_ratio = float(height) / float(width) + diff_ratios = ratios - aspect_ratio + + if aspect_ratio >= 1: + indices = [(index, x) for index, x in enumerate(diff_ratios) if x <= 0] + else: + indices = [(index, x) for index, x in enumerate(diff_ratios) if x >= 0] + + closest_ratio_id = min(indices, key=lambda pair: abs(pair[1]))[0] + closest_size = buckets[closest_ratio_id] + closest_ratio = ratios[closest_ratio_id] + + return closest_size, closest_ratio + + +class HunyuanVideo15ImageProcessor(VideoProcessor): + r""" + Image/video processor to preproces/postprocess the reference image/generatedvideo for the HunyuanVideo1.5 model. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept + `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method. + vae_scale_factor (`int`, *optional*, defaults to `16`): + VAE (spatial) scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of + this factor. + vae_latent_channels (`int`, *optional*, defaults to `32`): + VAE latent channels. + do_convert_rgb (`bool`, *optional*, defaults to `True`): + Whether to convert the image to RGB. + """ + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 16, + vae_latent_channels: int = 32, + do_convert_rgb: bool = True, + ): + super().__init__( + do_resize=do_resize, + vae_scale_factor=vae_scale_factor, + vae_latent_channels=vae_latent_channels, + do_convert_rgb=do_convert_rgb, + ) + + def calculate_default_height_width(self, height: int, width: int, target_size: int): + crop_size_list = generate_crop_size_list(base_size=target_size, patch_size=self.config.vae_scale_factor) + aspect_ratios = np.array([round(float(h) / float(w), 5) for h, w in crop_size_list]) + height, width = get_closest_ratio(height, width, aspect_ratios, crop_size_list)[0] + + return height, width diff --git a/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5.py b/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5.py new file mode 100644 index 0000000000000000000000000000000000000000..a0adff493ac0e7243f64a8763be74f4a5daeb9a8 --- /dev/null +++ b/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5.py @@ -0,0 +1,837 @@ +# Copyright 2025 The HunyuanVideo Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import re +from typing import Any + +import numpy as np +import torch +from transformers import ByT5Tokenizer, Qwen2_5_VLTextModel, Qwen2Tokenizer, T5EncoderModel + +from ...guiders import ClassifierFreeGuidance +from ...models import AutoencoderKLHunyuanVideo15, HunyuanVideo15Transformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import HunyuanVideo15ImageProcessor +from .pipeline_output import HunyuanVideo15PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import HunyuanVideo15Pipeline + >>> from diffusers.utils import export_to_video + + >>> model_id = "hunyuanvideo-community/HunyuanVideo-1.5-480p_t2v" + >>> pipe = HunyuanVideo15Pipeline.from_pretrained(model_id, torch_dtype=torch.float16) + >>> pipe.vae.enable_tiling() + >>> pipe.to("cuda") + + >>> output = pipe( + ... prompt="A cat walks on the grass, realistic", + ... num_inference_steps=50, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=15) + ``` +""" + + +def format_text_input(prompt: list[str], system_message: str) -> list[dict[str, Any]]: + """ + Apply text to template. + + Args: + prompt (list[str]): Input text. + system_message (str): System message. + + Returns: + list[dict[str, Any]]: List of chat conversation. + """ + + template = [ + [{"role": "system", "content": system_message}, {"role": "user", "content": p if p else " "}] for p in prompt + ] + + return template + + +def extract_glyph_texts(prompt: str) -> list[str]: + """ + Extract glyph texts from prompt using regex pattern. + + Args: + prompt: Input prompt string + + Returns: + List of extracted glyph texts + """ + pattern = r"\"(.*?)\"|“(.*?)”" + matches = re.findall(pattern, prompt) + result = [match[0] or match[1] for match in matches] + result = list(dict.fromkeys(result)) if len(result) > 1 else result + + if result: + formatted_result = ". ".join([f'Text "{text}"' for text in result]) + ". " + else: + formatted_result = None + + return formatted_result + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class HunyuanVideo15Pipeline(DiffusionPipeline): + r""" + Pipeline for text-to-video generation using HunyuanVideo1.5. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + transformer ([`HunyuanVideo15Transformer3DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded video latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded video latents. + vae ([`AutoencoderKLHunyuanVideo15`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): Tokenizer of class [Qwen2Tokenizer]. + text_encoder_2 ([`T5EncoderModel`]): + [T5EncoderModel](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel) + variant. + tokenizer_2 (`ByT5Tokenizer`): Tokenizer of class [ByT5Tokenizer] + guider ([`ClassifierFreeGuidance`]): + [ClassifierFreeGuidance]for classifier free guidance. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + + def __init__( + self, + text_encoder: Qwen2_5_VLTextModel, + tokenizer: Qwen2Tokenizer, + transformer: HunyuanVideo15Transformer3DModel, + vae: AutoencoderKLHunyuanVideo15, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder_2: T5EncoderModel, + tokenizer_2: ByT5Tokenizer, + guider: ClassifierFreeGuidance, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + guider=guider, + ) + + self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 16 + self.video_processor = HunyuanVideo15ImageProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.target_size = self.transformer.config.target_size if getattr(self, "transformer", None) else 640 + self.vision_states_dim = ( + self.transformer.config.image_embed_dim if getattr(self, "transformer", None) else 1152 + ) + self.num_channels_latents = self.vae.config.latent_channels if hasattr(self, "vae") else 32 + # fmt: off + self.system_message = "You are a helpful assistant. Describe the video by detailing the following aspects: \ + 1. The main content and theme of the video. \ + 2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects. \ + 3. Actions, events, behaviors temporal relationships, physical movement changes of the objects. \ + 4. background environment, light, style and atmosphere. \ + 5. camera angles, movements, and transitions used in the video." + # fmt: on + self.prompt_template_encode_start_idx = 108 + self.tokenizer_max_length = 1000 + self.tokenizer_2_max_length = 256 + self.vision_num_semantic_tokens = 729 + self.default_aspect_ratio = (16, 9) # (width: height) + + @staticmethod + def _get_mllm_prompt_embeds( + text_encoder: Qwen2_5_VLTextModel, + tokenizer: Qwen2Tokenizer, + prompt: str | list[str], + device: torch.device, + tokenizer_max_length: int = 1000, + num_hidden_layers_to_skip: int = 2, + # fmt: off + system_message: str = "You are a helpful assistant. Describe the video by detailing the following aspects: \ + 1. The main content and theme of the video. \ + 2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects. \ + 3. Actions, events, behaviors temporal relationships, physical movement changes of the objects. \ + 4. background environment, light, style and atmosphere. \ + 5. camera angles, movements, and transitions used in the video.", + # fmt: on + crop_start: int = 108, + ) -> tuple[torch.Tensor, torch.Tensor]: + prompt = [prompt] if isinstance(prompt, str) else prompt + + prompt = format_text_input(prompt, system_message) + + text_inputs = tokenizer.apply_chat_template( + prompt, + add_generation_prompt=True, + tokenize=True, + return_dict=True, + padding="max_length", + max_length=tokenizer_max_length + crop_start, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device=device) + prompt_attention_mask = text_inputs.attention_mask.to(device=device) + + prompt_embeds = text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-(num_hidden_layers_to_skip + 1)] + + if crop_start is not None and crop_start > 0: + prompt_embeds = prompt_embeds[:, crop_start:] + prompt_attention_mask = prompt_attention_mask[:, crop_start:] + + return prompt_embeds, prompt_attention_mask + + @staticmethod + def _get_byt5_prompt_embeds( + tokenizer: ByT5Tokenizer, + text_encoder: T5EncoderModel, + prompt: str | list[str], + device: torch.device, + tokenizer_max_length: int = 256, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + + glyph_texts = [extract_glyph_texts(p) for p in prompt] + + prompt_embeds_list = [] + prompt_embeds_mask_list = [] + + for glyph_text in glyph_texts: + if glyph_text is None: + glyph_text_embeds = torch.zeros( + (1, tokenizer_max_length, text_encoder.config.d_model), device=device, dtype=text_encoder.dtype + ) + glyph_text_embeds_mask = torch.zeros((1, tokenizer_max_length), device=device, dtype=torch.int64) + else: + txt_tokens = tokenizer( + glyph_text, + padding="max_length", + max_length=tokenizer_max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ).to(device) + + glyph_text_embeds = text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask.float(), + )[0] + glyph_text_embeds = glyph_text_embeds.to(device=device) + glyph_text_embeds_mask = txt_tokens.attention_mask.to(device=device) + + prompt_embeds_list.append(glyph_text_embeds) + prompt_embeds_mask_list.append(glyph_text_embeds_mask) + + prompt_embeds = torch.cat(prompt_embeds_list, dim=0) + prompt_embeds_mask = torch.cat(prompt_embeds_mask_list, dim=0) + + return prompt_embeds, prompt_embeds_mask + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + dtype: torch.dtype | None = None, + batch_size: int = 1, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + prompt_embeds_mask_2: torch.Tensor | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + batch_size (`int`): + batch size of prompts, defaults to 1 + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. If not provided, text embeddings will be generated from `prompt` input + argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated text mask. If not provided, text mask will be generated from `prompt` input argument. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text embeddings from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text mask from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + if prompt is None: + prompt = [""] * batch_size + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_mllm_prompt_embeds( + tokenizer=self.tokenizer, + text_encoder=self.text_encoder, + prompt=prompt, + device=device, + tokenizer_max_length=self.tokenizer_max_length, + system_message=self.system_message, + crop_start=self.prompt_template_encode_start_idx, + ) + + if prompt_embeds_2 is None: + prompt_embeds_2, prompt_embeds_mask_2 = self._get_byt5_prompt_embeds( + tokenizer=self.tokenizer_2, + text_encoder=self.text_encoder_2, + prompt=prompt, + device=device, + tokenizer_max_length=self.tokenizer_2_max_length, + ) + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_videos_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_videos_per_prompt, seq_len) + + _, seq_len_2, _ = prompt_embeds_2.shape + prompt_embeds_2 = prompt_embeds_2.repeat(1, num_videos_per_prompt, 1) + prompt_embeds_2 = prompt_embeds_2.view(batch_size * num_videos_per_prompt, seq_len_2, -1) + prompt_embeds_mask_2 = prompt_embeds_mask_2.repeat(1, num_videos_per_prompt, 1) + prompt_embeds_mask_2 = prompt_embeds_mask_2.view(batch_size * num_videos_per_prompt, seq_len_2) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds_mask = prompt_embeds_mask.to(dtype=dtype, device=device) + prompt_embeds_2 = prompt_embeds_2.to(dtype=dtype, device=device) + prompt_embeds_mask_2 = prompt_embeds_mask_2.to(dtype=dtype, device=device) + + return prompt_embeds, prompt_embeds_mask, prompt_embeds_2, prompt_embeds_mask_2 + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + prompt_embeds_2=None, + prompt_embeds_mask_2=None, + negative_prompt_embeds_2=None, + negative_prompt_embeds_mask_2=None, + ): + if height is None and width is not None: + raise ValueError("If `width` is provided, `height` also have to be provided.") + elif width is None and height is not None: + raise ValueError("If `height` is provided, `width` also have to be provided.") + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if prompt is None and prompt_embeds_2 is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_2`. Cannot leave both `prompt` and `prompt_embeds_2` undefined." + ) + + if prompt_embeds_2 is not None and prompt_embeds_mask_2 is None: + raise ValueError( + "If `prompt_embeds_2` are provided, `prompt_embeds_mask_2` also have to be passed. Make sure to generate `prompt_embeds_mask_2` from the same text encoder that was used to generate `prompt_embeds_2`." + ) + if negative_prompt_embeds_2 is not None and negative_prompt_embeds_mask_2 is None: + raise ValueError( + "If `negative_prompt_embeds_2` are provided, `negative_prompt_embeds_mask_2` also have to be passed. Make sure to generate `negative_prompt_embeds_mask_2` from the same text encoder that was used to generate `negative_prompt_embeds_2`." + ) + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 32, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + def prepare_cond_latents_and_mask(self, latents, dtype: torch.dtype | None, device: torch.device | None): + """ + Prepare conditional latents and mask for t2v generation. + + Args: + latents: Main latents tensor (B, C, F, H, W) + + Returns: + tuple: (cond_latents_concat, mask_concat) - both are zero tensors for t2v + """ + batch, channels, frames, height, width = latents.shape + + cond_latents_concat = torch.zeros(batch, channels, frames, height, width, dtype=dtype, device=device) + + mask_concat = torch.zeros(batch, 1, frames, height, width, dtype=dtype, device=device) + + return cond_latents_concat, mask_concat + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_frames: int = 121, + num_inference_steps: int = 50, + sigmas: list[float] = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + prompt_embeds_mask_2: torch.Tensor | None = None, + negative_prompt_embeds_2: torch.Tensor | None = None, + negative_prompt_embeds_mask_2: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. + height (`int`, *optional*): + The height in pixels of the generated video. + width (`int`, *optional*): + The width in pixels of the generated video. + num_frames (`int`, defaults to `121`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality video at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated mask for prompt embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated mask for negative prompt embeddings. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated text embeddings from the second text encoder. Can be used to easily tweak text inputs. + prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated mask for prompt embeddings from the second text encoder. + negative_prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings from the second text encoder. + negative_prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated mask for negative prompt embeddings from the second text encoder. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between "np", "pt", or "latent". + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HunyuanVideo15PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + + Examples: + + Returns: + [`~HunyuanVideo15PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HunyuanVideo15PipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated videos. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + prompt_embeds_2=prompt_embeds_2, + prompt_embeds_mask_2=prompt_embeds_mask_2, + negative_prompt_embeds_2=negative_prompt_embeds_2, + negative_prompt_embeds_mask_2=negative_prompt_embeds_mask_2, + ) + + if height is None and width is None: + height, width = self.video_processor.calculate_default_height_width( + self.default_aspect_ratio[1], self.default_aspect_ratio[0], self.target_size + ) + + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, prompt_embeds_mask, prompt_embeds_2, prompt_embeds_mask_2 = self.encode_prompt( + prompt=prompt, + device=device, + dtype=self.transformer.dtype, + batch_size=batch_size, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + prompt_embeds_2=prompt_embeds_2, + prompt_embeds_mask_2=prompt_embeds_mask_2, + ) + + if self.guider._enabled and self.guider.num_conditions > 1: + ( + negative_prompt_embeds, + negative_prompt_embeds_mask, + negative_prompt_embeds_2, + negative_prompt_embeds_mask_2, + ) = self.encode_prompt( + prompt=negative_prompt, + device=device, + dtype=self.transformer.dtype, + batch_size=batch_size, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + prompt_embeds_2=negative_prompt_embeds_2, + prompt_embeds_mask_2=negative_prompt_embeds_mask_2, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas) + + # 5. Prepare latent variables + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + self.num_channels_latents, + height, + width, + num_frames, + self.transformer.dtype, + device, + generator, + latents, + ) + cond_latents_concat, mask_concat = self.prepare_cond_latents_and_mask(latents, self.transformer.dtype, device) + image_embeds = torch.zeros( + batch_size, + self.vision_num_semantic_tokens, + self.vision_states_dim, + dtype=self.transformer.dtype, + device=device, + ) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = torch.cat([latents, cond_latents_concat, mask_concat], dim=1) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype) + + # Step 1: Collect model inputs needed for the guidance method + # conditional inputs should always be first element in the tuple + guider_inputs = { + "encoder_hidden_states": (prompt_embeds, negative_prompt_embeds), + "encoder_attention_mask": (prompt_embeds_mask, negative_prompt_embeds_mask), + "encoder_hidden_states_2": (prompt_embeds_2, negative_prompt_embeds_2), + "encoder_attention_mask_2": (prompt_embeds_mask_2, negative_prompt_embeds_mask_2), + } + + # Step 2: Update guider's internal state for this denoising step + self.guider.set_state(step=i, num_inference_steps=num_inference_steps, timestep=t) + + # Step 3: Prepare batched model inputs based on the guidance method + # The guider splits model inputs into separate batches for conditional/unconditional predictions. + # For CFG with guider_inputs = {"encoder_hidden_states": (prompt_embeds, negative_prompt_embeds)}: + # you will get a guider_state with two batches: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "__guidance_identifier__": "pred_cond"}, # conditional batch + # {"encoder_hidden_states": negative_prompt_embeds, "__guidance_identifier__": "pred_uncond"}, # unconditional batch + # ] + # Other guidance methods may return 1 batch (no guidance) or 3+ batches (e.g., PAG, APG). + guider_state = self.guider.prepare_inputs(guider_inputs) + # Step 4: Run the denoiser for each batch + # Each batch in guider_state represents a different conditioning (conditional, unconditional, etc.). + # We run the model once per batch and store the noise prediction in guider_state_batch.noise_pred. + for guider_state_batch in guider_state: + self.guider.prepare_models(self.transformer) + + # Extract conditioning kwargs for this batch (e.g., encoder_hidden_states) + cond_kwargs = { + input_name: getattr(guider_state_batch, input_name) for input_name in guider_inputs.keys() + } + + # e.g. "pred_cond"/"pred_uncond" + context_name = getattr(guider_state_batch, self.guider._identifier_key) + with self.transformer.cache_context(context_name): + # Run denoiser and store noise prediction in this batch + guider_state_batch.noise_pred = self.transformer( + hidden_states=latent_model_input, + image_embeds=image_embeds, + timestep=timestep, + attention_kwargs=self.attention_kwargs, + return_dict=False, + **cond_kwargs, + )[0] + + # Cleanup model (e.g., remove hooks) + self.guider.cleanup_models(self.transformer) + + # Step 5: Combine predictions using the guidance method + # The guider takes all noise predictions from guider_state and combines them according to the guidance algorithm. + # Continuing the CFG example, the guider receives: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "noise_pred": noise_pred_cond, "__guidance_identifier__": "pred_cond"}, # batch 0 + # {"encoder_hidden_states": negative_prompt_embeds, "noise_pred": noise_pred_uncond, "__guidance_identifier__": "pred_uncond"}, # batch 1 + # ] + # And extracts predictions using the __guidance_identifier__: + # pred_cond = guider_state[0]["noise_pred"] # extracts noise_pred_cond + # pred_uncond = guider_state[1]["noise_pred"] # extracts noise_pred_uncond + # Then applies CFG formula: + # noise_pred = pred_uncond + guidance_scale * (pred_cond - pred_uncond) + # Returns GuiderOutput(pred=noise_pred, pred_cond=pred_cond, pred_uncond=pred_uncond) + noise_pred = self.guider(guider_state)[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + # 8. decode the latents to video and postprocess + if not output_type == "latent": + latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HunyuanVideo15PipelineOutput(frames=video) diff --git a/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5_image2video.py b/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5_image2video.py new file mode 100644 index 0000000000000000000000000000000000000000..1d33c2ae188f93d6e24e4c13bc6c38c994e94f32 --- /dev/null +++ b/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5_image2video.py @@ -0,0 +1,960 @@ +# Copyright 2025 The HunyuanVideo Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import re +from typing import Any + +import numpy as np +import PIL +import torch +from transformers import ( + ByT5Tokenizer, + Qwen2_5_VLTextModel, + Qwen2Tokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, +) + +from ...guiders import ClassifierFreeGuidance +from ...models import AutoencoderKLHunyuanVideo15, HunyuanVideo15Transformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import HunyuanVideo15ImageProcessor +from .pipeline_output import HunyuanVideo15PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import HunyuanVideo15ImageToVideoPipeline + >>> from diffusers.utils import export_to_video + + >>> model_id = "hunyuanvideo-community/HunyuanVideo-1.5-480p_i2v" + >>> pipe = HunyuanVideo15ImageToVideoPipeline.from_pretrained(model_id, torch_dtype=torch.float16) + >>> pipe.vae.enable_tiling() + >>> pipe.to("cuda") + + >>> image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG") + + >>> output = pipe( + ... prompt="Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside.", + ... image=image, + ... num_inference_steps=50, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=24) + ``` +""" + + +# Copied from diffusers.pipelines.hunyuan_video1_5.pipeline_hunyuan_video1_5.format_text_input +def format_text_input(prompt: list[str], system_message: str) -> list[dict[str, Any]]: + """ + Apply text to template. + + Args: + prompt (list[str]): Input text. + system_message (str): System message. + + Returns: + list[dict[str, Any]]: List of chat conversation. + """ + + template = [ + [{"role": "system", "content": system_message}, {"role": "user", "content": p if p else " "}] for p in prompt + ] + + return template + + +# Copied from diffusers.pipelines.hunyuan_video1_5.pipeline_hunyuan_video1_5.extract_glyph_texts +def extract_glyph_texts(prompt: str) -> list[str]: + """ + Extract glyph texts from prompt using regex pattern. + + Args: + prompt: Input prompt string + + Returns: + List of extracted glyph texts + """ + pattern = r"\"(.*?)\"|“(.*?)”" + matches = re.findall(pattern, prompt) + result = [match[0] or match[1] for match in matches] + result = list(dict.fromkeys(result)) if len(result) > 1 else result + + if result: + formatted_result = ". ".join([f'Text "{text}"' for text in result]) + ". " + else: + formatted_result = None + + return formatted_result + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class HunyuanVideo15ImageToVideoPipeline(DiffusionPipeline): + r""" + Pipeline for image-to-video generation using HunyuanVideo1.5. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + transformer ([`HunyuanVideo15Transformer3DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded video latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded video latents. + vae ([`AutoencoderKLHunyuanVideo15`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): Tokenizer of class [Qwen2Tokenizer]. + text_encoder_2 ([`T5EncoderModel`]): + [T5EncoderModel](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel) + variant. + tokenizer_2 (`ByT5Tokenizer`): Tokenizer of class [ByT5Tokenizer] + guider ([`ClassifierFreeGuidance`]): + [ClassifierFreeGuidance]for classifier free guidance. + image_encoder ([`SiglipVisionModel`]): + [SiglipVisionModel](https://huggingface.co/docs/transformers/en/model_doc/siglip#transformers.SiglipVisionModel) + variant. + feature_extractor ([`SiglipImageProcessor`]): + [SiglipImageProcessor](https://huggingface.co/docs/transformers/en/model_doc/siglip#transformers.SiglipImageProcessor) + variant. + """ + + model_cpu_offload_seq = "image_encoder->text_encoder->transformer->vae" + + def __init__( + self, + text_encoder: Qwen2_5_VLTextModel, + tokenizer: Qwen2Tokenizer, + transformer: HunyuanVideo15Transformer3DModel, + vae: AutoencoderKLHunyuanVideo15, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder_2: T5EncoderModel, + tokenizer_2: ByT5Tokenizer, + guider: ClassifierFreeGuidance, + image_encoder: SiglipVisionModel, + feature_extractor: SiglipImageProcessor, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + guider=guider, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + + self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 16 + self.video_processor = HunyuanVideo15ImageProcessor( + vae_scale_factor=self.vae_scale_factor_spatial, do_resize=False, do_convert_rgb=True + ) + self.target_size = self.transformer.config.target_size if getattr(self, "transformer", None) else 640 + self.vision_states_dim = ( + self.transformer.config.image_embed_dim if getattr(self, "transformer", None) else 1152 + ) + self.num_channels_latents = self.vae.config.latent_channels if hasattr(self, "vae") else 32 + # fmt: off + self.system_message = "You are a helpful assistant. Describe the video by detailing the following aspects: \ + 1. The main content and theme of the video. \ + 2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects. \ + 3. Actions, events, behaviors temporal relationships, physical movement changes of the objects. \ + 4. background environment, light, style and atmosphere. \ + 5. camera angles, movements, and transitions used in the video." + # fmt: on + self.prompt_template_encode_start_idx = 108 + self.tokenizer_max_length = 1000 + self.tokenizer_2_max_length = 256 + self.vision_num_semantic_tokens = 729 + + @staticmethod + # Copied from diffusers.pipelines.hunyuan_video1_5.pipeline_hunyuan_video1_5.HunyuanVideo15Pipeline._get_mllm_prompt_embeds + def _get_mllm_prompt_embeds( + text_encoder: Qwen2_5_VLTextModel, + tokenizer: Qwen2Tokenizer, + prompt: str | list[str], + device: torch.device, + tokenizer_max_length: int = 1000, + num_hidden_layers_to_skip: int = 2, + # fmt: off + system_message: str = "You are a helpful assistant. Describe the video by detailing the following aspects: \ + 1. The main content and theme of the video. \ + 2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects. \ + 3. Actions, events, behaviors temporal relationships, physical movement changes of the objects. \ + 4. background environment, light, style and atmosphere. \ + 5. camera angles, movements, and transitions used in the video.", + # fmt: on + crop_start: int = 108, + ) -> tuple[torch.Tensor, torch.Tensor]: + prompt = [prompt] if isinstance(prompt, str) else prompt + + prompt = format_text_input(prompt, system_message) + + text_inputs = tokenizer.apply_chat_template( + prompt, + add_generation_prompt=True, + tokenize=True, + return_dict=True, + padding="max_length", + max_length=tokenizer_max_length + crop_start, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device=device) + prompt_attention_mask = text_inputs.attention_mask.to(device=device) + + prompt_embeds = text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_attention_mask, + output_hidden_states=True, + ).hidden_states[-(num_hidden_layers_to_skip + 1)] + + if crop_start is not None and crop_start > 0: + prompt_embeds = prompt_embeds[:, crop_start:] + prompt_attention_mask = prompt_attention_mask[:, crop_start:] + + return prompt_embeds, prompt_attention_mask + + @staticmethod + # Copied from diffusers.pipelines.hunyuan_video1_5.pipeline_hunyuan_video1_5.HunyuanVideo15Pipeline._get_byt5_prompt_embeds + def _get_byt5_prompt_embeds( + tokenizer: ByT5Tokenizer, + text_encoder: T5EncoderModel, + prompt: str | list[str], + device: torch.device, + tokenizer_max_length: int = 256, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + + glyph_texts = [extract_glyph_texts(p) for p in prompt] + + prompt_embeds_list = [] + prompt_embeds_mask_list = [] + + for glyph_text in glyph_texts: + if glyph_text is None: + glyph_text_embeds = torch.zeros( + (1, tokenizer_max_length, text_encoder.config.d_model), device=device, dtype=text_encoder.dtype + ) + glyph_text_embeds_mask = torch.zeros((1, tokenizer_max_length), device=device, dtype=torch.int64) + else: + txt_tokens = tokenizer( + glyph_text, + padding="max_length", + max_length=tokenizer_max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ).to(device) + + glyph_text_embeds = text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask.float(), + )[0] + glyph_text_embeds = glyph_text_embeds.to(device=device) + glyph_text_embeds_mask = txt_tokens.attention_mask.to(device=device) + + prompt_embeds_list.append(glyph_text_embeds) + prompt_embeds_mask_list.append(glyph_text_embeds_mask) + + prompt_embeds = torch.cat(prompt_embeds_list, dim=0) + prompt_embeds_mask = torch.cat(prompt_embeds_mask_list, dim=0) + + return prompt_embeds, prompt_embeds_mask + + @staticmethod + def _get_image_latents( + vae: AutoencoderKLHunyuanVideo15, + image_processor: HunyuanVideo15ImageProcessor, + image: PIL.Image.Image, + height: int, + width: int, + device: torch.device, + ) -> torch.Tensor: + vae_dtype = vae.dtype + image_tensor = image_processor.preprocess(image, height=height, width=width).to(device, dtype=vae_dtype) + image_tensor = image_tensor.unsqueeze(2) + image_latents = retrieve_latents(vae.encode(image_tensor), sample_mode="argmax") + image_latents = image_latents * vae.config.scaling_factor + return image_latents + + @staticmethod + def _get_image_embeds( + image_encoder: SiglipVisionModel, + feature_extractor: SiglipImageProcessor, + image: PIL.Image.Image, + device: torch.device, + ) -> torch.Tensor: + image_encoder_dtype = next(image_encoder.parameters()).dtype + image = feature_extractor.preprocess(images=image, do_resize=True, return_tensors="pt", do_convert_rgb=True) + image = image.to(device=device, dtype=image_encoder_dtype) + image_enc_hidden_states = image_encoder(**image).last_hidden_state + + return image_enc_hidden_states + + def encode_image( + self, + image: PIL.Image.Image, + batch_size: int, + device: torch.device, + dtype: torch.dtype, + ) -> torch.Tensor: + image_embeds = self._get_image_embeds( + image_encoder=self.image_encoder, + feature_extractor=self.feature_extractor, + image=image, + device=device, + ) + image_embeds = image_embeds.repeat(batch_size, 1, 1) + image_embeds = image_embeds.to(device=device, dtype=dtype) + return image_embeds + + # Copied from diffusers.pipelines.hunyuan_video1_5.pipeline_hunyuan_video1_5.HunyuanVideo15Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + dtype: torch.dtype | None = None, + batch_size: int = 1, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + prompt_embeds_mask_2: torch.Tensor | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + batch_size (`int`): + batch size of prompts, defaults to 1 + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. If not provided, text embeddings will be generated from `prompt` input + argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated text mask. If not provided, text mask will be generated from `prompt` input argument. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text embeddings from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated glyph text mask from ByT5. If not provided, will be generated from `prompt` input + argument using self.tokenizer_2 and self.text_encoder_2. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + if prompt is None: + prompt = [""] * batch_size + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_mllm_prompt_embeds( + tokenizer=self.tokenizer, + text_encoder=self.text_encoder, + prompt=prompt, + device=device, + tokenizer_max_length=self.tokenizer_max_length, + system_message=self.system_message, + crop_start=self.prompt_template_encode_start_idx, + ) + + if prompt_embeds_2 is None: + prompt_embeds_2, prompt_embeds_mask_2 = self._get_byt5_prompt_embeds( + tokenizer=self.tokenizer_2, + text_encoder=self.text_encoder_2, + prompt=prompt, + device=device, + tokenizer_max_length=self.tokenizer_2_max_length, + ) + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_videos_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_videos_per_prompt, seq_len) + + _, seq_len_2, _ = prompt_embeds_2.shape + prompt_embeds_2 = prompt_embeds_2.repeat(1, num_videos_per_prompt, 1) + prompt_embeds_2 = prompt_embeds_2.view(batch_size * num_videos_per_prompt, seq_len_2, -1) + prompt_embeds_mask_2 = prompt_embeds_mask_2.repeat(1, num_videos_per_prompt, 1) + prompt_embeds_mask_2 = prompt_embeds_mask_2.view(batch_size * num_videos_per_prompt, seq_len_2) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds_mask = prompt_embeds_mask.to(dtype=dtype, device=device) + prompt_embeds_2 = prompt_embeds_2.to(dtype=dtype, device=device) + prompt_embeds_mask_2 = prompt_embeds_mask_2.to(dtype=dtype, device=device) + + return prompt_embeds, prompt_embeds_mask, prompt_embeds_2, prompt_embeds_mask_2 + + def check_inputs( + self, + prompt, + image: PIL.Image.Image, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + prompt_embeds_2=None, + prompt_embeds_mask_2=None, + negative_prompt_embeds_2=None, + negative_prompt_embeds_mask_2=None, + ): + if not isinstance(image, PIL.Image.Image): + raise ValueError(f"`image` has to be of type `PIL.Image.Image` but is {type(image)}") + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if prompt is None and prompt_embeds_2 is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_2`. Cannot leave both `prompt` and `prompt_embeds_2` undefined." + ) + + if prompt_embeds_2 is not None and prompt_embeds_mask_2 is None: + raise ValueError( + "If `prompt_embeds_2` are provided, `prompt_embeds_mask_2` also have to be passed. Make sure to generate `prompt_embeds_mask_2` from the same text encoder that was used to generate `prompt_embeds_2`." + ) + if negative_prompt_embeds_2 is not None and negative_prompt_embeds_mask_2 is None: + raise ValueError( + "If `negative_prompt_embeds_2` are provided, `negative_prompt_embeds_mask_2` also have to be passed. Make sure to generate `negative_prompt_embeds_mask_2` from the same text encoder that was used to generate `negative_prompt_embeds_2`." + ) + + # Copied from diffusers.pipelines.hunyuan_video1_5.pipeline_hunyuan_video1_5.HunyuanVideo15Pipeline.prepare_latents + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 32, + height: int = 720, + width: int = 1280, + num_frames: int = 129, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + def prepare_cond_latents_and_mask( + self, + latents: torch.Tensor, + image: PIL.Image.Image, + batch_size: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + ): + """ + Prepare conditional latents and mask for t2v generation. + + Args: + latents: Main latents tensor (B, C, F, H, W) + + Returns: + tuple: (cond_latents_concat, mask_concat) - both are zero tensors for t2v + """ + + batch, channels, frames, latent_height, latent_width = latents.shape + + image_latents = self._get_image_latents( + vae=self.vae, + image_processor=self.video_processor, + image=image, + height=height, + width=width, + device=device, + ) + + latent_condition = image_latents.repeat(batch_size, 1, frames, 1, 1) + latent_condition[:, :, 1:, :, :] = 0 + latent_condition = latent_condition.to(device=device, dtype=dtype) + + latent_mask = torch.zeros(batch, 1, frames, latent_height, latent_width, dtype=dtype, device=device) + latent_mask[:, :, 0, :, :] = 1.0 + + return latent_condition, latent_mask + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PIL.Image.Image, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + num_frames: int = 121, + num_inference_steps: int = 50, + sigmas: list[float] = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + prompt_embeds_mask_2: torch.Tensor | None = None, + negative_prompt_embeds_2: torch.Tensor | None = None, + negative_prompt_embeds_mask_2: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image`): + The input image to condition video generation on. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds` + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the video generation. If not defined, one has to pass + `negative_prompt_embeds` instead. + num_frames (`int`, defaults to `121`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality video at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated mask for prompt embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated mask for negative prompt embeddings. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated text embeddings from the second text encoder. Can be used to easily tweak text inputs. + prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated mask for prompt embeddings from the second text encoder. + negative_prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings from the second text encoder. + negative_prompt_embeds_mask_2 (`torch.Tensor`, *optional*): + Pre-generated mask for negative prompt embeddings from the second text encoder. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between "np", "pt", or "latent". + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`HunyuanVideo15PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + + Examples: + + Returns: + [`~HunyuanVideo15PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`HunyuanVideo15PipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated videos. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + image=image, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + prompt_embeds_2=prompt_embeds_2, + prompt_embeds_mask_2=prompt_embeds_mask_2, + negative_prompt_embeds_2=negative_prompt_embeds_2, + negative_prompt_embeds_mask_2=negative_prompt_embeds_mask_2, + ) + + height, width = self.video_processor.calculate_default_height_width( + height=image.size[1], width=image.size[0], target_size=self.target_size + ) + image = self.video_processor.resize(image, height=height, width=width, resize_mode="crop") + + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode image + image_embeds = self.encode_image( + image=image, + batch_size=batch_size * num_videos_per_prompt, + device=device, + dtype=self.transformer.dtype, + ) + + # 4. Encode input prompt + prompt_embeds, prompt_embeds_mask, prompt_embeds_2, prompt_embeds_mask_2 = self.encode_prompt( + prompt=prompt, + device=device, + dtype=self.transformer.dtype, + batch_size=batch_size, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + prompt_embeds_2=prompt_embeds_2, + prompt_embeds_mask_2=prompt_embeds_mask_2, + ) + + if self.guider._enabled and self.guider.num_conditions > 1: + ( + negative_prompt_embeds, + negative_prompt_embeds_mask, + negative_prompt_embeds_2, + negative_prompt_embeds_mask_2, + ) = self.encode_prompt( + prompt=negative_prompt, + device=device, + dtype=self.transformer.dtype, + batch_size=batch_size, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + prompt_embeds_2=negative_prompt_embeds_2, + prompt_embeds_mask_2=negative_prompt_embeds_mask_2, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas) + + # 6. Prepare latent variables + latents = self.prepare_latents( + batch_size=batch_size * num_videos_per_prompt, + num_channels_latents=self.num_channels_latents, + height=height, + width=width, + num_frames=num_frames, + dtype=self.transformer.dtype, + device=device, + generator=generator, + latents=latents, + ) + + cond_latents_concat, mask_concat = self.prepare_cond_latents_and_mask( + latents=latents, + image=image, + batch_size=batch_size * num_videos_per_prompt, + height=height, + width=width, + dtype=self.transformer.dtype, + device=device, + ) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = torch.cat([latents, cond_latents_concat, mask_concat], dim=1) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype) + + if self.transformer.config.use_meanflow: + if i == len(timesteps) - 1: + timestep_r = torch.tensor([0.0], device=device) + else: + timestep_r = timesteps[i + 1] + timestep_r = timestep_r.expand(latents.shape[0]).to(latents.dtype) + else: + timestep_r = None + + # Step 1: Collect model inputs needed for the guidance method + # conditional inputs should always be first element in the tuple + guider_inputs = { + "encoder_hidden_states": (prompt_embeds, negative_prompt_embeds), + "encoder_attention_mask": (prompt_embeds_mask, negative_prompt_embeds_mask), + "encoder_hidden_states_2": (prompt_embeds_2, negative_prompt_embeds_2), + "encoder_attention_mask_2": (prompt_embeds_mask_2, negative_prompt_embeds_mask_2), + } + + # Step 2: Update guider's internal state for this denoising step + self.guider.set_state(step=i, num_inference_steps=num_inference_steps, timestep=t) + + # Step 3: Prepare batched model inputs based on the guidance method + # The guider splits model inputs into separate batches for conditional/unconditional predictions. + # For CFG with guider_inputs = {"encoder_hidden_states": (prompt_embeds, negative_prompt_embeds)}: + # you will get a guider_state with two batches: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "__guidance_identifier__": "pred_cond"}, # conditional batch + # {"encoder_hidden_states": negative_prompt_embeds, "__guidance_identifier__": "pred_uncond"}, # unconditional batch + # ] + # Other guidance methods may return 1 batch (no guidance) or 3+ batches (e.g., PAG, APG). + guider_state = self.guider.prepare_inputs(guider_inputs) + # Step 4: Run the denoiser for each batch + # Each batch in guider_state represents a different conditioning (conditional, unconditional, etc.). + # We run the model once per batch and store the noise prediction in guider_state_batch.noise_pred. + for guider_state_batch in guider_state: + self.guider.prepare_models(self.transformer) + + # Extract conditioning kwargs for this batch (e.g., encoder_hidden_states) + cond_kwargs = { + input_name: getattr(guider_state_batch, input_name) for input_name in guider_inputs.keys() + } + + # e.g. "pred_cond"/"pred_uncond" + context_name = getattr(guider_state_batch, self.guider._identifier_key) + with self.transformer.cache_context(context_name): + # Run denoiser and store noise prediction in this batch + guider_state_batch.noise_pred = self.transformer( + hidden_states=latent_model_input, + image_embeds=image_embeds, + timestep=timestep, + timestep_r=timestep_r, + attention_kwargs=self.attention_kwargs, + return_dict=False, + **cond_kwargs, + )[0] + + # Cleanup model (e.g., remove hooks) + self.guider.cleanup_models(self.transformer) + + # Step 5: Combine predictions using the guidance method + # The guider takes all noise predictions from guider_state and combines them according to the guidance algorithm. + # Continuing the CFG example, the guider receives: + # guider_state = [ + # {"encoder_hidden_states": prompt_embeds, "noise_pred": noise_pred_cond, "__guidance_identifier__": "pred_cond"}, # batch 0 + # {"encoder_hidden_states": negative_prompt_embeds, "noise_pred": noise_pred_uncond, "__guidance_identifier__": "pred_uncond"}, # batch 1 + # ] + # And extracts predictions using the __guidance_identifier__: + # pred_cond = guider_state[0]["noise_pred"] # extracts noise_pred_cond + # pred_uncond = guider_state[1]["noise_pred"] # extracts noise_pred_uncond + # Then applies CFG formula: + # noise_pred = pred_uncond + guidance_scale * (pred_cond - pred_uncond) + # Returns GuiderOutput(pred=noise_pred, pred_cond=pred_cond, pred_uncond=pred_uncond) + noise_pred = self.guider(guider_state)[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return HunyuanVideo15PipelineOutput(frames=video) diff --git a/diffusers/pipelines/hunyuan_video1_5/pipeline_output.py b/diffusers/pipelines/hunyuan_video1_5/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..ab8767299e55c961bed4d8394e01154029c00544 --- /dev/null +++ b/diffusers/pipelines/hunyuan_video1_5/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class HunyuanVideo15PipelineOutput(BaseOutput): + r""" + Output class for HunyuanVideo1.5 pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/diffusers/pipelines/hunyuandit/__init__.py b/diffusers/pipelines/hunyuandit/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8337399106f0585d15fa0a35f607baa2c04b203b --- /dev/null +++ b/diffusers/pipelines/hunyuandit/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_hunyuandit"] = ["HunyuanDiTPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_hunyuandit import HunyuanDiTPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/hunyuandit/pipeline_hunyuandit.py b/diffusers/pipelines/hunyuandit/pipeline_hunyuandit.py new file mode 100644 index 0000000000000000000000000000000000000000..5d656a3c370ac989a0aa5e0453f6f8ccda76ab96 --- /dev/null +++ b/diffusers/pipelines/hunyuandit/pipeline_hunyuandit.py @@ -0,0 +1,912 @@ +# Copyright 2025 HunyuanDiT Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import torch +from transformers import BertModel, BertTokenizer, CLIPImageProcessor, T5EncoderModel, T5Tokenizer + +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKL, HunyuanDiT2DModel +from ...models.embeddings import get_2d_rotary_pos_embed +from ...pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from ...schedulers import DDPMScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import HunyuanDiTPipeline + + >>> pipe = HunyuanDiTPipeline.from_pretrained( + ... "Tencent-Hunyuan/HunyuanDiT-Diffusers", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + + >>> # You may also use English prompt as HunyuanDiT supports both English and Chinese + >>> # prompt = "An astronaut riding a horse" + >>> prompt = "一个宇航员在骑马" + >>> image = pipe(prompt).images[0] + ``` +""" + +STANDARD_RATIO = np.array( + [ + 1.0, # 1:1 + 4.0 / 3.0, # 4:3 + 3.0 / 4.0, # 3:4 + 16.0 / 9.0, # 16:9 + 9.0 / 16.0, # 9:16 + ] +) +STANDARD_SHAPE = [ + [(1024, 1024), (1280, 1280)], # 1:1 + [(1024, 768), (1152, 864), (1280, 960)], # 4:3 + [(768, 1024), (864, 1152), (960, 1280)], # 3:4 + [(1280, 768)], # 16:9 + [(768, 1280)], # 9:16 +] +STANDARD_AREA = [np.array([w * h for w, h in shapes]) for shapes in STANDARD_SHAPE] +SUPPORTED_SHAPE = [ + (1024, 1024), + (1280, 1280), # 1:1 + (1024, 768), + (1152, 864), + (1280, 960), # 4:3 + (768, 1024), + (864, 1152), + (960, 1280), # 3:4 + (1280, 768), # 16:9 + (768, 1280), # 9:16 +] + + +def map_to_standard_shapes(target_width, target_height): + target_ratio = target_width / target_height + closest_ratio_idx = np.argmin(np.abs(STANDARD_RATIO - target_ratio)) + closest_area_idx = np.argmin(np.abs(STANDARD_AREA[closest_ratio_idx] - target_width * target_height)) + width, height = STANDARD_SHAPE[closest_ratio_idx][closest_area_idx] + return width, height + + +def get_resize_crop_region_for_grid(src, tgt_size): + th = tw = tgt_size + h, w = src + + r = h / w + + # resize + if r > 1: + resize_height = th + resize_width = int(round(th / h * w)) + else: + resize_width = tw + resize_height = int(round(tw / w * h)) + + crop_top = int(round((th - resize_height) / 2.0)) + crop_left = int(round((tw - resize_width) / 2.0)) + + return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class HunyuanDiTPipeline(DiffusionPipeline): + r""" + Pipeline for English/Chinese-to-image generation using HunyuanDiT. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + HunyuanDiT uses two text encoders: [mT5](https://huggingface.co/google/mt5-base) and [bilingual CLIP](fine-tuned by + ourselves) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. We use + `sdxl-vae-fp16-fix`. + text_encoder (`~transformers.BertModel`, `~transformers.CLIPTextModel` | None): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + HunyuanDiT uses a fine-tuned [bilingual CLIP]. + tokenizer (`~transformers.BertTokenizer`, `~transformers.CLIPTokenizer` | None): + A `BertTokenizer` or `CLIPTokenizer` to tokenize text. + transformer ([`HunyuanDiT2DModel`]): + The HunyuanDiT model designed by Tencent Hunyuan. + text_encoder_2 (`T5EncoderModel`): + The mT5 embedder. Specifically, it is 't5-v1_1-xxl'. + tokenizer_2 (`T5Tokenizer`): + The tokenizer for the mT5 embedder. + scheduler ([`DDPMScheduler`]): + A scheduler to be used in combination with HunyuanDiT to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [ + "safety_checker", + "feature_extractor", + "text_encoder_2", + "tokenizer_2", + "text_encoder", + "tokenizer", + ] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "prompt_embeds_2", + "negative_prompt_embeds_2", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: BertModel, + tokenizer: BertTokenizer, + transformer: HunyuanDiT2DModel, + scheduler: DDPMScheduler, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + text_encoder_2: T5EncoderModel | None = None, + tokenizer_2: T5Tokenizer | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + text_encoder_2=text_encoder_2, + ) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + + def encode_prompt( + self, + prompt: str, + device: torch.device = None, + dtype: torch.dtype = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int | None = None, + text_encoder_index: int = 0, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + dtype (`torch.dtype`): + torch dtype + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds` is passed directly. + max_sequence_length (`int`, *optional*): maximum sequence length to use for the prompt. + text_encoder_index (`int`, *optional*): + Index of the text encoder to use. `0` for clip and `1` for T5. + """ + if dtype is None: + if self.text_encoder_2 is not None: + dtype = self.text_encoder_2.dtype + elif self.transformer is not None: + dtype = self.transformer.dtype + else: + dtype = None + + if device is None: + device = self._execution_device + + tokenizers = [self.tokenizer, self.tokenizer_2] + text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = tokenizers[text_encoder_index] + text_encoder = text_encoders[text_encoder_index] + + if max_sequence_length is None: + if text_encoder_index == 0: + max_length = 77 + if text_encoder_index == 1: + max_length = 256 + else: + max_length = max_sequence_length + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_attention_mask = text_inputs.attention_mask.to(device) + prompt_embeds = text_encoder( + text_input_ids.to(device), + attention_mask=prompt_attention_mask, + ) + prompt_embeds = prompt_embeds[0] + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_attention_mask = uncond_input.attention_mask.to(device) + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + attention_mask=negative_prompt_attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds, negative_prompt_embeds, prompt_attention_mask, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + prompt_embeds_2=None, + negative_prompt_embeds_2=None, + prompt_attention_mask_2=None, + negative_prompt_attention_mask_2=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is None and prompt_embeds_2 is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_2`. Cannot leave both `prompt` and `prompt_embeds_2` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if prompt_embeds_2 is not None and prompt_attention_mask_2 is None: + raise ValueError("Must provide `prompt_attention_mask_2` when specifying `prompt_embeds_2`.") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if negative_prompt_embeds_2 is not None and negative_prompt_attention_mask_2 is None: + raise ValueError( + "Must provide `negative_prompt_attention_mask_2` when specifying `negative_prompt_embeds_2`." + ) + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_embeds_2 is not None and negative_prompt_embeds_2 is not None: + if prompt_embeds_2.shape != negative_prompt_embeds_2.shape: + raise ValueError( + "`prompt_embeds_2` and `negative_prompt_embeds_2` must have the same shape when passed directly, but" + f" got: `prompt_embeds_2` {prompt_embeds_2.shape} != `negative_prompt_embeds_2`" + f" {negative_prompt_embeds_2.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_2: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + prompt_attention_mask_2: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask_2: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + guidance_rescale: float = 0.0, + original_size: tuple[int, int] | None = (1024, 1024), + target_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + use_resolution_binning: bool = True, + ): + r""" + The call function to the pipeline for generation with HunyuanDiT. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`): + The height in pixels of the generated image. + width (`int`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + negative_prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds` is passed directly. + prompt_attention_mask_2 (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds_2` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds` is passed directly. + negative_prompt_attention_mask_2 (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds_2` is passed directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback_on_step_end (`Callable[[int, int], None]`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A callback function or a list of callback functions to be called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + A list of tensor inputs that should be passed to the callback function. If not defined, all tensor + inputs will be passed. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Rescale the noise_cfg according to `guidance_rescale`. Based on findings of [Common Diffusion Noise + Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891). See Section 3.4 + original_size (`tuple[int, int]`, *optional*, defaults to `(1024, 1024)`): + The original size of the image. Used to calculate the time ids. + target_size (`tuple[int, int]`, *optional*): + The target size of the image. Used to calculate the time ids. + crops_coords_top_left (`tuple[int, int]`, *optional*, defaults to `(0, 0)`): + The top left coordinates of the crop. Used to calculate the time ids. + use_resolution_binning (`bool`, *optional*, defaults to `True`): + Whether to use resolution binning or not. If `True`, the input resolution will be mapped to the closest + standard resolution. Supported resolutions are 1024x1024, 1280x1280, 1024x768, 1152x864, 1280x960, + 768x1024, 864x1152, 960x1280, 1280x768, and 768x1280. It is recommended to set this to `True`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. default height and width + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + height = int((height // 16) * 16) + width = int((width // 16) * 16) + + if use_resolution_binning and (height, width) not in SUPPORTED_SHAPE: + width, height = map_to_standard_shapes(width, height) + height = int(height) + width = int(width) + logger.warning(f"Reshaped to (height, width)=({height}, {width}), Supported shapes are {SUPPORTED_SHAPE}") + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + prompt_embeds_2, + negative_prompt_embeds_2, + prompt_attention_mask_2, + negative_prompt_attention_mask_2, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + + ( + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + device=device, + dtype=self.transformer.dtype, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=77, + text_encoder_index=0, + ) + ( + prompt_embeds_2, + negative_prompt_embeds_2, + prompt_attention_mask_2, + negative_prompt_attention_mask_2, + ) = self.encode_prompt( + prompt=prompt, + device=device, + dtype=self.transformer.dtype, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds_2, + negative_prompt_embeds=negative_prompt_embeds_2, + prompt_attention_mask=prompt_attention_mask_2, + negative_prompt_attention_mask=negative_prompt_attention_mask_2, + max_sequence_length=256, + text_encoder_index=1, + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7 create image_rotary_emb, style embedding & time ids + grid_height = height // 8 // self.transformer.config.patch_size + grid_width = width // 8 // self.transformer.config.patch_size + base_size = 512 // 8 // self.transformer.config.patch_size + grid_crops_coords = get_resize_crop_region_for_grid((grid_height, grid_width), base_size) + image_rotary_emb = get_2d_rotary_pos_embed( + self.transformer.inner_dim // self.transformer.num_heads, + grid_crops_coords, + (grid_height, grid_width), + device=device, + output_type="pt", + ) + + style = torch.tensor([0], device=device) + + target_size = target_size or (height, width) + add_time_ids = list(original_size + target_size + crops_coords_top_left) + add_time_ids = torch.tensor([add_time_ids], dtype=prompt_embeds.dtype) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask]) + prompt_embeds_2 = torch.cat([negative_prompt_embeds_2, prompt_embeds_2]) + prompt_attention_mask_2 = torch.cat([negative_prompt_attention_mask_2, prompt_attention_mask_2]) + add_time_ids = torch.cat([add_time_ids] * 2, dim=0) + style = torch.cat([style] * 2, dim=0) + + prompt_embeds = prompt_embeds.to(device=device) + prompt_attention_mask = prompt_attention_mask.to(device=device) + prompt_embeds_2 = prompt_embeds_2.to(device=device) + prompt_attention_mask_2 = prompt_attention_mask_2.to(device=device) + add_time_ids = add_time_ids.to(dtype=prompt_embeds.dtype, device=device).repeat( + batch_size * num_images_per_prompt, 1 + ) + style = style.to(device=device).repeat(batch_size * num_images_per_prompt) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # expand scalar t to 1-D tensor to match the 1st dim of latent_model_input + t_expand = torch.tensor([t] * latent_model_input.shape[0], device=device).to( + dtype=latent_model_input.dtype + ) + + # predict the noise residual + noise_pred = self.transformer( + latent_model_input, + t_expand, + encoder_hidden_states=prompt_embeds, + text_embedding_mask=prompt_attention_mask, + encoder_hidden_states_t5=prompt_embeds_2, + text_embedding_mask_t5=prompt_attention_mask_2, + image_meta_size=add_time_ids, + style=style, + image_rotary_emb=image_rotary_emb, + return_dict=False, + )[0] + + noise_pred, _ = noise_pred.chunk(2, dim=1) + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + prompt_embeds_2 = callback_outputs.pop("prompt_embeds_2", prompt_embeds_2) + negative_prompt_embeds_2 = callback_outputs.pop( + "negative_prompt_embeds_2", negative_prompt_embeds_2 + ) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/ideogram4/__init__.py b/diffusers/pipelines/ideogram4/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ed5eb596f31984bce2a2ff0c9408d3b5029a7b1b --- /dev/null +++ b/diffusers/pipelines/ideogram4/__init__.py @@ -0,0 +1,49 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa: F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_ideogram4"] = ["Ideogram4Pipeline"] + _import_structure["pipeline_output"] = ["Ideogram4PipelineOutput"] + _import_structure["prompt_enhancer"] = ["Ideogram4PromptEnhancerHead"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_ideogram4 import Ideogram4Pipeline + from .pipeline_output import Ideogram4PipelineOutput + from .prompt_enhancer import Ideogram4PromptEnhancerHead +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/ideogram4/pipeline_ideogram4.py b/diffusers/pipelines/ideogram4/pipeline_ideogram4.py new file mode 100644 index 0000000000000000000000000000000000000000..541259ed1f039f40c89bf455d789b26794a89424 --- /dev/null +++ b/diffusers/pipelines/ideogram4/pipeline_ideogram4.py @@ -0,0 +1,747 @@ +# Copyright 2026 Ideogram AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from typing import Any, Callable + +import torch +from transformers import AutoTokenizer, PreTrainedModel +from transformers.masking_utils import create_causal_mask + +from ...image_processor import VaeImageProcessor +from ...loaders import Ideogram4LoraLoaderMixin +from ...models.autoencoders import AutoencoderKLFlux2 +from ...models.transformers.transformer_ideogram4 import ( + IMAGE_POSITION_OFFSET, + LLM_TOKEN_INDICATOR, + OUTPUT_IMAGE_INDICATOR, + SEQUENCE_PADDING_INDICATOR, + Ideogram4Transformer2DModel, +) +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_outlines_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import Ideogram4PipelineOutput +from .prompt_enhancer import ( + PROMPT_UPSAMPLE_TEMPERATURE, + Ideogram4PromptEnhancerHead, + build_caption_logits_processor, + build_prompt_enhancer, + generate_captions, +) + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Hidden states of these Qwen3-VL decoder layers are concatenated to form the per-token +# text conditioning consumed by the Ideogram4 transformer. +QWEN3_VL_ACTIVATION_LAYERS = (0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 35) + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import Ideogram4Pipeline + + >>> pipe = Ideogram4Pipeline.from_pretrained("ideogram-ai/ideogram-v4", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "A photo of a cat holding a sign that says hello world" + >>> # The defaults are the recommended settings for best quality. + >>> image = pipe(prompt, height=2048, width=2048, generator=torch.Generator("cuda").manual_seed(0)).images[0] + >>> image.save("ideogram4.png") + ``` +""" + + +def _logit_normal_sigmas( + num_inference_steps: int, + mu: float, + std: float = 1.0, + logsnr_min: float = -15.0, + logsnr_max: float = 18.0, + device: torch.device | None = None, +) -> torch.Tensor: + r""" + Build a length-`num_inference_steps` sigma schedule using the Ideogram4 logit-normal flow-matching schedule. + + Sigmas are returned in `[0, 1]` in decreasing order (sigma close to 1 corresponds to pure noise, sigma close to 0 + to clean data), matching diffusers conventions. + + The Ideogram4 schedule applies `sigma(s) = 1 - logit_normal_cdf_inverse(1 - s)` to `s = linspace(0, 1, N + 1)` and + keeps the first `N` entries; a terminal zero is appended downstream by the scheduler. + """ + intervals = torch.linspace(0.0, 1.0, num_inference_steps + 1, dtype=torch.float64) + # Apply the inverse CDF of a normal then push through the logistic to obtain a logit-normal CDF inverse. + z = torch.special.ndtri(intervals) + y = mu + std * z + t = 1.0 - torch.special.expit(y) + t_min = 1.0 / (1.0 + math.exp(0.5 * logsnr_max)) + t_max = 1.0 / (1.0 + math.exp(0.5 * logsnr_min)) + t = t.clamp(t_min, t_max) + # Convert from model time (0 = noise, 1 = data) to diffusers sigma (1 = noise, 0 = data) and reverse. + sigmas = (1.0 - t).flip(0) + # Drop the trailing 0; FlowMatchEulerDiscreteScheduler.set_timesteps appends one back internally. + sigmas = sigmas[:-1].to(dtype=torch.float32, device=device) + return sigmas + + +def _resolution_aware_mu( + height: int, + width: int, + base_mu: float, + base_resolution: tuple[int, int] = (512, 512), +) -> float: + """Shift the schedule mean as a function of image resolution.""" + num_pixels = height * width + base_pixels = base_resolution[0] * base_resolution[1] + return base_mu + 0.5 * math.log(num_pixels / base_pixels) + + +def _expand_tensor_to_effective_batch( + tensor: torch.Tensor, + batch_size: int, + num_per_prompt: int, + tensor_name: str | None = None, +) -> torch.Tensor: + """Replicate `tensor` along dim 0 from `batch_size` (or 1) to `batch_size * num_per_prompt`.""" + target_batch_size = batch_size * num_per_prompt + + if tensor.shape[0] == target_batch_size: + return tensor + + if tensor.shape[0] == 1: + repeat_by = target_batch_size + elif tensor.shape[0] == batch_size: + repeat_by = num_per_prompt + else: + tensor_name = f"`{tensor_name}`" if tensor_name is not None else "Tensor" + raise ValueError( + f"{tensor_name} batch size must be 1, `batch_size` ({batch_size}), or " + f"`batch_size * num_*_per_prompt` ({target_batch_size}), but got {tensor.shape[0]}." + ) + + return torch.repeat_interleave(tensor, repeats=repeat_by, dim=0, output_size=tensor.shape[0] * repeat_by) + + +class Ideogram4Pipeline(DiffusionPipeline, Ideogram4LoraLoaderMixin): + r""" + Text-to-image pipeline for Ideogram4. + + Ideogram4 is a flow-matching model trained with asymmetric classifier-free guidance: a `transformer` consumes + text-conditioned features alongside the image latents, while a separate `unconditional_transformer` denoises with + zeroed text features. The two velocity predictions are linearly blended each step. + + Args: + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + Flow-matching scheduler. The pipeline overrides the default sigma schedule with a resolution-aware + logit-normal schedule. + vae ([`AutoencoderKLFlux2`]): + Variational auto-encoder used to decode latents back into images. + text_encoder ([`PreTrainedModel`]): + Multimodal text encoder. The pipeline consumes hidden states from a fixed set of intermediate decoder + layers (see `QWEN3_VL_ACTIVATION_LAYERS`). + tokenizer ([`AutoTokenizer`]): + Tokenizer paired with `text_encoder`. + transformer ([`Ideogram4Transformer2DModel`]): + Conditional flow-matching transformer. + unconditional_transformer ([`Ideogram4Transformer2DModel`]): + Unconditional (asymmetric-CFG) flow-matching transformer. + """ + + model_cpu_offload_seq = "prompt_enhancer_head->text_encoder->transformer->unconditional_transformer->vae" + _optional_components = ["prompt_enhancer_head"] + _callback_tensor_inputs = ["latents"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLFlux2, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer, + transformer: Ideogram4Transformer2DModel, + unconditional_transformer: Ideogram4Transformer2DModel, + prompt_enhancer_head: Ideogram4PromptEnhancerHead | None = None, + ) -> None: + super().__init__() + + self.register_modules( + scheduler=scheduler, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + unconditional_transformer=unconditional_transformer, + prompt_enhancer_head=prompt_enhancer_head, + ) + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) is not None else 8 + ) + # Ideogram4 patchifies the VAE output by a factor of 2 before feeding into the transformer. + self.patch_size = 2 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * self.patch_size) + + # Built lazily on first upsample: the head-less encoder body + `prompt_enhancer_head`, combined. + self._prompt_enhancer = None + # Outlines logits processor for schema-constrained captions; built lazily on first upsample. + self._caption_logits_processor = None + + def upsample_prompt( + self, + prompt: str | list[str], + height: int = 2048, + width: int = 2048, + temperature: float = PROMPT_UPSAMPLE_TEMPERATURE, + max_new_tokens: int = 1024, + generator: torch.Generator | list[torch.Generator] | None = None, + device: torch.device | None = None, + ) -> list[str]: + """Rewrite each prompt into Ideogram4's native structured JSON caption. + + Requires the optional `prompt_enhancer_head` component, which is grafted onto the shared `text_encoder` body to + make it generative. Generation is schema-constrained when `outlines` is installed, otherwise it runs + unconstrained. Pass `generator` (the same one accepted by `__call__`) to make sampling reproducible. + """ + if self.prompt_enhancer_head is None: + raise ValueError( + "Prompt upsampling requires the `prompt_enhancer_head` component, which is not loaded. Load it and " + "pass it in, e.g.:\n" + " from diffusers import Ideogram4PromptEnhancerHead\n" + " head = Ideogram4PromptEnhancerHead.from_pretrained('diffusers/qwen3-vl-8b-instruct-lm-head')\n" + " pipe = Ideogram4Pipeline.from_pretrained(model_id, prompt_enhancer_head=head)" + ) + if self._prompt_enhancer is None: + self._prompt_enhancer = build_prompt_enhancer(self.text_encoder, self.prompt_enhancer_head) + if self._caption_logits_processor is None and is_outlines_available(): + self._caption_logits_processor = build_caption_logits_processor(self._prompt_enhancer, self.tokenizer) + if self._caption_logits_processor is None: + logger.warning_once( + "`outlines` is not installed; prompt upsampling runs unconstrained and may not return schema-valid " + "JSON. Install with `pip install outlines` for structured captions." + ) + + return generate_captions( + self._prompt_enhancer, + self.tokenizer, + self._caption_logits_processor, + prompt, + height, + width, + temperature=temperature, + max_new_tokens=max_new_tokens, + generator=generator, + device=device, + ) + + @staticmethod + def _prepare_ids( + text_lengths: list[int], + grid_h: int, + grid_w: int, + max_text_tokens: int, + device: torch.device, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Build the packed `[left-pad][text][image]` layout from the per-prompt text lengths and the image grid. + + Returns `position_ids` (3-axis MRoPE), `segment_ids` (block-diagonal attention) and `indicator` (per-token + text/image/pad role). + """ + batch_size = len(text_lengths) + num_image_tokens = grid_h * grid_w + total_seq_len = max_text_tokens + num_image_tokens + + # Image position ids (t=0, h, w); offset keeps them disjoint from text positions. + h_idx = torch.arange(grid_h).view(-1, 1).expand(grid_h, grid_w).reshape(-1) + w_idx = torch.arange(grid_w).view(1, -1).expand(grid_h, grid_w).reshape(-1) + t_idx = torch.zeros_like(h_idx) + image_pos = torch.stack([t_idx, h_idx, w_idx], dim=1) + IMAGE_POSITION_OFFSET + + position_ids = torch.zeros(batch_size, total_seq_len, 3, dtype=torch.long) + segment_ids = torch.full((batch_size, total_seq_len), SEQUENCE_PADDING_INDICATOR, dtype=torch.long) + indicator = torch.zeros(batch_size, total_seq_len, dtype=torch.long) + + for b, num_text in enumerate(text_lengths): + offset = max_text_tokens - num_text + + text_pos = torch.arange(num_text) + text_pos_3d = torch.stack([text_pos, text_pos, text_pos], dim=1) + position_ids[b, offset : offset + num_text] = text_pos_3d + position_ids[b, offset + num_text :] = image_pos + + indicator[b, offset : offset + num_text] = LLM_TOKEN_INDICATOR + indicator[b, offset + num_text :] = OUTPUT_IMAGE_INDICATOR + + segment_ids[b, offset : offset + num_text + num_image_tokens] = 1 + + return position_ids.to(device), segment_ids.to(device), indicator.to(device) + + @staticmethod + def _get_text_encoder_hidden_states( + text_encoder, + token_ids: torch.Tensor, + attention_mask: torch.Tensor, + pos_2d: torch.Tensor, + ) -> list[torch.Tensor]: + """Run the text encoder's decoder layers, returning the hidden states tapped at each activation layer.""" + + language_model = text_encoder.language_model + + inputs_embeds = language_model.embed_tokens(token_ids) + + position_ids_4d = pos_2d[None, ...].expand(4, pos_2d.shape[0], -1) + text_position_ids = position_ids_4d[0] + mrope_position_ids = position_ids_4d[1:] + + causal_mask = create_causal_mask( + config=language_model.config, + inputs_embeds=inputs_embeds, + attention_mask=attention_mask, + past_key_values=None, + position_ids=text_position_ids, + ) + position_embeddings = language_model.rotary_emb(inputs_embeds, mrope_position_ids) + + tap_set = set(QWEN3_VL_ACTIVATION_LAYERS) + captured: dict[int, torch.Tensor] = {} + hidden_states = inputs_embeds + for layer_idx, decoder_layer in enumerate(language_model.layers): + hidden_states = decoder_layer( + hidden_states, + attention_mask=causal_mask, + position_ids=text_position_ids, + past_key_values=None, + position_embeddings=position_embeddings, + ) + if layer_idx in tap_set: + captured[layer_idx] = hidden_states + + return [captured[i] for i in QWEN3_VL_ACTIVATION_LAYERS] + + def encode_prompt( + self, + prompt: str | list[str], + grid_h: int, + grid_w: int, + max_sequence_length: int, + device: torch.device, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Prepare the conditioning for the packed text+image sequence (one entry per prompt). + + Returns a flat tuple `(prompt_embeds, position_ids, segment_ids, indicator)`. The unconditional branch carries + no text, so the pipeline builds its (zeroed) inputs directly rather than encoding a negative prompt. + """ + prompts = [prompt] if isinstance(prompt, str) else list(prompt) + batch_size = len(prompts) + num_image_tokens = grid_h * grid_w + + # Tokenize each chat-formatted prompt and left-pad to `max_sequence_length`. Only the text region is fed to + # the encoder: the packed image tokens come after the text and the encoder is causal, so they never affect it. + token_ids = torch.zeros(batch_size, max_sequence_length, dtype=torch.long) + attention_mask = torch.zeros(batch_size, max_sequence_length, dtype=torch.long) + text_position_ids = torch.zeros(batch_size, max_sequence_length, dtype=torch.long) + text_lengths = [] + for b, text_prompt in enumerate(prompts): + messages = [{"role": "user", "content": [{"type": "text", "text": text_prompt}]}] + text = self.tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False) + toks = self.tokenizer(text, return_tensors="pt", add_special_tokens=False)["input_ids"][0] + n = int(toks.shape[0]) + if n > max_sequence_length: + raise ValueError(f"prompt has {n} tokens, exceeds max_sequence_length={max_sequence_length}") + text_lengths.append(n) + offset = max_sequence_length - n + token_ids[b, offset:] = toks + attention_mask[b, offset:] = 1 + text_position_ids[b, offset:] = torch.arange(n) + + # To support enable_model_cpu_offload, we need to move the text_encoder inputs to the text encoder's actual + # device te_device. This is necessary because the `CpuOffload` model offload hook attaches to a component's + # `forward` method, but we call text_encoder's submodules directly below, so the hook never fires to onload the + # model to the execution device. Other offloading techniques (group, sequential) would work without te_device + # because they hook submodules, not just the top-level component module. Note that in the + # enable_model_cpu_offload case te_device will actually be the offload device (e.g. CPU). + te_device = self.text_encoder.device + token_ids = token_ids.to(te_device) + attention_mask = attention_mask.to(te_device) + text_position_ids = text_position_ids.to(te_device) + + # Concatenate the tapped activation-layer hidden states into per-token text features, zeroing padding. + selected = self._get_text_encoder_hidden_states( + self.text_encoder, token_ids, attention_mask, text_position_ids + ) + text_features = torch.stack(selected, dim=0).permute(1, 2, 3, 0).reshape(batch_size, max_sequence_length, -1) + text_features = (text_features * attention_mask.to(text_features.dtype).unsqueeze(-1)).to(torch.float32) + text_features = text_features.to(device) + + position_ids, segment_ids, indicator = self._prepare_ids( + text_lengths, grid_h, grid_w, max_sequence_length, device + ) + + # Pack the text features into the full sequence; image positions carry no text features. + image_feature_padding = torch.zeros( + batch_size, num_image_tokens, text_features.shape[-1], dtype=text_features.dtype, device=device + ) + prompt_embeds = torch.cat([text_features, image_feature_padding], dim=1) + + return prompt_embeds, position_ids, segment_ids, indicator + + def prepare_latents( + self, + batch_size: int, + num_image_tokens: int, + latent_dim: int, + dtype: torch.dtype, + device: torch.device, + generator: torch.Generator | list[torch.Generator] | None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + shape = (batch_size, num_image_tokens, latent_dim) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self) -> float | None: + return self._guidance_scale + + @property + def num_timesteps(self) -> int: + return self._num_timesteps + + @property + def attention_kwargs(self) -> dict[str, Any] | None: + return self._attention_kwargs + + @property + def interrupt(self) -> bool: + return self._interrupt + + def check_inputs( + self, + prompt, + height, + width, + num_inference_steps, + guidance_scale, + guidance_schedule, + callback_on_step_end_tensor_inputs=None, + ): + if prompt is None: + raise ValueError("`prompt` must be provided.") + if not isinstance(prompt, (str, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` ({height}) and `width` ({width}) must both be divisible by {self.vae_scale_factor * self.patch_size} " + f"(vae_scale_factor * patch_size)." + ) + + # Guidance is controlled by either a constant `guidance_scale` or a per-step `guidance_schedule`; exactly + # one must be set (the `guidance_schedule` default makes the no-arg call use the recommended schedule). + if guidance_scale is not None and guidance_schedule is not None: + raise ValueError("Only one of `guidance_scale` and `guidance_schedule` may be set.") + if guidance_scale is None and guidance_schedule is None: + raise ValueError("One of `guidance_scale` and `guidance_schedule` must be set.") + if guidance_schedule is not None and len(guidance_schedule) != num_inference_steps: + raise ValueError( + f"`guidance_schedule` must have length `num_inference_steps` ({num_inference_steps}), " + f"got {len(guidance_schedule)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found " + f"{[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + height: int = 2048, + width: int = 2048, + num_inference_steps: int = 48, + guidance_scale: float | None = None, + guidance_schedule: list[float] | torch.Tensor | None = (7.0,) * 45 + (3.0,) * 3, + mu: float = 0.0, + std: float = 1.5, + prompt_upsampling: bool = False, + prompt_upsampling_temperature: float = PROMPT_UPSAMPLE_TEMPERATURE, + max_sequence_length: int = 2048, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[["Ideogram4Pipeline", int, int, dict[str, Any]], dict[str, Any]] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ) -> Ideogram4PipelineOutput | tuple[Any]: + r""" + Run text-to-image generation. + + Args: + prompt (`str` or `list[str]`): + Prompt(s) to guide image generation. + height (`int`, *optional*, defaults to 2048): + Output image height in pixels; must be a multiple of `vae_scale_factor * patch_size`. + width (`int`, *optional*, defaults to 2048): + Output image width in pixels; must be a multiple of `vae_scale_factor * patch_size`. + num_inference_steps (`int`, *optional*, defaults to 48): + Number of flow-matching steps. The default is the recommended setting for best quality. + guidance_scale (`float`, *optional*): + Constant classifier-free guidance scale applied at every step. The conditional and unconditional + velocity predictions are blended as `v = guidance_scale * v_pos + (1 - guidance_scale) * v_neg`. + Mutually exclusive with `guidance_schedule` (setting both raises). Defaults to `None`. + guidance_schedule (`list[float]` or `torch.Tensor`, *optional*): + Per-step guidance scale schedule; must have length `num_inference_steps`. The first entry corresponds + to the first step (largest noise level). Mutually exclusive with `guidance_scale`; exactly one must be + set. Defaults to the recommended schedule (7.0 for the main steps, dropping to 3.0 for the final 3 + "polish" steps). To use a constant scale instead, pass `guidance_scale` and `guidance_schedule=None`. + mu (`float`, *optional*, defaults to 0.0): + Base mean of the logit-normal flow-matching schedule. The schedule mean is shifted by half the log of + the resolution ratio relative to 512x512. + std (`float`, *optional*, defaults to 1.5): + Standard deviation of the logit-normal flow-matching schedule. + prompt_upsampling (`bool`, *optional*, defaults to `False`): + If `True`, rewrite `prompt` into Ideogram4's native structured JSON caption via + [`~Ideogram4Pipeline.upsample_prompt`] before encoding. Requires the optional `prompt_enhancer_head` + component; install `outlines` for schema-constrained captions. `generator` is reused to make the + upsampling reproducible. + prompt_upsampling_temperature (`float`, *optional*, defaults to 1.0): + Sampling temperature for prompt upsampling when `prompt_upsampling=True`. + max_sequence_length (`int`, *optional*, defaults to 2048): + Maximum number of text tokens per prompt. + num_images_per_prompt (`int`, *optional*, defaults to 1): + Number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + Generator(s) used to make sampling deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noise of shape `(batch_size, num_image_tokens, latent_dim)`. + output_type (`str`, *optional*, defaults to `"pil"`): + One of `"pil"`, `"np"`, `"pt"`, or `"latent"`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return an [`~pipelines.ideogram4.Ideogram4PipelineOutput`]. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary passed along to the attention processor of each transformer. A `"scale"` entry + scales the loaded LoRA weights (e.g. `{"scale": 0.7}`) when the PEFT backend is active. + callback_on_step_end (`Callable`, *optional*): + Callback invoked at the end of every denoising step. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + Names of tensors to expose to the callback via `callback_kwargs`. + + Examples: + + Returns: + [`~pipelines.ideogram4.Ideogram4PipelineOutput`] or `tuple`. + """ + self.check_inputs( + prompt=prompt, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + guidance_schedule=guidance_schedule, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + + device = self._execution_device + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 0. Optionally rewrite the prompt(s) into Ideogram4's native structured JSON caption. + if prompt_upsampling: + prompt = self.upsample_prompt( + prompt, + height=height, + width=width, + temperature=prompt_upsampling_temperature, + max_new_tokens=max_sequence_length, + generator=generator, + device=device, + ) + + # 1. Image grid (drives both the packed layout and the latent shape). + grid_h, grid_w = ( + height // (self.vae_scale_factor * self.patch_size), + width // (self.vae_scale_factor * self.patch_size), + ) + num_image_tokens = grid_h * grid_w + + # 2. Encode prompts into the packed conditioning (one entry per prompt). + llm_features, position_ids, segment_ids, indicator = self.encode_prompt( + prompt=prompt, + grid_h=grid_h, + grid_w=grid_w, + max_sequence_length=max_sequence_length, + device=device, + ) + + # 3. Replicate the conditioning for num_images_per_prompt. + llm_features = _expand_tensor_to_effective_batch(llm_features, batch_size, num_images_per_prompt) + position_ids = _expand_tensor_to_effective_batch(position_ids, batch_size, num_images_per_prompt) + segment_ids = _expand_tensor_to_effective_batch(segment_ids, batch_size, num_images_per_prompt) + indicator = _expand_tensor_to_effective_batch(indicator, batch_size, num_images_per_prompt) + + # 4. Unconditional (image-only) branch, derived from the conditioning: zeroed text features and the + # image-region slices of the layout. + neg_llm_features = torch.zeros( + batch_size * num_images_per_prompt, + num_image_tokens, + llm_features.shape[-1], + dtype=llm_features.dtype, + device=device, + ) + neg_position_ids = position_ids[:, max_sequence_length:] + neg_segment_ids = segment_ids[:, max_sequence_length:] + neg_indicator = indicator[:, max_sequence_length:] + + # 4. Set up the resolution-aware logit-normal schedule on the scheduler. + schedule_mu = _resolution_aware_mu(height=height, width=width, base_mu=mu) + sigmas = _logit_normal_sigmas(num_inference_steps, schedule_mu, std=std, device=device) + self.scheduler.set_timesteps(sigmas=sigmas.tolist(), device=device) + timesteps = self.scheduler.timesteps + self._num_timesteps = len(timesteps) + + # 5. Resolve the per-step guidance schedule (a constant `guidance_scale` broadcasts to every step, otherwise + # use the provided `guidance_schedule`, validated by `check_inputs`) and the tensor of per-step weights `gw`. + if guidance_scale is not None: + guidance_schedule = [float(guidance_scale)] * num_inference_steps + gw = torch.as_tensor(guidance_schedule, dtype=torch.float32, device=device) + + # 6. Prepare latents in the packed (B, num_image_tokens, latent_dim) layout. + latent_dim = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_image_tokens=num_image_tokens, + latent_dim=latent_dim, + dtype=torch.float32, + device=device, + generator=generator, + latents=latents, + ) + + # 7. Padding for the text region of the conditional packed sequence (image latents are appended after it). + max_text_tokens = max_sequence_length + text_z_padding = torch.zeros( + batch_size * num_images_per_prompt, + max_text_tokens, + latent_dim, + dtype=torch.float32, + device=device, + ) + + # The transformers run in their loaded compute dtype; cast the (otherwise float32) text features to match. + # `latents` stay float32 for scheduler precision and are cast per-step at the transformer call below. + llm_features = llm_features.to(self.transformer.dtype) + neg_llm_features = neg_llm_features.to(self.unconditional_transformer.dtype) + + # 8. Denoising loop. The scheduler stores `num_train_timesteps`-scaled timesteps; convert back to model time. + num_train_timesteps = self.scheduler.config.num_train_timesteps + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # Map sigma-domain timestep to model time `t` in [0, 1] (0 = noise, 1 = clean data). + t_model = 1.0 - (t.float() / num_train_timesteps) + t_model = t_model.expand(batch_size * num_images_per_prompt).to(self.transformer.dtype) + + # Conditional pass operates on the full packed sequence. + pos_z = torch.cat([text_z_padding, latents], dim=1).to(self.transformer.dtype) + pos_out = self.transformer( + hidden_states=pos_z, + timestep=t_model, + encoder_hidden_states=llm_features, + position_ids=position_ids, + segment_ids=segment_ids, + indicator=indicator, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + # Velocity (and guidance) is computed in float32 for scheduler precision; the transformers + # return their compute dtype, so cast the predicted velocities up here. + pos_v = pos_out[:, max_text_tokens:].to(torch.float32) + + # Unconditional pass uses image-only positions with zeroed text features. + neg_v = self.unconditional_transformer( + hidden_states=latents.to(self.unconditional_transformer.dtype), + timestep=t_model, + encoder_hidden_states=neg_llm_features, + position_ids=neg_position_ids, + segment_ids=neg_segment_ids, + indicator=neg_indicator, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0].to(torch.float32) + + # Expose the current step's guidance weight via `self.guidance_scale` so callbacks can read it. + self._guidance_scale = guidance_schedule[i] + gw_i = gw[i] + v = gw_i * pos_v + (1.0 - gw_i) * neg_v + + latents = self.scheduler.step(-v, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {k: locals()[k] for k in callback_on_step_end_tensor_inputs} + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + latents = callback_outputs.pop("latents", latents) + + progress_bar.update() + + # 9. Decode: unpatch the latents, denormalize with the VAE batch-norm stats, and decode through the VAE. + if output_type == "latent": + image = latents + else: + z = latents + # VAE bn stores per-channel statistics on the packed-channel latent space (ae_channels * patch ** 2). + bn_mean = self.vae.bn.running_mean.view(1, 1, -1).to(device=z.device, dtype=z.dtype) + bn_std = torch.sqrt(self.vae.bn.running_var + self.vae.config.batch_norm_eps).view(1, 1, -1) + bn_std = bn_std.to(device=z.device, dtype=z.dtype) + z = z * bn_std + bn_mean + + patch = self.patch_size + ae_channels = z.shape[-1] // (patch * patch) + z = z.view(batch_size * num_images_per_prompt, grid_h, grid_w, patch, patch, ae_channels) + z = z.permute(0, 5, 1, 3, 2, 4).contiguous() + z = z.view(batch_size * num_images_per_prompt, ae_channels, grid_h * patch, grid_w * patch) + + decoded = self.vae.decode(z.to(self.vae.dtype), return_dict=False)[0] + image = self.image_processor.postprocess(decoded.float(), output_type=output_type) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + return Ideogram4PipelineOutput(images=image) diff --git a/diffusers/pipelines/ideogram4/pipeline_output.py b/diffusers/pipelines/ideogram4/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..1f6b65e44575e8fceae1c4334a6834d5662eb371 --- /dev/null +++ b/diffusers/pipelines/ideogram4/pipeline_output.py @@ -0,0 +1,34 @@ +# Copyright 2026 Ideogram AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class Ideogram4PipelineOutput(BaseOutput): + """ + Output class for the Ideogram 4 pipeline. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`): + List of denoised PIL images of length `batch_size`, or numpy array of shape `(batch_size, height, width, + num_channels)`. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/ideogram4/prompt_enhancer.py b/diffusers/pipelines/ideogram4/prompt_enhancer.py new file mode 100644 index 0000000000000000000000000000000000000000..2b7867afe650288999eb5baf5ba7346d145e7fd5 --- /dev/null +++ b/diffusers/pipelines/ideogram4/prompt_enhancer.py @@ -0,0 +1,218 @@ +# Copyright 2026 Ideogram AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""Prompt-enhancement assets for Ideogram4. + +Ideogram4 is trained on a *structured JSON caption* rather than a free-form prompt. The optional prompt enhancer +rewrites a short user idea into that native caption schema by combining the head-less Qwen3-VL text encoder with the +optional `Ideogram4PromptEnhancerHead` component to form a generative model. + +This mirrors the role of Flux2's `system_messages.py`, but the target is a constrained JSON object instead of free +text, so `outlines` (an optional dependency) is used to guarantee a schema-valid result when available. + +The caption helpers here are shared by `Ideogram4Pipeline` and the modular `Ideogram4PromptUpsampleStep`. +""" + +import math + +import torch +from torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin +from ...utils import logging + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +PROMPT_UPSAMPLE_TEMPERATURE = 1.0 + + +class Ideogram4PromptEnhancerHead(ModelMixin, ConfigMixin): + """LM head that makes the head-less Qwen3-VL `text_encoder` generative for prompt upsampling. + + An optional pipeline component (`prompt_enhancer_head`): its weights load via a normal `from_pretrained` (its own + small repo, or bundled in the model repo) rather than an in-pipeline download. At upsample time the pipeline + combines it with the shared `text_encoder` body to form the generative model. + """ + + @register_to_config + def __init__(self, hidden_size: int = 4096, vocab_size: int = 151936): + super().__init__() + self.lm_head = nn.Linear(hidden_size, vocab_size, bias=False) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + return self.lm_head(hidden_states) + + +def build_prompt_enhancer(text_encoder, prompt_enhancer_head): + """Combine the head-less Qwen3-VL `text_encoder` body with `prompt_enhancer_head` into a generative model. + + The body is shared by reference (no second copy in memory); only the small head is extra. + """ + from accelerate import init_empty_weights + from transformers import Qwen3VLForConditionalGeneration + + with init_empty_weights(): + enhancer = Qwen3VLForConditionalGeneration(text_encoder.config) + enhancer.model = text_encoder + enhancer.lm_head = prompt_enhancer_head.lm_head + return enhancer.eval() + + +# System message that instructs the encoder to emit Ideogram4's native single-line JSON caption. +# docstyle-ignore +CAPTION_SYSTEM_MESSAGE = """You convert a short user idea into a structured JSON caption for an image renderer. Output ONE minified single-line JSON object and NOTHING else (no markdown, no commentary). + +SCHEMA — keys in this exact order: +{"high_level_description":"...","compositional_deconstruction":{"background":"...","elements":[ ... ]}} +- object element: {"type":"obj","desc":"..."} +- text element: {"type":"text","text":"VERBATIM CHARS","desc":"..."} + +STEP 1 — PICK THE MEDIUM. It decides what `background` and `elements` mean. Honor any medium or style the user implies; default to photograph only when nothing else fits. Render ANY subject faithfully — real, fantastical, sci-fi, surreal, abstract — in the chosen medium. + +A) DESIGNED ARTIFACT — poster, logo, album/book cover, flyer, banner, sticker, packaging, app icon, infographic, menu, card, wordmark. THE FRAME IS THE ARTIFACT — never a photo of it hanging in a room. + - high_level_description: name it as graphic design (e.g. "a minimalist jazz poster, flat graphic design..."). + - background: the design's OWN backdrop only — a flat color, gradient, or simple texture filling the frame. No room, wall, floor, easel, depth, or camera/photo language. + - elements: the design's parts as a flat 2D layout — a `text` element for every headline/label (verbatim), `obj` elements for the central graphic/illustration/shapes/badges. Place by region (top / center / bottom). + +B) SCENE — a photograph, illustration, painting, 3D render, anime frame, etc. of a real or imagined place or subject. + - high_level_description: one sentence naming the subject and the medium/style. + - background: the scene SHELL — surroundings, ground/sky/walls, atmosphere, ambient light; concrete and specific. The ground/floor/water surface lives here, never as an element. + - elements: the main subject FIRST as an `obj`, then supporting `obj` elements (props, secondary subjects) that plausibly belong. Add `text` elements only where the scene would really carry text (signs, labels, brands). + +C) ABSTRACT / CONCEPTUAL — "nostalgia", "chaos and order", "sound waves", pure pattern. Concretize the idea into a deliberate visual composition. + - background: the dominant color field, gradient, or texture of the composition. + - elements: the shapes, forms, motifs, or symbolic objects that carry the concept, as `obj` elements. Add `text` only if the idea calls for words. + +UNIVERSAL RULES (every medium): +1. The user's core subject/concept MUST appear among the elements (as an `obj`, normally first). Naming it only in high_level_description or background is NOT enough. +2. Commit to ONE concrete value each (one color, one style, one count). No hedging: ban "various", "such as", "e.g.", "or similar", "maybe", "X or Y" for one property. +3. NEVER use a transparent, empty, or plain white background UNLESS the user explicitly says "transparent", "isolated", "sticker", or "cutout". +4. A coherent subject (one animal, person, vehicle, object) is exactly ONE element; its parts go inside its `desc`. Use separate elements for genuinely separate subjects. +5. Each `desc` is 25-55 words, identity-first, standalone. Do not mention shadows, depth of field, bokeh, lens, focus, or grain. +6. high_level_description: one sentence, at most 40 words, starts with the subject, names the medium. Preserve non-ASCII characters as-is. +7. Output STRICTLY VALID JSON: double quotes around every key and string, NO trailing commas, each element object closes with "}" right after its last value. +8. Catch the "warm" impulse. Only when you are about to describe light as "warm", "golden", "amber", or "honey", stop and check: is there a specific physical source in the scene casting that colour (candle, sunset, lamp, neon, fire)? If YES, name the source and the colour it casts instead of the mood word. If NO, you are just reaching for warmth as ambience — drop it and leave the light neutral ("soft" or "even"). Don't recolour or relight anything else; this only intercepts the warm reach, every other scene and mood the user wants is untouched. +9. Describe physical reality, not impressions. Avoid mood-words — "luminous", "radiant", "vibrant", "lush", "dynamic", "gorgeous", "stunning", "breathtaking", "mesmerizing", and metaphorical "glowing" — they produce a generic AI look (the same trap as "warm"). Use observable properties: "the cheekbone catches a small highlight", not "luminous complexion". +10. Every named thing must appear as its own element. Each subject, object, sign, and quoted phrase the user names gets its own element — quoted text (single or double quotes) becomes its own verbatim `text` element. Count the named units in the prompt; the element list must hold at least that many. Don't drop or merge them. +11. Don't add what wasn't asked for. No glitch art, wireframe overlay, body fragmentation, double-exposure, "dissolving", or extra stylization unless the prompt requests it. Asked for a cinematic photo of a journalist → render that, not a glitch-art composite. +12. Name attributes concretely, anchored to landmarks. People: skin tone, hair (colour + style), each visible garment with colour, expression, pose, one distinguishing feature. Objects: shape, material, colour, a distinctive part. Place things against named references — "resting on the lower-right corner of the table", not "on the surface". +13. Name real references by name. If the user names a brand, product, character, place, or person (Nike Dunk Low, Spider-Man, the Eiffel Tower), keep that exact name in the `desc`; don't swap it for a generic look-alike unless they ask for an anonymous one. +14. "Professional photo/headshot" of a person means professional CONTEXT — neutral attire, soft even daylight, neutral backdrop, friendly expression — not dramatic studio gear; no heavy rim-light or creamy bokeh unless asked. + +EXAMPLES + +User idea: a cup of coffee on a table +Output: {"high_level_description":"A white ceramic cup of black coffee on a worn wooden cafe table, a casual overcast-daylight phone photograph with an off-center composition.","compositional_deconstruction":{"background":"Scratched oak cafe table filling the lower frame, a pale grey mortar-lined brick wall a few feet behind slightly out of focus, a tall window on the left spilling soft overcast daylight across the table, neutral white balance, muted brown and green tones.","elements":[{"type":"obj","desc":"White ceramic cup of black coffee with a thin curved handle turned to the right and a faint crema ring at the rim, resting on a matching round saucer near the center of the table, a thin wisp of steam at the surface."},{"type":"obj","desc":"Brushed-steel teaspoon lying on the saucer to the right of the cup, handle angled toward the lower-right corner, a single small water droplet on the bowl of the spoon."}]}} + +User idea: a minimalist poster for a jazz festival +Output: {"high_level_description":"A minimalist jazz festival poster, flat graphic design with bold typography and a single abstract saxophone motif on a deep teal background.","compositional_deconstruction":{"background":"Solid deep teal background filling the entire frame with a subtle fine paper-grain texture and a thin mustard-yellow keyline border just inside the edges, no scene and no depth.","elements":[{"type":"obj","desc":"A large flat geometric saxophone in mustard yellow and cream, centered in the upper two-thirds, built from simple bold shapes with no shading, angled diagonally from lower-left to upper-right."},{"type":"text","text":"JAZZ\\nFESTIVAL","desc":"Large bold condensed sans-serif headline in cream, stacked on two lines across the center of the poster, slightly overlapping the saxophone motif."},{"type":"text","text":"NOV 15 · CITY HALL","desc":"Small uppercase mustard-yellow caption centered near the bottom edge with wide letter spacing."}]}}""" + +# User turn. `{aspect_ratio}` and `{original_prompt}` are filled in by `Ideogram4Pipeline.upsample_prompt`. +# docstyle-ignore +CAPTION_USER_TEMPLATE = """TARGET IMAGE ASPECT RATIO: {aspect_ratio} (width:height). +User idea: {original_prompt}""" + + +def build_caption_logits_processor(model, tokenizer): + """Build an `outlines` logits processor that constrains generation to the Ideogram4 caption schema. + + Returns a logits processor compatible with `transformers` `generate(logits_processor=[...])`. The caller is + responsible for checking `is_outlines_available()` first; `outlines` (and its `pydantic` dependency) are imported + lazily here so they remain optional. The schema mirrors Ideogram's native caption / caption_verifier: a high-level + description plus a compositional deconstruction of background + typed elements. + """ + from typing import List, Literal, Union + + import outlines + from pydantic import BaseModel, Field + + class ObjElement(BaseModel): + type: Literal["obj"] + desc: str + + class TextElement(BaseModel): + type: Literal["text"] + text: str + desc: str + + class Composition(BaseModel): + background: str + elements: List[Union[ObjElement, TextElement]] = Field(min_length=1) + + class Caption(BaseModel): + high_level_description: str + compositional_deconstruction: Composition + + outlines_model = outlines.from_transformers(model, tokenizer) + return outlines.Generator(outlines_model, Caption).logits_processor + + +def generate_captions( + prompt_enhancer, + tokenizer, + logits_processor, + prompt: str | list[str], + height: int, + width: int, + temperature: float = PROMPT_UPSAMPLE_TEMPERATURE, + max_new_tokens: int = 1024, + generator: torch.Generator | list[torch.Generator] | None = None, + device: torch.device | None = None, +) -> list[str]: + """Rewrite each prompt into the native structured JSON caption with the grafted `prompt_enhancer`. + + Pass `generator` to make sampling reproducible (a seed is derived from it and used inside a forked RNG so the + caller's own RNG stream is untouched). + """ + device = device or prompt_enhancer.device + prompts = [prompt] if isinstance(prompt, str) else list(prompt) + divisor = math.gcd(width, height) or 1 + aspect_ratio = f"{width // divisor}:{height // divisor}" + + sampling_seed = None + if generator is not None: + gen = generator[0] if isinstance(generator, list) else generator + sampling_seed = int(torch.randint(0, 2**63 - 1, (1,), generator=gen, device=gen.device).item()) + fork_devices = [device] if getattr(device, "type", None) == "cuda" else [] + + captions = [] + for i, text_prompt in enumerate(prompts): + messages = [ + {"role": "system", "content": CAPTION_SYSTEM_MESSAGE}, + { + "role": "user", + "content": CAPTION_USER_TEMPLATE.format(aspect_ratio=aspect_ratio, original_prompt=text_prompt), + }, + ] + inputs = tokenizer.apply_chat_template( + messages, add_generation_prompt=True, tokenize=True, return_tensors="pt", return_dict=True + ).to(device) + generate_kwargs = { + "max_new_tokens": max_new_tokens, + "do_sample": temperature > 0, + "temperature": temperature, + "use_cache": True, + } + if logits_processor is not None: + logits_processor.reset() + generate_kwargs["logits_processor"] = [logits_processor] + with torch.random.fork_rng(devices=fork_devices, enabled=sampling_seed is not None): + if sampling_seed is not None: + torch.manual_seed(sampling_seed + i) + generated = prompt_enhancer.generate(**inputs, **generate_kwargs) + new_tokens = generated[:, inputs["input_ids"].shape[1] :] + captions.append(tokenizer.decode(new_tokens[0], skip_special_tokens=True).strip()) + return captions diff --git a/diffusers/pipelines/joyimage/__init__.py b/diffusers/pipelines/joyimage/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a5faea9d9763b66c776385bba63731a4d6b5f155 --- /dev/null +++ b/diffusers/pipelines/joyimage/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa: F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_joyimage_edit"] = ["JoyImageEditPipeline"] + _import_structure["pipeline_joyimage_edit_plus"] = ["JoyImageEditPlusPipeline"] + _import_structure["pipeline_output"] = ["JoyImageEditPipelineOutput", "JoyImageEditPlusPipelineOutput"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_joyimage_edit import JoyImageEditPipeline + from .pipeline_joyimage_edit_plus import JoyImageEditPlusPipeline + from .pipeline_output import JoyImageEditPipelineOutput, JoyImageEditPlusPipelineOutput +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/joyimage/image_processor.py b/diffusers/pipelines/joyimage/image_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..3aa7da1a0dccd792286102334ca25c98bd28842a --- /dev/null +++ b/diffusers/pipelines/joyimage/image_processor.py @@ -0,0 +1,149 @@ +# Copyright 2025 The JoyImage Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from typing import Tuple + +from PIL import Image + +from ...configuration_utils import register_to_config +from ...image_processor import VaeImageProcessor + + +# fmt: off +BUCKETS = { + 1024: [ + (512, 1792), (512, 1856), (512, 1920), (512, 1984), (512, 2048), + (576, 1600), (576, 1664), (576, 1728), (576, 1792), + (640, 1472), (640, 1536), (640, 1600), + (704, 1344), (704, 1408), (704, 1472), + (768, 1216), (768, 1280), (768, 1344), + (832, 1152), (832, 1216), + (896, 1088), (896, 1152), + (960, 1024), (960, 1088), + (1024, 960), (1024, 1024), + (1088, 896), (1088, 960), + (1152, 832), (1152, 896), + (1216, 768), (1216, 832), + (1280, 768), + (1344, 704), (1344, 768), + (1408, 704), + (1472, 640), (1472, 704), + (1536, 640), + (1600, 576), (1600, 640), + (1664, 576), + (1728, 576), + (1792, 512), (1792, 576), + (1856, 512), + (1920, 512), + (1984, 512), + (2048, 512), + ], +} +# fmt: on + + +def find_best_bucket(height: int, width: int, basesize: int) -> Tuple[int, int]: + """Return the (h, w) bucket whose aspect ratio is closest to height/width.""" + target_ratio = height / width + return min( + BUCKETS[basesize], + key=lambda hw: abs(hw[0] / hw[1] - target_ratio), + ) + + +class JoyImageEditImageProcessor(VaeImageProcessor): + """ + Image processor for the JoyImage Edit pipeline. + + Handles bucket-based resolution selection and resize-center-crop preprocessing. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to resize the image. + vae_scale_factor (`int`, *optional*, defaults to `8`): + VAE spatial scale factor. + basesize (`int`, *optional*, defaults to `1024`): + Base resolution for bucket generation. + resample (`str`, *optional*, defaults to `bilinear`): + Resampling filter for resizing. + do_normalize (`bool`, *optional*, defaults to `True`): + Whether to normalize the image to [-1,1]. + do_binarize (`bool`, *optional*, defaults to `False`): + Whether to binarize the image to 0/1. + do_convert_rgb (`bool`, *optional*, defaults to `False`): + Whether to convert the images to RGB format. + do_convert_grayscale (`bool`, *optional*, defaults to `False`): + Whether to convert the images to grayscale format. + """ + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 8, + basesize: int = 1024, + resample: str = "bilinear", + do_normalize: bool = True, + do_binarize: bool = False, + do_convert_rgb: bool = False, + do_convert_grayscale: bool = False, + ): + super().__init__() + + def get_default_height_width( + self, + image: Image.Image, + height: int | None = None, + width: int | None = None, + ) -> Tuple[int, int]: + if height is not None and width is not None: + src_w, src_h = width, height + elif image is None: + src_w, src_h = self.config.basesize, self.config.basesize + elif isinstance(image, list): + src_w, src_h = image[0].size + else: + src_w, src_h = image.size + + return find_best_bucket(src_h, src_w, self.config.basesize) + + def resize_center_crop( + self, + img, + target_size: Tuple[int, int], + ): + """ + Scale image to cover target_size, then center-crop. + + Args: + img: Input PIL image or list of PIL images. + target_size: (height, width) to crop to. + + Returns: + Resized and center-cropped PIL image(s), matching the input type. + """ + if isinstance(img, list): + return [self.resize_center_crop(i, target_size) for i in img] + + w, h = img.size + bh, bw = target_size + scale = max(bh / h, bw / w) + resize_h = math.ceil(h * scale) + resize_w = math.ceil(w * scale) + img = img.resize((resize_w, resize_h), Image.BILINEAR) + left = (resize_w - bw) // 2 + top = (resize_h - bh) // 2 + img = img.crop((left, top, left + bw, top + bh)) + return img diff --git a/diffusers/pipelines/joyimage/pipeline_joyimage_edit.py b/diffusers/pipelines/joyimage/pipeline_joyimage_edit.py new file mode 100644 index 0000000000000000000000000000000000000000..bf9f12a34c21aca7b399447d59db11961f226e6c --- /dev/null +++ b/diffusers/pipelines/joyimage/pipeline_joyimage_edit.py @@ -0,0 +1,877 @@ +import inspect +from typing import Callable, Dict, List, Optional, Union + +import numpy as np +import torch +from PIL import Image +from transformers import ( + Qwen2Tokenizer, + Qwen3VLForConditionalGeneration, + Qwen3VLProcessor, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...models import AutoencoderKLWan, JoyImageEditTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import JoyImageEditImageProcessor +from .pipeline_output import JoyImageEditPipelineOutput + + +EXAMPLE_DOC_STRING = """ +Examples: + ```python + >>> import torch + >>> from diffusers import JoyImageEditPipeline + >>> from diffusers.utils import load_image + + >>> model_id = "jdopensource/JoyAI-Image-Edit-Diffusers" + >>> pipe = JoyImageEditPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> image = load_image("https://huggingface.co/datasets/diffusers/docs-images/resolve/main/astronaut.jpg") + >>> output = pipe( + ... image=image, # pass an image for editing; omit for text-to-image generation + ... prompt="Add wings to the astronaut.", + ... num_inference_steps=40, + ... guidance_scale=4.0, + ... generator=torch.manual_seed(0), + ... ) + >>> output.images[0].save("joyimage_edit.png") + ``` +""" + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + """ + Configure the scheduler and return its timestep sequence. + + Exactly one of ``timesteps``, ``sigmas``, or ``num_inference_steps`` should be provided to control the denoising + schedule. + + Args: + scheduler: The diffusion scheduler. + num_inference_steps: Number of denoising steps (used when neither + ``timesteps`` nor ``sigmas`` is given). + device: Target device for the timestep tensor. + timesteps: Custom discrete timesteps. + sigmas: Custom sigma values (alternative to ``timesteps``). + **kwargs: Additional keyword arguments forwarded to ``set_timesteps``. + + Returns: + Tuple of (timesteps tensor, num_inference_steps int). + + Raises: + ValueError: If both ``timesteps`` and ``sigmas`` are provided, or if the + scheduler does not support the requested schedule parameterisation. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed.") + + if timesteps is not None: + if "timesteps" not in set(inspect.signature(scheduler.set_timesteps).parameters.keys()): + raise ValueError(f"{scheduler.__class__} does not support custom timesteps.") + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + if "sigmas" not in set(inspect.signature(scheduler.set_timesteps).parameters.keys()): + raise ValueError(f"{scheduler.__class__} does not support custom sigmas.") + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + + return timesteps, num_inference_steps + + +class JoyImageEditPipeline(DiffusionPipeline): + """ + Diffusion pipeline for image editing using the JoyImage architecture. + + The pipeline encodes text and image conditioning via a Qwen3-VL text encoder, denoises latents with a 3-D + transformer, and decodes the result with a WAN VAE. + + Model offloading order: text_encoder -> transformer -> vae. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLWan, + text_encoder: Qwen3VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: JoyImageEditTransformer3DModel, + processor: Qwen3VLProcessor, + text_token_max_length: int = 2048, + ): + """ + Initialise the pipeline and register all sub-modules. + + Args: + scheduler: Noise scheduler for the denoising process. + vae: Variational autoencoder used for encoding / decoding latents. + text_encoder: Qwen3-VL multimodal language model for prompt encoding. + tokenizer: Tokenizer paired with the text encoder. + transformer: 3-D transformer denoising network. + processor: Qwen3-VL processor for multi-image prompt preparation. + text_token_max_length: Maximum number of text tokens for the encoder. + """ + super().__init__() + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + processor=processor, + ) + + self.text_token_max_length = text_token_max_length + + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.vae_image_processor = JoyImageEditImageProcessor( + vae_scale_factor=self.vae_scale_factor_spatial, + ) + + # Prompt templates used when encoding text with / without image tokens. + self.prompt_template_encode = { + "image": ( + "<|im_start|>system\n \\nDescribe the image by detailing the color, shape, size, texture, " + "quantity, text, spatial relationships of the objects and background:<|im_end|>\n" + "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + ), + "multiple_images": ( + "<|im_start|>system\n \\nDescribe the image by detailing the color, shape, size, texture, " + "quantity, text, spatial relationships of the objects and background:<|im_end|>\n" + "{}<|im_start|>assistant\n" + ), + } + # Number of system-prompt tokens to drop from the beginning of hidden states. + self.prompt_template_encode_start_idx = { + "image": 34, + "multiple_images": 34, + } + + # ------------------------------------------------------------------ + # Internal helpers + # ------------------------------------------------------------------ + + def _get_last_decoder_hidden_states(self, forward_fn, **kwargs): + """ + Run ``forward_fn(**kwargs)`` while capturing the **pre-norm** output of the last decoder layer via a forward + hook. + + This model was trained on transformers 4.57, where ``Qwen3VLForConditionalGeneration``'s + ``@check_model_inputs`` decorator monkey-patched each decoder layer to collect ``hidden_states``. Because + ``Qwen3VLCausalLMOutputWithPast`` has no ``last_hidden_state`` field, ``tie_last_hidden_states`` had no effect + and ``hidden_states[-1]`` was the **pre-norm** output of the last decoder layer. + + Starting from https://github.com/huggingface/transformers/pull/42609 the CausalLM forward explicitly returns + ``hidden_states=outputs.hidden_states`` from the inner model. Combined with the subsequent + ``@check_model_inputs`` → ``@capture_outputs`` migration (transformers 5.x), ``hidden_states`` is now captured + at the ``Qwen3VLTextModel`` level where ``tie_last_hidden_states=True`` replaces ``hidden_states[-1]`` with the + **post-norm** ``last_hidden_state``. The CausalLM simply passes this through, so ``hidden_states[-1]`` becomes + post-norm – a ~10× scale difference (std ≈ 2 vs ≈ 21) that breaks inference. + + This helper bypasses both mechanisms by hooking the last decoder layer directly, returning the raw pre-norm + output regardless of the transformers version. + """ + captured = {} + + def _hook(_module, _input, output): + captured["hidden_states"] = output[0] if isinstance(output, tuple) else output + + handle = self.text_encoder.model.language_model.layers[-1].register_forward_hook(_hook) + try: + forward_fn(**kwargs) + finally: + handle.remove() + return captured["hidden_states"] + + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor) -> tuple[torch.Tensor, ...]: + """ + Extract valid (non-padded) hidden states for each sequence in the batch. + + Args: + hidden_states: Shape (B, T, D). + mask: Binary attention mask of shape (B, T). + + Returns: + Tuple of tensors, one per batch element, each of shape (valid_T, D). + """ + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + return torch.split(selected, valid_lengths.tolist(), dim=0) + + def _get_qwen_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + template_type: str = "image", + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + """ + Encode text prompts using the Qwen tokenizer (text-only path). + + Args: + prompt: A single prompt string or a list of prompt strings. + template_type: Key into ``prompt_template_encode`` / ``prompt_template_encode_start_idx``. + device: Target device. + dtype: Target floating-point dtype. + + Returns: + Tuple of (prompt_embeds, encoder_attention_mask) where both tensors have shape (B, max_seq_len, D) and (B, + max_seq_len) respectively, zero-padded to the same length. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + template = self.prompt_template_encode[template_type] + drop_idx = self.prompt_template_encode_start_idx[template_type] + + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, + max_length=self.text_token_max_length + drop_idx, + padding=True, + truncation=True, + return_tensors="pt", + ).to(device) + + hidden_states = self._get_last_decoder_hidden_states( + self.text_encoder, + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + ) + + # Drop system-prompt prefix tokens and re-pack into a padded batch. + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + + max_seq_len = min( + self.text_token_max_length, + max(u.size(0) for u in split_hidden_states), + max(u.size(0) for u in attn_mask_list), + ) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + return prompt_embeds, encoder_attention_mask + + def encode_prompt_multiple_images( + self, + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + images: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + prompt_embeds_mask: Optional[torch.Tensor] = None, + template_type: Optional[str] = "multiple_images", + max_sequence_length: Optional[int] = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + """ + Encode prompts that contain inline image tokens via the Qwen processor. + + ``\\n`` placeholders in each prompt string are replaced by the Qwen vision special tokens before being + fed to the multimodal encoder. + + Args: + prompt: Prompt string(s), optionally containing ``\\n`` tokens. + device: Target device. + num_images_per_prompt: Number of outputs to generate per prompt. + images: Pixel tensors corresponding to the inline image tokens. + prompt_embeds: Pre-computed prompt embeddings. + prompt_embeds_mask: Attention mask for pre-computed embeddings. + template_type: Must be ``"multiple_images"``. + max_sequence_length: If set, truncate the output to this length + (keeping the last ``max_sequence_length`` tokens). + + Returns: + Tuple of (prompt_embeds, prompt_embeds_mask). + """ + if template_type != "multiple_images": + raise ValueError(f"Expected template_type 'multiple_images', but got '{template_type}'") + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + template = self.prompt_template_encode[template_type] + drop_idx = self.prompt_template_encode_start_idx[template_type] + + prompt = [f"\n{p}" for p in prompt] + prompt = [f"<|im_start|>user\n{p}<|im_end|>\n" for p in prompt] + + prompt = [p.replace("\n", "<|vision_start|><|image_pad|><|vision_end|>") for p in prompt] + prompt = [template.format(p) for p in prompt] + + if images is not None: + if not isinstance(images, list): + images = [images] * len(prompt) + elif len(images) < len(prompt) and len(prompt) % len(images) == 0: + images = images * (len(prompt) // len(images)) + + inputs = self.processor( + text=prompt, + images=images, + padding=True, + return_tensors="pt", + ).to(device) + + last_hidden_states = self._get_last_decoder_hidden_states(self.text_encoder, **inputs) + + prompt_embeds = last_hidden_states[:, drop_idx:] + prompt_embeds_mask = inputs["attention_mask"][:, drop_idx:] + + if max_sequence_length is not None and prompt_embeds.shape[1] > max_sequence_length: + prompt_embeds = prompt_embeds[:, -max_sequence_length:, :] + prompt_embeds_mask = prompt_embeds_mask[:, -max_sequence_length:] + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + return prompt_embeds, prompt_embeds_mask + + def encode_prompt( + self, + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + prompt_embeds_mask: Optional[torch.Tensor] = None, + max_sequence_length: int = 1024, + template_type: str = "image", + ) -> tuple[torch.Tensor, torch.Tensor]: + """ + Encode a text prompt into embeddings (text-only path). + + Pre-computed ``prompt_embeds`` bypass encoding entirely. + + Args: + prompt: Prompt string or list of prompt strings. + device: Target device. + num_images_per_prompt: Number of outputs to generate per prompt. + prompt_embeds: Pre-computed prompt embeddings. + prompt_embeds_mask: Attention mask for pre-computed embeddings. + max_sequence_length: Maximum output sequence length. + template_type: Prompt template key (``"image"`` or ``"multiple_images"``). + + Returns: + Tuple of (prompt_embeds, prompt_embeds_mask). + """ + device = device or self._execution_device + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, template_type, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + ): + """ + Validate pipeline inputs before the forward pass. + + Raises: + ValueError: On any invalid combination of arguments. + """ + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError("`callback_on_step_end_tensor_inputs` has invalid keys.") + + if prompt is not None and prompt_embeds is not None: + raise ValueError("Cannot forward both `prompt` and `prompt_embeds`.") + elif prompt is None and prompt_embeds is None: + raise ValueError("Provide either `prompt` or `prompt_embeds`.") + elif prompt is not None and not isinstance(prompt, (str, list)): + raise ValueError("`prompt` has to be of type `str` or `list`.") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError("Cannot forward both `negative_prompt` and `negative_prompt_embeds`.") + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError("If `prompt_embeds` are provided, `prompt_embeds_mask` is required.") + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError("If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` is required.") + + def normalize_latents(self, latent: torch.Tensor) -> torch.Tensor: + """ + Normalise latents using per-channel statistics from the VAE config. + + Uses (latent - mean) / std when the VAE exposes ``latents_mean`` and ``latents_std``; otherwise falls back to + scaling by ``scaling_factor``. + + Args: + latent: Raw latent tensor from ``vae.encode``. + + Returns: + Normalised latent tensor. + """ + if hasattr(self.vae.config, "latents_mean") and hasattr(self.vae.config, "latents_std"): + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, -1, 1, 1, 1) + .to(device=latent.device, dtype=latent.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, -1, 1, 1, 1) + .to(device=latent.device, dtype=latent.dtype) + ) + latent = (latent - latents_mean) / latents_std + else: + latent = latent * self.vae.config.scaling_factor + return latent + + def denormalize_latents(self, latent: torch.Tensor) -> torch.Tensor: + """ + Invert :meth:`normalize_latents` to recover the original latent scale. + + Args: + latent: Normalised latent tensor. + + Returns: + Latent tensor in the scale expected by ``vae.decode``. + """ + if hasattr(self.vae.config, "latents_mean") and hasattr(self.vae.config, "latents_std"): + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, -1, 1, 1, 1) + .to(device=latent.device, dtype=latent.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, -1, 1, 1, 1) + .to(device=latent.device, dtype=latent.dtype) + ) + latent = latent * latents_std + latents_mean + else: + latent = latent / self.vae.config.scaling_factor + return latent + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + video_length: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[Union[torch.Generator, List[torch.Generator]]], + latents: Optional[torch.Tensor] = None, + image: Optional[List[Image.Image]] = None, + enable_denormalization: bool = True, + ) -> tuple[torch.Tensor, Optional[torch.Tensor]]: + """ + Prepare the initial noisy latent tensor for the denoising loop. + + Args: + batch_size: Number of samples in the batch. + num_channels_latents: Latent channel dimension from the transformer config. + height: Spatial height in pixels. + width: Spatial width in pixels. + video_length: Number of frames (1 for image inference). + dtype: Floating-point dtype for the latent tensor. + device: Target device. + generator: RNG generator(s) for reproducible sampling. + latents: Optional user-provided initial noise for the target slot. When ``None`` random noise is sampled. + image: Optional list of PIL reference images to VAE-encode as conditioning slots. + enable_denormalization: Whether to normalise encoded reference latents. + + Returns: + Tuple of ``(latents, image_latents)`` where ``latents`` has shape ``(B, 1, C, T, H', W')`` and + ``image_latents`` has shape ``(B, N_ref, C, T, H', W')`` or ``None`` when no reference images are given. + + Raises: + ValueError: If ``generator`` is a list whose length differs from ``batch_size``. + """ + noise_shape = ( + batch_size, + 1, + num_channels_latents, + (video_length - 1) // self.vae_scale_factor_temporal + 1, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError("Generator list length must match batch size.") + + if latents is None: + latents = randn_tensor(noise_shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + image_latents = None + if image is not None: + if batch_size > len(image) and batch_size % len(image) == 0: + image = image * (batch_size // len(image)) + elif batch_size > len(image): + raise ValueError(f"Cannot duplicate `image` of batch size {len(image)} to {batch_size} text prompts.") + ref_img = [torch.from_numpy(np.array(x.convert("RGB"))) for x in image] + ref_img = torch.stack(ref_img).to(device=device, dtype=dtype) + ref_img = ref_img / 127.5 - 1.0 + ref_img = ref_img.permute(0, 3, 1, 2).unsqueeze(2) + image_latents = self.vae.encode(ref_img).latent_dist.sample() + if enable_denormalization: + image_latents = self.normalize_latents(image_latents) + image_latents = image_latents.unsqueeze(1) # (B, 1, C, T, H', W') + + return latents, image_latents + + # ------------------------------------------------------------------ + # Pipeline properties + # ------------------------------------------------------------------ + + @property + def guidance_scale(self) -> float: + """Classifier-free guidance scale used in the current forward pass.""" + return self._guidance_scale + + @property + def do_classifier_free_guidance(self) -> bool: + """True when guidance_scale > 1, enabling classifier-free guidance.""" + return self._guidance_scale > 1 + + @property + def num_timesteps(self) -> int: + """Total number of denoising timesteps in the current forward pass.""" + return self._num_timesteps + + @property + def interrupt(self) -> bool: + """When True, the denoising loop is interrupted at the next step.""" + return self._interrupt + + # ------------------------------------------------------------------ + # Forward pass + # ------------------------------------------------------------------ + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput | None = None, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 40, + timesteps: List[int] = None, + sigmas: List[float] = None, + guidance_scale: float = 4.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + prompt_embeds_mask: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds_mask: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback_on_step_end: Optional[ + Union[ + Callable[[int, int, Dict], None], + PipelineCallback, + MultiPipelineCallbacks, + ] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 4096, + enable_denormalization: bool = True, + ): + r""" + Generate an edited image conditioned on a reference image and a text prompt. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide generation. + height (`int`): + Height of the generated output in pixels. + width (`int`): + Width of the generated output in pixels. + image (`PipelineImageInput`, *optional*): + Reference image used for conditioning. When provided the pipeline operates in image-editing mode with + ``num_items=2``. + num_inference_steps (`int`, *optional*, defaults to 40): + Number of denoising steps. More steps generally improve quality at the cost of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps for the denoising process. When provided, ``num_inference_steps`` is inferred from the + list length. + sigmas (`List[float]`, *optional*): + Custom sigmas for the denoising process. Mutually exclusive with ``timesteps``. + guidance_scale (`float`, *optional*, defaults to 4.0): + Classifier-free guidance scale. + negative_prompt (`str` or `List[str]`, *optional*): + Negative prompt(s) used to suppress undesired content. + num_images_per_prompt (`int`, *optional*, defaults to 1): + Number of generated samples per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + RNG generator(s) for deterministic sampling. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents for the target slot. Sampled from a Gaussian distribution when not + provided. Can be used to seed generation from a specific starting noise tensor. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-computed prompt embeddings. When provided ``prompt`` can be omitted. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for ``prompt_embeds``. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-computed negative prompt embeddings. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for ``negative_prompt_embeds``. + output_type (`str`, *optional*, defaults to ``"pil"``): + Output format. Pass ``"latent"`` to return raw latents. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a :class:`JoyImageEditPipelineOutput` or a plain tensor. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + Callback invoked at the end of each denoising step with signature ``(self, step: int, timestep: int, + callback_kwargs: Dict)``. + callback_on_step_end_tensor_inputs (`List[str]`, *optional*, defaults to ``["latents"]``): + Tensor keys included in ``callback_kwargs`` for ``callback_on_step_end``. + max_sequence_length (`int`, *optional*, defaults to 4096): + Maximum sequence length for prompt encoding. + enable_denormalization (`bool`, *optional*, defaults to `True`): + Denormalise latents before VAE decoding. + + Examples: + + Returns: + [`~pipelines.joyimage.JoyImageEditPipelineOutput`] or `torch.Tensor`: + If ``return_dict`` is ``True``, returns a pipeline output object containing the generated image(s). + Otherwise returns the image tensor directly. + """ + # Resize the input image to the nearest bucket resolution. + # Or resize the specified height and width to the nearest bucket resolution. + height, width = self.vae_image_processor.get_default_height_width(image, height, width) + processed_image = None + if image is not None: + processed_image = self.vae_image_processor.resize_center_crop(image, (height, width)) + + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._interrupt = False + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # num_items: 1 for unconditional generation, 2 for reference-image editing. + num_items = 1 if image is None else 2 + + # Encode the conditioning prompt. + if processed_image is not None: + prompt_embeds, prompt_embeds_mask = self.encode_prompt_multiple_images( + prompt=prompt, + images=processed_image, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + else: + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + if self.do_classifier_free_guidance: + # Build default negative prompts when none are provided. + if negative_prompt is None and negative_prompt_embeds is None: + negative_prompt = [""] * batch_size + + if processed_image is not None: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt_multiple_images( + prompt=negative_prompt, + images=processed_image, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + else: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas, + ) + + num_channels_latents = self.transformer.config.in_channels + noise_latents, image_latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + 1, # video_length = 1 for image inference + prompt_embeds.dtype, + device, + generator, + latents, + image=( + (processed_image if isinstance(processed_image, list) else [processed_image]) + if processed_image is not None + else None + ), + enable_denormalization=enable_denormalization, + ) + + if image_latents is not None: + latents = torch.cat([image_latents, noise_latents], dim=1) + else: + latents = noise_latents + + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # Restore reference latents so they are never overwritten by the scheduler. + if image_latents is not None: + latents[:, : (num_items - 1)] = image_latents + + latent_model_input = latents + t_expand = t.repeat(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=t_expand, + encoder_hidden_states=prompt_embeds, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_pred_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=t_expand, + encoder_hidden_states=negative_prompt_embeds, + return_dict=False, + )[0] + + comb_pred = noise_pred_uncond + self.guidance_scale * (noise_pred - noise_pred_uncond) + # Rescale to match the conditional prediction norm (guidance rescaling). + cond_norm = torch.norm(noise_pred, dim=2, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=2, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm.clamp_min(1e-6)) + + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + if progress_bar is not None: + progress_bar.update() + + if output_type != "latent": + latents = latents.flatten(0, 1) + if enable_denormalization: + latents = self.denormalize_latents(latents) + + image = self.vae.decode(latents, return_dict=False)[0] + image = image.unflatten(0, (batch_size * num_images_per_prompt, -1)) + else: + image = latents + + # Extract the target slot (last item) from each batch element. + # (B, num_items, C, T, H, W) -> permute -> (B, num_items, T, C, H, W) -> [:, -1] -> (B, T, C, H, W) + image = image.float().permute(0, 1, 3, 2, 4, 5)[:, -1].squeeze(1) + + image = self.image_processor.postprocess(image, output_type=output_type) + + self.maybe_free_model_hooks() + + if not return_dict: + return image + + return JoyImageEditPipelineOutput(images=image) diff --git a/diffusers/pipelines/joyimage/pipeline_joyimage_edit_plus.py b/diffusers/pipelines/joyimage/pipeline_joyimage_edit_plus.py new file mode 100644 index 0000000000000000000000000000000000000000..ac8e01278e3eae93436d2fe9e3ff4c7575915fa3 --- /dev/null +++ b/diffusers/pipelines/joyimage/pipeline_joyimage_edit_plus.py @@ -0,0 +1,770 @@ +# Copyright 2025 The JoyImage Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import torch +from PIL import Image +from transformers import ( + Qwen2Tokenizer, + Qwen3VLForConditionalGeneration, + Qwen3VLProcessor, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...models import AutoencoderKLWan +from ...models.transformers.transformer_joyimage_edit_plus import JoyImageEditPlusTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import JoyImageEditImageProcessor +from .pipeline_output import JoyImageEditPlusPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ +Examples: + ```python + >>> import torch + >>> from diffusers import JoyImageEditPlusPipeline + >>> from diffusers.utils import load_image + + >>> model_id = "jdopensource/JoyAI-Image-Edit-Plus-Diffusers" + >>> pipe = JoyImageEditPlusPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> images = [ + ... load_image("dog.png"), + ... load_image("person.png"), + ... ] + >>> output = pipe( + ... images=images, + ... prompt="Let the person lovingly play with the dog.", + ... height=1024, + ... width=1024, + ... num_inference_steps=30, + ... guidance_scale=4.0, + ... generator=torch.manual_seed(42), + ... ) + >>> output.images[0].save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class JoyImageEditPlusPipeline(DiffusionPipeline): + r""" + Diffusion pipeline for multi-image instruction-guided editing using JoyImage Edit Plus. + + Supports multiple reference images with different resolutions. Each reference image is independently VAE-encoded + and patchified, then concatenated with the target noise patches for joint denoising. + + Args: + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`Qwen3VLForConditionalGeneration`]): + Multimodal text encoder for prompt encoding with inline image understanding. + tokenizer ([`Qwen2Tokenizer`]): + Tokenizer for text processing. + transformer ([`JoyImageEditPlusTransformer3DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + processor ([`Qwen3VLProcessor`]): + Processor for multimodal inputs (text + images). + text_token_max_length (`int`, defaults to `2048`): + Maximum token length for text encoding. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLWan, + text_encoder: Qwen3VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: JoyImageEditPlusTransformer3DModel, + processor: Qwen3VLProcessor, + text_token_max_length: int = 2048, + ): + super().__init__() + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + processor=processor, + ) + + self.text_token_max_length = text_token_max_length + + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8 + self.image_processor = JoyImageEditImageProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + self.prompt_template_encode = { + "multiple_images": ( + "<|im_start|>system\n \\nDescribe the image by detailing the color, shape, size, texture, " + "quantity, text, spatial relationships of the objects and background:<|im_end|>\n" + "{}<|im_start|>assistant\n" + ), + } + self.prompt_template_encode_start_idx = { + "multiple_images": 34, + } + + # ------------------------------------------------------------------ + # Internal helpers + # ------------------------------------------------------------------ + + def _get_last_decoder_hidden_states(self, forward_fn, **kwargs): + """ + Run ``forward_fn(**kwargs)`` while capturing the **pre-norm** output of the last decoder layer via a forward + hook. + + This model was trained on transformers 4.57, where ``Qwen3VLForConditionalGeneration``'s + ``@check_model_inputs`` decorator monkey-patched each decoder layer to collect ``hidden_states``. Because + ``Qwen3VLCausalLMOutputWithPast`` has no ``last_hidden_state`` field, ``tie_last_hidden_states`` had no effect + and ``hidden_states[-1]`` was the **pre-norm** output of the last decoder layer. + + Starting from https://github.com/huggingface/transformers/pull/42609 the CausalLM forward explicitly returns + ``hidden_states=outputs.hidden_states`` from the inner model. Combined with the subsequent + ``@check_model_inputs`` → ``@capture_outputs`` migration (transformers 5.x), ``hidden_states`` is now captured + at the ``Qwen3VLTextModel`` level where ``tie_last_hidden_states=True`` replaces ``hidden_states[-1]`` with the + **post-norm** ``last_hidden_state``. The CausalLM simply passes this through, so ``hidden_states[-1]`` becomes + post-norm – a ~10x scale difference (std ~2 vs ~21) that breaks inference. + + This helper bypasses both mechanisms by hooking the last decoder layer directly, returning the raw pre-norm + output regardless of the transformers version. + """ + captured = {} + + def _hook(_module, _input, output): + captured["hidden_states"] = output[0] if isinstance(output, tuple) else output + + handle = self.text_encoder.model.language_model.layers[-1].register_forward_hook(_hook) + try: + forward_fn(**kwargs) + finally: + handle.remove() + return captured["hidden_states"] + + def encode_prompt_multiple_images( + self, + prompt: str | list[str], + device: torch.device | None = None, + images: list[Image.Image] | None = None, + max_sequence_length: int | None = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + """Encode prompts with inline tokens via the Qwen3-VL processor.""" + device = device or self._execution_device + template = self.prompt_template_encode["multiple_images"] + drop_idx = self.prompt_template_encode_start_idx["multiple_images"] + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [p.replace("\n", "<|vision_start|><|image_pad|><|vision_end|>") for p in prompt] + prompt = [template.format(p) for p in prompt] + + inputs = self.processor( + text=prompt, + images=images, + padding=True, + return_tensors="pt", + ).to(device) + + last_hidden_states = self._get_last_decoder_hidden_states(self.text_encoder, **inputs) + + prompt_embeds = last_hidden_states[:, drop_idx:] + prompt_embeds_mask = inputs["attention_mask"][:, drop_idx:] + + if max_sequence_length is not None and prompt_embeds.shape[1] > max_sequence_length: + prompt_embeds = prompt_embeds[:, -max_sequence_length:, :] + prompt_embeds_mask = prompt_embeds_mask[:, -max_sequence_length:] + + return prompt_embeds, prompt_embeds_mask + + def _pad_sequence(self, x: torch.Tensor, target_length: int) -> torch.Tensor: + current_length = x.shape[1] + if current_length >= target_length: + return x[:, -target_length:] + padding_length = target_length - current_length + if x.ndim >= 3: + padding = torch.zeros((x.shape[0], padding_length, *x.shape[2:]), dtype=x.dtype, device=x.device) + else: + padding = torch.zeros((x.shape[0], padding_length), dtype=x.dtype, device=x.device) + return torch.cat([x, padding], dim=1) + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: torch.Generator | list[torch.Generator] | None, + reference_images: list[list[Image.Image]] | None = None, + latents: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, list[list[tuple[int, int, int]]]]: + """Prepare 6D padded latent tensor with target noise + reference image latents. + + Args: + latents: Optional pre-computed noise for the target slot. Shape ``(B, C, 1, H', W')`` where + ``H'`` and ``W'`` are the latent-space dimensions. When ``None``, random noise is sampled. + + Returns: + padded_latents: [B, max_patches, C, pt, ph, pw] target_mask: [B, max_patches] (True for target patches) + shape_list: per-sample list of (t, h, w) tuples for each component + """ + pt, ph, pw = self.transformer.config.patch_size + + all_patches = [] + all_target_masks = [] + all_shape_lists = [] + max_patches = 0 + + for i in range(batch_size): + sample_gen = generator[i] if isinstance(generator, list) else generator + + # Target noise + t_target = 1 + h_target = int(height) // self.vae_scale_factor_spatial + w_target = int(width) // self.vae_scale_factor_spatial + if latents is None: + noise_shape = (num_channels_latents, t_target, h_target, w_target) + noise_block = randn_tensor(noise_shape, generator=sample_gen, device=device, dtype=dtype) + else: + noise_block = latents[i].to(device=device, dtype=dtype) + + sample_items = [noise_block] + + # Reference images + if reference_images is not None and reference_images[i]: + for ref_img_pil in reference_images[i]: + ref_tensor = self.image_processor.preprocess(ref_img_pil).to(device=device, dtype=dtype) + ref_tensor = ref_tensor.unsqueeze(2) # [B, C, H, W] -> [B, C, 1, H, W] + + ref_latent = self.vae.encode(ref_tensor.to(self.vae.dtype)).latent_dist.mode() + ref_latent = ref_latent.to(dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, -1, 1, 1, 1) + .to(ref_latent.device, ref_latent.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, -1, 1, 1, 1) + .to(ref_latent.device, ref_latent.dtype) + ) + ref_latent = (ref_latent - latents_mean) / latents_std + ref_latent = ref_latent.squeeze(0) # [C, 1, H', W'] + sample_items.append(ref_latent) + + # Patchify each item and build shape_list + sample_patches = [] + sample_masks = [] + sample_shapes = [] + + for j, item in enumerate(sample_items): + c, t, h, w = item.shape + l_t, l_h, l_w = t // pt, h // ph, w // pw + sample_shapes.append((l_t, l_h, l_w)) + + patches = item.reshape(c, l_t, pt, l_h, ph, l_w, pw) + patches = patches.permute(1, 3, 5, 0, 2, 4, 6).reshape(-1, c, pt, ph, pw) + sample_patches.append(patches) + sample_masks.append(torch.full((patches.shape[0],), j == 0, device=device, dtype=torch.bool)) + + combined_patches = torch.cat(sample_patches, dim=0) + combined_masks = torch.cat(sample_masks, dim=0) + + all_patches.append(combined_patches) + all_target_masks.append(combined_masks) + all_shape_lists.append(sample_shapes) + max_patches = max(max_patches, combined_patches.shape[0]) + + # Pad to uniform size + padded_latents = torch.zeros( + (batch_size, max_patches, num_channels_latents, pt, ph, pw), device=device, dtype=dtype + ) + target_mask = torch.zeros((batch_size, max_patches), device=device, dtype=torch.bool) + + for i in range(batch_size): + n = all_patches[i].shape[0] + padded_latents[i, :n] = all_patches[i] + target_mask[i, :n] = all_target_masks[i] + + return padded_latents, target_mask, all_shape_lists + + # ------------------------------------------------------------------ + # Properties + # ------------------------------------------------------------------ + + @property + def guidance_scale(self) -> float: + return self._guidance_scale + + @property + def do_classifier_free_guidance(self) -> bool: + return self._guidance_scale > 1 + + @property + def num_timesteps(self) -> int: + return self._num_timesteps + + @property + def interrupt(self) -> bool: + return self._interrupt + + # ------------------------------------------------------------------ + # Forward pass + # ------------------------------------------------------------------ + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height is not None and height % self.vae_scale_factor_spatial != 0: + raise ValueError(f"`height` must be divisible by {self.vae_scale_factor_spatial} but is {height}.") + if width is not None and width % self.vae_scale_factor_spatial != 0: + raise ValueError(f"`width` must be divisible by {self.vae_scale_factor_spatial} but is {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found" + f" {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + images: list[Image.Image] | list[list[Image.Image]] | None = None, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 30, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 4.0, + negative_prompt: str | list[str] | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int, dict], None] + | PipelineCallback + | MultiPipelineCallbacks + | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 4096, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + images (`list[Image.Image]` or `list[list[Image.Image]]`, *optional*): + Reference images for editing. Each image can have a different resolution. If a flat list is provided, + it is treated as one sample with multiple references. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + height (`int`, *optional*): + The height in pixels of the generated image. If `None`, determined from the last reference image. + width (`int`, *optional*): + The width in pixels of the generated image. If `None`, determined from the last reference image. + num_inference_steps (`int`, *optional*, defaults to `30`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spacing is used. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process. + guidance_scale (`float`, *optional*, defaults to `4.0`): + Classifier-free guidance scale. Higher values encourage the model to generate images more aligned with + the `prompt` at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, a blank prompt is used for + classifier-free guidance. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents to be used as inputs for image generation. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for pre-generated text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for pre-generated negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `"pil"` (`PIL.Image.Image`), `"np"` + (`np.ndarray`), `"pt"` (`torch.Tensor`), or `"latent"` for raw latent output. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`JoyImageEditPlusPipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function called at the end of each denoising step with arguments: the pipeline, step index, timestep, + and a dict of callback tensor inputs. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*, defaults to `["latents"]`): + The list of tensor inputs for the `callback_on_step_end` function. + max_sequence_length (`int`, *optional*, defaults to `4096`): + Maximum sequence length for the text encoder. + + Examples: + + Returns: + [`JoyImageEditPlusPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`JoyImageEditPlusPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list of generated images. + """ + # Normalize images input to List[List[Image]] + if images is not None: + if isinstance(images[0], Image.Image): + images = [images] # single sample + + self.check_inputs( + prompt=prompt, + height=height, + width=width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._interrupt = False + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Determine output resolution from last reference image if not specified + if height is None or width is None: + if images is not None and len(images[0]) > 0: + last_img = images[0][-1] + height, width = self.image_processor.get_default_height_width(last_img) + else: + height = height or 1024 + width = width or 1024 + + device = self._execution_device + + # Pre-process images: bucket-resize each reference image (matching original pipeline) + if images is not None: + processed_images = [] + for sample_imgs in images: + processed_sample = [] + for img in sample_imgs: + ref_h, ref_w = self.image_processor.get_default_height_width(img) + resize_img = self.image_processor.resize_center_crop(img, (ref_h, ref_w)) + processed_sample.append(resize_img) + processed_images.append(processed_sample) + images = processed_images + + # Construct prompts with tokens + prompt = [prompt] if isinstance(prompt, str) else prompt + if images is not None: + formatted_prompts = [] + for i in range(batch_size): + num_refs = len(images[i]) if i < len(images) else 0 + image_tags = "".join(["\n" for _ in range(num_refs)]) + p = prompt[i] if i < len(prompt) else prompt[0] + formatted_prompts.append(f"<|im_start|>user\n{image_tags}{p}<|im_end|>\n") + else: + formatted_prompts = [f"<|im_start|>user\n{p}<|im_end|>\n" for p in prompt] + + # Flatten all images for the processor + flattened_images = None + if images is not None: + flattened_images = [img for sublist in images for img in sublist] + + # Encode prompt + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self.encode_prompt_multiple_images( + prompt=formatted_prompts, + images=flattened_images, + device=device, + max_sequence_length=max_sequence_length, + ) + + # Encode negative prompt for CFG + if self.do_classifier_free_guidance: + if negative_prompt is None and negative_prompt_embeds is None: + neg_prompts = [] + for i in range(batch_size): + num_refs = len(images[i]) if images is not None and i < len(images) else 0 + image_tags = "".join(["\n" for _ in range(num_refs)]) + neg_prompts.append(f"<|im_start|>user\n{image_tags} <|im_end|>\n") + negative_prompt = neg_prompts + elif negative_prompt is not None and negative_prompt_embeds is None: + neg_list = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + neg_prompts = [] + for i in range(batch_size): + num_refs = len(images[i]) if images is not None and i < len(images) else 0 + image_tags = "".join(["\n" for _ in range(num_refs)]) + n = neg_list[i] if i < len(neg_list) else neg_list[0] + neg_prompts.append(f"<|im_start|>user\n{image_tags}{n}<|im_end|>\n") + negative_prompt = neg_prompts + + if negative_prompt_embeds is None: + neg_prompt_list = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt_multiple_images( + prompt=neg_prompt_list, + images=flattened_images, + device=device, + max_sequence_length=max_sequence_length, + ) + + # Pad and concatenate [negative, positive] + max_seq_len = max(prompt_embeds.shape[1], negative_prompt_embeds.shape[1]) + prompt_embeds = torch.cat( + [ + self._pad_sequence(negative_prompt_embeds, max_seq_len), + self._pad_sequence(prompt_embeds, max_seq_len), + ] + ) + if prompt_embeds_mask is not None and negative_prompt_embeds_mask is not None: + prompt_embeds_mask = torch.cat( + [ + self._pad_sequence(negative_prompt_embeds_mask, max_seq_len), + self._pad_sequence(prompt_embeds_mask, max_seq_len), + ] + ) + + # Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # Prepare latents (patchified) + num_channels_latents = self.transformer.config.in_channels + latents, target_mask, shape_list = self.prepare_latents( + batch_size=batch_size, + num_channels_latents=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + reference_images=images, + latents=latents, + ) + + # Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + clean_reference_backup = latents.clone() + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # Restore reference patches + latents[~target_mask] = clean_reference_backup[~target_mask] + + model_input = latents + + # CFG expansion + if self.do_classifier_free_guidance: + model_input_cfg = torch.cat([model_input] * 2) + t_expand = t.repeat(model_input_cfg.shape[0]) + cfg_shape_list = shape_list * 2 + else: + model_input_cfg = model_input + t_expand = t.repeat(batch_size) + cfg_shape_list = shape_list + + # Transformer forward + noise_pred = self.transformer( + hidden_states=model_input_cfg, + timestep=t_expand, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_mask=prompt_embeds_mask, + shape_list=cfg_shape_list, + return_dict=False, + )[0] + + # CFG combination with norm rescaling + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + comb_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + cond_norm = torch.norm(noise_pred_text, dim=2, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=2, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm.clamp_min(1e-6)) + + # Scheduler step + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0].to( + dtype=prompt_embeds.dtype + ) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + if progress_bar is not None: + progress_bar.update() + + # Post-processing: decode target latents + if output_type != "latent": + latents[~target_mask] = clean_reference_backup[~target_mask] + pt, ph, pw = self.transformer.config.patch_size + + image_list = [] + for b_idx in range(batch_size): + l_t, l_h, l_w = shape_list[b_idx][0] + target_len = l_t * l_h * l_w + + target_patches = latents[b_idx, :target_len] + c_lat = target_patches.shape[1] + video_latent = target_patches.reshape(l_t, l_h, l_w, c_lat, pt, ph, pw) + video_latent = video_latent.permute(3, 0, 4, 1, 5, 2, 6).reshape( + 1, c_lat, l_t * pt, l_h * ph, l_w * pw + ) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, -1, 1, 1, 1) + .to(video_latent.device, video_latent.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, -1, 1, 1, 1) + .to(video_latent.device, video_latent.dtype) + ) + video_latent = video_latent * latents_std + latents_mean + + sample_image = self.vae.decode(video_latent.to(self.vae.dtype), return_dict=False)[0] + # [1, C, T=1, H, W] -> [C, H, W] + sample_image = sample_image.float().squeeze(0).squeeze(1) + image_list.append(sample_image) + + image = torch.stack(image_list) # [B, C, H, W] + image = self.image_processor.postprocess(image, output_type=output_type) + else: + image = latents + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return JoyImageEditPlusPipelineOutput(images=image) diff --git a/diffusers/pipelines/joyimage/pipeline_output.py b/diffusers/pipelines/joyimage/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..4ffb3e53a103cc10c9edbe3bae818ecdf83d356d --- /dev/null +++ b/diffusers/pipelines/joyimage/pipeline_output.py @@ -0,0 +1,25 @@ +from dataclasses import dataclass +from typing import List, Union + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class JoyImageEditPipelineOutput(BaseOutput): + """ + Output class for JoyImageEdit generation pipelines. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +@dataclass +class JoyImageEditPlusPipelineOutput(BaseOutput): + """ + Output class for JoyImage Edit Plus multi-image editing pipelines. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] diff --git a/diffusers/pipelines/kandinsky/__init__.py b/diffusers/pipelines/kandinsky/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..606f7b378a79489bbcbaa87db2040bd4196bbd8a --- /dev/null +++ b/diffusers/pipelines/kandinsky/__init__.py @@ -0,0 +1,66 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_kandinsky"] = ["KandinskyPipeline"] + _import_structure["pipeline_kandinsky_combined"] = [ + "KandinskyCombinedPipeline", + "KandinskyImg2ImgCombinedPipeline", + "KandinskyInpaintCombinedPipeline", + ] + _import_structure["pipeline_kandinsky_img2img"] = ["KandinskyImg2ImgPipeline"] + _import_structure["pipeline_kandinsky_inpaint"] = ["KandinskyInpaintPipeline"] + _import_structure["pipeline_kandinsky_prior"] = ["KandinskyPriorPipeline", "KandinskyPriorPipelineOutput"] + _import_structure["text_encoder"] = ["MultilingualCLIP"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + + else: + from .pipeline_kandinsky import KandinskyPipeline + from .pipeline_kandinsky_combined import ( + KandinskyCombinedPipeline, + KandinskyImg2ImgCombinedPipeline, + KandinskyInpaintCombinedPipeline, + ) + from .pipeline_kandinsky_img2img import KandinskyImg2ImgPipeline + from .pipeline_kandinsky_inpaint import KandinskyInpaintPipeline + from .pipeline_kandinsky_prior import KandinskyPriorPipeline, KandinskyPriorPipelineOutput + from .text_encoder import MultilingualCLIP + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/kandinsky/pipeline_kandinsky.py b/diffusers/pipelines/kandinsky/pipeline_kandinsky.py new file mode 100644 index 0000000000000000000000000000000000000000..aa6006ddc082711b668f39b01d8a2587835c2329 --- /dev/null +++ b/diffusers/pipelines/kandinsky/pipeline_kandinsky.py @@ -0,0 +1,415 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import torch +from transformers import ( + XLMRobertaTokenizer, +) + +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDIMScheduler, DDPMScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from .text_encoder import MultilingualCLIP + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyPipeline, KandinskyPriorPipeline + >>> import torch + + >>> pipe_prior = KandinskyPriorPipeline.from_pretrained("kandinsky-community/Kandinsky-2-1-prior") + >>> pipe_prior.to("cuda") + + >>> prompt = "red cat, 4k photo" + >>> out = pipe_prior(prompt) + >>> image_emb = out.image_embeds + >>> negative_image_emb = out.negative_image_embeds + + >>> pipe = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1") + >>> pipe.to("cuda") + + >>> image = pipe( + ... prompt, + ... image_embeds=image_emb, + ... negative_image_embeds=negative_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=100, + ... ).images + + >>> image[0].save("cat.png") + ``` +""" + + +def get_new_h_w(h, w, scale_factor=8): + new_h = h // scale_factor**2 + if h % scale_factor**2 != 0: + new_h += 1 + new_w = w // scale_factor**2 + if w % scale_factor**2 != 0: + new_w += 1 + return new_h * scale_factor, new_w * scale_factor + + +class KandinskyPipeline(DiffusionPipeline): + """ + Pipeline for text-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + text_encoder ([`MultilingualCLIP`]): + Frozen text-encoder. + tokenizer ([`XLMRobertaTokenizer`]): + Tokenizer of class + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + """ + + model_cpu_offload_seq = "text_encoder->unet->movq" + + def __init__( + self, + text_encoder: MultilingualCLIP, + tokenizer: XLMRobertaTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler | DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + ) + self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + truncation=True, + max_length=77, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_input_ids = text_input_ids.to(device) + text_mask = text_inputs.attention_mask.to(device) + + prompt_embeds, text_encoder_hidden_states = self.text_encoder( + input_ids=text_input_ids, attention_mask=text_mask + ) + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=77, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + uncond_text_input_ids = uncond_input.input_ids.to(device) + uncond_text_mask = uncond_input.attention_mask.to(device) + + negative_prompt_embeds, uncond_text_encoder_hidden_states = self.text_encoder( + input_ids=uncond_text_input_ids, attention_mask=uncond_text_mask + ) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image_embeds: torch.Tensor | list[torch.Tensor], + negative_image_embeds: torch.Tensor | list[torch.Tensor], + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds, text_encoder_hidden_states, _ = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + + if do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=prompt_embeds.dtype, device=device + ) + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps_tensor = self.scheduler.timesteps + + num_channels_latents = self.unet.config.in_channels + + height, width = get_new_h_w(height, width, self.movq_scale_factor) + + # create initial latent + latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + text_encoder_hidden_states.dtype, + device, + generator, + latents, + self.scheduler, + ) + + for i, t in enumerate(self.progress_bar(timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + added_cond_kwargs = {"text_embeds": prompt_embeds, "image_embeds": image_embeds} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_encoder_hidden_states, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + ).prev_sample + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # post-processing + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + + self.maybe_free_model_hooks() + + if output_type not in ["pt", "np", "pil"]: + raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}") + + if output_type in ["np", "pil"]: + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky/pipeline_kandinsky_combined.py b/diffusers/pipelines/kandinsky/pipeline_kandinsky_combined.py new file mode 100644 index 0000000000000000000000000000000000000000..3a0c3c07e8caaa7ee3f9d086eb7815d51e9f7592 --- /dev/null +++ b/diffusers/pipelines/kandinsky/pipeline_kandinsky_combined.py @@ -0,0 +1,817 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Callable + +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, + XLMRobertaTokenizer, +) + +from ...models import PriorTransformer, UNet2DConditionModel, VQModel +from ...schedulers import DDIMScheduler, DDPMScheduler, UnCLIPScheduler +from ...utils import ( + replace_example_docstring, +) +from ..pipeline_utils import DiffusionPipeline +from .pipeline_kandinsky import KandinskyPipeline +from .pipeline_kandinsky_img2img import KandinskyImg2ImgPipeline +from .pipeline_kandinsky_inpaint import KandinskyInpaintPipeline +from .pipeline_kandinsky_prior import KandinskyPriorPipeline +from .text_encoder import MultilingualCLIP + + +TEXT2IMAGE_EXAMPLE_DOC_STRING = """ + Examples: + ```py + from diffusers import AutoPipelineForText2Image + import torch + + pipe = AutoPipelineForText2Image.from_pretrained( + "kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16 + ) + pipe.enable_model_cpu_offload() + + prompt = "A lion in galaxies, spirals, nebulae, stars, smoke, iridescent, intricate detail, octane render, 8k" + + image = pipe(prompt=prompt, num_inference_steps=25).images[0] + ``` +""" + +IMAGE2IMAGE_EXAMPLE_DOC_STRING = """ + Examples: + ```py + from diffusers import AutoPipelineForImage2Image + import torch + import requests + from io import BytesIO + from PIL import Image + import os + + pipe = AutoPipelineForImage2Image.from_pretrained( + "kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16 + ) + pipe.enable_model_cpu_offload() + + prompt = "A fantasy landscape, Cinematic lighting" + negative_prompt = "low quality, bad quality" + + url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + + response = requests.get(url) + image = Image.open(BytesIO(response.content)).convert("RGB") + image.thumbnail((768, 768)) + + image = pipe(prompt=prompt, image=original_image, num_inference_steps=25).images[0] + ``` +""" + +INPAINT_EXAMPLE_DOC_STRING = """ + Examples: + ```py + from diffusers import AutoPipelineForInpainting + from diffusers.utils import load_image + import torch + import numpy as np + + pipe = AutoPipelineForInpainting.from_pretrained( + "kandinsky-community/kandinsky-2-1-inpaint", torch_dtype=torch.float16 + ) + pipe.enable_model_cpu_offload() + + prompt = "A fantasy landscape, Cinematic lighting" + negative_prompt = "low quality, bad quality" + + original_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" "/kandinsky/cat.png" + ) + + mask = np.zeros((768, 768), dtype=np.float32) + # Let's mask out an area above the cat's head + mask[:250, 250:-250] = 1 + + image = pipe(prompt=prompt, image=original_image, mask_image=mask, num_inference_steps=25).images[0] + ``` +""" + + +class KandinskyCombinedPipeline(DiffusionPipeline): + """ + Combined Pipeline for text-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + text_encoder ([`MultilingualCLIP`]): + Frozen text-encoder. + tokenizer ([`XLMRobertaTokenizer`]): + Tokenizer of class + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + prior_prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + prior_tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + prior_scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + """ + + _load_connected_pipes = True + model_cpu_offload_seq = "text_encoder->unet->movq->prior_prior->prior_image_encoder->prior_text_encoder" + _exclude_from_cpu_offload = ["prior_prior"] + + def __init__( + self, + text_encoder: MultilingualCLIP, + tokenizer: XLMRobertaTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler | DDPMScheduler, + movq: VQModel, + prior_prior: PriorTransformer, + prior_image_encoder: CLIPVisionModelWithProjection, + prior_text_encoder: CLIPTextModelWithProjection, + prior_tokenizer: CLIPTokenizer, + prior_scheduler: UnCLIPScheduler, + prior_image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + prior_prior=prior_prior, + prior_image_encoder=prior_image_encoder, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + prior_image_processor=prior_image_processor, + ) + self.prior_pipe = KandinskyPriorPipeline( + prior=prior_prior, + image_encoder=prior_image_encoder, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + image_processor=prior_image_processor, + ) + self.decoder_pipe = KandinskyPipeline( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗 + Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a + GPU only when their specific submodule's `forward` method is called. Offloading happens on a submodule basis. + Memory savings are higher than using `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.enable_model_cpu_offload() + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + height: int = 512, + width: int = 512, + prior_guidance_scale: float = 4.0, + prior_num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + prior_outputs = self.prior_pipe( + prompt=prompt, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + num_inference_steps=prior_num_inference_steps, + generator=generator, + latents=latents, + guidance_scale=prior_guidance_scale, + output_type="pt", + return_dict=False, + ) + image_embeds = prior_outputs[0] + negative_image_embeds = prior_outputs[1] + + prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt + + if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0: + prompt = (image_embeds.shape[0] // len(prompt)) * prompt + + outputs = self.decoder_pipe( + prompt=prompt, + image_embeds=image_embeds, + negative_image_embeds=negative_image_embeds, + width=width, + height=height, + num_inference_steps=num_inference_steps, + generator=generator, + guidance_scale=guidance_scale, + output_type=output_type, + callback=callback, + callback_steps=callback_steps, + return_dict=return_dict, + ) + + self.maybe_free_model_hooks() + + return outputs + + +class KandinskyImg2ImgCombinedPipeline(DiffusionPipeline): + """ + Combined Pipeline for image-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + text_encoder ([`MultilingualCLIP`]): + Frozen text-encoder. + tokenizer ([`XLMRobertaTokenizer`]): + Tokenizer of class + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + prior_prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + prior_tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + prior_scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + """ + + _load_connected_pipes = True + model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->prior_prior->text_encoder->unet->movq" + _exclude_from_cpu_offload = ["prior_prior"] + + def __init__( + self, + text_encoder: MultilingualCLIP, + tokenizer: XLMRobertaTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler | DDPMScheduler, + movq: VQModel, + prior_prior: PriorTransformer, + prior_image_encoder: CLIPVisionModelWithProjection, + prior_text_encoder: CLIPTextModelWithProjection, + prior_tokenizer: CLIPTokenizer, + prior_scheduler: UnCLIPScheduler, + prior_image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + prior_prior=prior_prior, + prior_image_encoder=prior_image_encoder, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + prior_image_processor=prior_image_processor, + ) + self.prior_pipe = KandinskyPriorPipeline( + prior=prior_prior, + image_encoder=prior_image_encoder, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + image_processor=prior_image_processor, + ) + self.decoder_pipe = KandinskyImg2ImgPipeline( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + Note that offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.enable_model_cpu_offload() + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(IMAGE2IMAGE_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + negative_prompt: str | list[str] | None = None, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + strength: float = 0.3, + height: int = 512, + width: int = 512, + prior_guidance_scale: float = 4.0, + prior_num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded + again. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + strength (`float`, *optional*, defaults to 0.3): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + prior_outputs = self.prior_pipe( + prompt=prompt, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + num_inference_steps=prior_num_inference_steps, + generator=generator, + latents=latents, + guidance_scale=prior_guidance_scale, + output_type="pt", + return_dict=False, + ) + image_embeds = prior_outputs[0] + negative_image_embeds = prior_outputs[1] + + prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt + image = [image] if isinstance(prompt, PIL.Image.Image) else image + + if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0: + prompt = (image_embeds.shape[0] // len(prompt)) * prompt + + if ( + isinstance(image, (list, tuple)) + and len(image) < image_embeds.shape[0] + and image_embeds.shape[0] % len(image) == 0 + ): + image = (image_embeds.shape[0] // len(image)) * image + + outputs = self.decoder_pipe( + prompt=prompt, + image=image, + image_embeds=image_embeds, + negative_image_embeds=negative_image_embeds, + strength=strength, + width=width, + height=height, + num_inference_steps=num_inference_steps, + generator=generator, + guidance_scale=guidance_scale, + output_type=output_type, + callback=callback, + callback_steps=callback_steps, + return_dict=return_dict, + ) + + self.maybe_free_model_hooks() + + return outputs + + +class KandinskyInpaintCombinedPipeline(DiffusionPipeline): + """ + Combined Pipeline for generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + text_encoder ([`MultilingualCLIP`]): + Frozen text-encoder. + tokenizer ([`XLMRobertaTokenizer`]): + Tokenizer of class + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + prior_prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + prior_tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + prior_scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + """ + + _load_connected_pipes = True + model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->prior_prior->text_encoder->unet->movq" + _exclude_from_cpu_offload = ["prior_prior"] + + def __init__( + self, + text_encoder: MultilingualCLIP, + tokenizer: XLMRobertaTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler | DDPMScheduler, + movq: VQModel, + prior_prior: PriorTransformer, + prior_image_encoder: CLIPVisionModelWithProjection, + prior_text_encoder: CLIPTextModelWithProjection, + prior_tokenizer: CLIPTokenizer, + prior_scheduler: UnCLIPScheduler, + prior_image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + prior_prior=prior_prior, + prior_image_encoder=prior_image_encoder, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + prior_image_processor=prior_image_processor, + ) + self.prior_pipe = KandinskyPriorPipeline( + prior=prior_prior, + image_encoder=prior_image_encoder, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + image_processor=prior_image_processor, + ) + self.decoder_pipe = KandinskyInpaintPipeline( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + Note that offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.enable_model_cpu_offload() + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(INPAINT_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + mask_image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + negative_prompt: str | list[str] | None = None, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + height: int = 512, + width: int = 512, + prior_guidance_scale: float = 4.0, + prior_num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded + again. + mask_image (`np.array`): + Tensor representing an image batch, to mask `image`. White pixels in the mask will be repainted, while + black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single + channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, + so the expected shape would be `(B, H, W, 1)`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + prior_outputs = self.prior_pipe( + prompt=prompt, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + num_inference_steps=prior_num_inference_steps, + generator=generator, + latents=latents, + guidance_scale=prior_guidance_scale, + output_type="pt", + return_dict=False, + ) + image_embeds = prior_outputs[0] + negative_image_embeds = prior_outputs[1] + + prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt + image = [image] if isinstance(prompt, PIL.Image.Image) else image + mask_image = [mask_image] if isinstance(mask_image, PIL.Image.Image) else mask_image + + if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0: + prompt = (image_embeds.shape[0] // len(prompt)) * prompt + + if ( + isinstance(image, (list, tuple)) + and len(image) < image_embeds.shape[0] + and image_embeds.shape[0] % len(image) == 0 + ): + image = (image_embeds.shape[0] // len(image)) * image + + if ( + isinstance(mask_image, (list, tuple)) + and len(mask_image) < image_embeds.shape[0] + and image_embeds.shape[0] % len(mask_image) == 0 + ): + mask_image = (image_embeds.shape[0] // len(mask_image)) * mask_image + + outputs = self.decoder_pipe( + prompt=prompt, + image=image, + mask_image=mask_image, + image_embeds=image_embeds, + negative_image_embeds=negative_image_embeds, + width=width, + height=height, + num_inference_steps=num_inference_steps, + generator=generator, + guidance_scale=guidance_scale, + output_type=output_type, + callback=callback, + callback_steps=callback_steps, + return_dict=return_dict, + ) + + self.maybe_free_model_hooks() + + return outputs diff --git a/diffusers/pipelines/kandinsky/pipeline_kandinsky_img2img.py b/diffusers/pipelines/kandinsky/pipeline_kandinsky_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..f374741bd9fb742f09c8c4360e041a58b8b0b88c --- /dev/null +++ b/diffusers/pipelines/kandinsky/pipeline_kandinsky_img2img.py @@ -0,0 +1,501 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import PIL.Image +import torch +from transformers import ( + XLMRobertaTokenizer, +) + +from ...image_processor import VaeImageProcessor +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDIMScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from .text_encoder import MultilingualCLIP + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyImg2ImgPipeline, KandinskyPriorPipeline + >>> from diffusers.utils import load_image + >>> import torch + + >>> pipe_prior = KandinskyPriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + + >>> prompt = "A red cartoon frog, 4k" + >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False) + + >>> pipe = KandinskyImg2ImgPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + + >>> init_image = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/frog.png" + ... ) + + >>> image = pipe( + ... prompt, + ... image=init_image, + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=100, + ... strength=0.2, + ... ).images + + >>> image[0].save("red_frog.png") + ``` +""" + + +def get_new_h_w(h, w, scale_factor=8): + new_h = h // scale_factor**2 + if h % scale_factor**2 != 0: + new_h += 1 + new_w = w // scale_factor**2 + if w % scale_factor**2 != 0: + new_w += 1 + return new_h * scale_factor, new_w * scale_factor + + +class KandinskyImg2ImgPipeline(DiffusionPipeline): + """ + Pipeline for image-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + text_encoder ([`MultilingualCLIP`]): + Frozen text-encoder. + tokenizer ([`XLMRobertaTokenizer`]): + Tokenizer of class + scheduler ([`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ image encoder and decoder + """ + + model_cpu_offload_seq = "text_encoder->unet->movq" + + def __init__( + self, + text_encoder: MultilingualCLIP, + movq: VQModel, + tokenizer: XLMRobertaTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + movq=movq, + ) + self.movq_scale_factor = ( + 2 ** (len(self.movq.config.block_out_channels) - 1) if getattr(self, "movq", None) else 8 + ) + movq_latent_channels = self.movq.config.latent_channels if getattr(self, "movq", None) else 4 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.movq_scale_factor, + vae_latent_channels=movq_latent_channels, + resample="bicubic", + reducing_gap=1, + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start:] + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, latents, latent_timestep, shape, dtype, device, generator, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + + shape = latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + latents = self.add_noise(latents, noise, latent_timestep) + return latents + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=77, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_input_ids = text_input_ids.to(device) + text_mask = text_inputs.attention_mask.to(device) + + prompt_embeds, text_encoder_hidden_states = self.text_encoder( + input_ids=text_input_ids, attention_mask=text_mask + ) + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=77, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + uncond_text_input_ids = uncond_input.input_ids.to(device) + uncond_text_mask = uncond_input.attention_mask.to(device) + + negative_prompt_embeds, uncond_text_encoder_hidden_states = self.text_encoder( + input_ids=uncond_text_input_ids, attention_mask=uncond_text_mask + ) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + # add_noise method to overwrite the one in schedule because it use a different beta schedule for adding noise vs sampling + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + betas = torch.linspace(0.0001, 0.02, 1000, dtype=torch.float32) + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_cumprod = alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + + return noisy_samples + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + image_embeds: torch.Tensor, + negative_image_embeds: torch.Tensor, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + strength: float = 0.3, + guidance_scale: float = 7.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`torch.Tensor`, `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + strength (`float`, *optional*, defaults to 0.3): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + # 1. Define call parameters + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = guidance_scale > 1.0 + + # 2. get text and image embeddings + prompt_embeds, text_encoder_hidden_states, _ = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + + if do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=prompt_embeds.dtype, device=device + ) + + # 3. pre-processing initial image + if not isinstance(image, list): + image = [image] + if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image): + raise ValueError( + f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor" + ) + + image = torch.cat([self.image_processor.preprocess(i, width, height) for i in image], dim=0) + image = image.to(dtype=prompt_embeds.dtype, device=device) + + latents = self.movq.encode(image)["latents"] + latents = latents.repeat_interleave(num_images_per_prompt, dim=0) + + # 4. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + + timesteps_tensor, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + # the formular to calculate timestep for add_noise is taken from the original kandinsky repo + latent_timestep = int(self.scheduler.config.num_train_timesteps * strength) - 2 + + latent_timestep = torch.tensor([latent_timestep] * batch_size, dtype=timesteps_tensor.dtype, device=device) + + num_channels_latents = self.unet.config.in_channels + + height, width = get_new_h_w(height, width, self.movq_scale_factor) + + # 5. Create initial latent + latents = self.prepare_latents( + latents, + latent_timestep, + (batch_size, num_channels_latents, height, width), + text_encoder_hidden_states.dtype, + device, + generator, + self.scheduler, + ) + + # 6. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + added_cond_kwargs = {"text_embeds": prompt_embeds, "image_embeds": image_embeds} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_encoder_hidden_states, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + ).prev_sample + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 7. post-processing + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + + self.maybe_free_model_hooks() + + if output_type not in ["pt", "np", "pil"]: + raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}") + + image = self.image_processor.postprocess(image, output_type) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky/pipeline_kandinsky_inpaint.py b/diffusers/pipelines/kandinsky/pipeline_kandinsky_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..9bc52ce871f5be608a7efb73106f02afa40d0c22 --- /dev/null +++ b/diffusers/pipelines/kandinsky/pipeline_kandinsky_inpaint.py @@ -0,0 +1,643 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from copy import deepcopy +from typing import Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from packaging import version +from PIL import Image +from transformers import ( + XLMRobertaTokenizer, +) + +from ... import __version__ +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDIMScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from .text_encoder import MultilingualCLIP + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyInpaintPipeline, KandinskyPriorPipeline + >>> from diffusers.utils import load_image + >>> import torch + >>> import numpy as np + + >>> pipe_prior = KandinskyPriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + + >>> prompt = "a hat" + >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False) + + >>> pipe = KandinskyInpaintPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-1-inpaint", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + + >>> init_image = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ) + + >>> mask = np.zeros((768, 768), dtype=np.float32) + >>> mask[:250, 250:-250] = 1 + + >>> out = pipe( + ... prompt, + ... image=init_image, + ... mask_image=mask, + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=50, + ... ) + + >>> image = out.images[0] + >>> image.save("cat_with_hat.png") + ``` +""" + + +def get_new_h_w(h, w, scale_factor=8): + new_h = h // scale_factor**2 + if h % scale_factor**2 != 0: + new_h += 1 + new_w = w // scale_factor**2 + if w % scale_factor**2 != 0: + new_w += 1 + return new_h * scale_factor, new_w * scale_factor + + +def prepare_mask(masks): + prepared_masks = [] + for mask in masks: + old_mask = deepcopy(mask) + for i in range(mask.shape[1]): + for j in range(mask.shape[2]): + if old_mask[0][i][j] == 1: + continue + if i != 0: + mask[:, i - 1, j] = 0 + if j != 0: + mask[:, i, j - 1] = 0 + if i != 0 and j != 0: + mask[:, i - 1, j - 1] = 0 + if i != mask.shape[1] - 1: + mask[:, i + 1, j] = 0 + if j != mask.shape[2] - 1: + mask[:, i, j + 1] = 0 + if i != mask.shape[1] - 1 and j != mask.shape[2] - 1: + mask[:, i + 1, j + 1] = 0 + prepared_masks.append(mask) + return torch.stack(prepared_masks, dim=0) + + +def prepare_mask_and_masked_image(image, mask, height, width): + r""" + Prepares a pair (mask, image) to be consumed by the Kandinsky inpaint pipeline. This means that those inputs will + be converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for + the ``image`` and ``1`` for the ``mask``. + + The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be + binarized (``mask > 0.5``) and cast to ``torch.float32`` too. + + Args: + image (np.array | PIL.Image | torch.Tensor): The image to inpaint. + It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width`` + ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``. + mask (_type_): The mask to apply to the image, i.e. regions to inpaint. + It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width`` + ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + + + Raises: + ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask + should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions. + TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not + (ot the other way around). + + Returns: + tuple[torch.Tensor]: The pair (mask, image) as ``torch.Tensor`` with 4 + dimensions: ``batch x channels x height x width``. + """ + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + if mask is None: + raise ValueError("`mask_image` input cannot be undefined.") + + if isinstance(image, torch.Tensor): + if not isinstance(mask, torch.Tensor): + raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not") + + # Batch single image + if image.ndim == 3: + assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)" + image = image.unsqueeze(0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions" + assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions" + assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size" + + # Check image is in [-1, 1] + if image.min() < -1 or image.max() > 1: + raise ValueError("Image should be in [-1, 1] range") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("Mask should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + # Image as float32 + image = image.to(dtype=torch.float32) + elif isinstance(mask, torch.Tensor): + raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not") + else: + # preprocess image + if isinstance(image, (PIL.Image.Image, np.ndarray)): + image = [image] + + if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): + # resize all images w.r.t passed height an width + image = [i.resize((width, height), resample=Image.BICUBIC, reducing_gap=1) for i in image] + image = [np.array(i.convert("RGB"))[None, :] for i in image] + image = np.concatenate(image, axis=0) + elif isinstance(image, list) and isinstance(image[0], np.ndarray): + image = np.concatenate([i[None, :] for i in image], axis=0) + + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): + mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask] + mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) + mask = mask.astype(np.float32) / 255.0 + elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): + mask = np.concatenate([m[None, None, :] for m in mask], axis=0) + + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + mask = 1 - mask + + return mask, image + + +class KandinskyInpaintPipeline(DiffusionPipeline): + """ + Pipeline for text-guided image inpainting using Kandinsky2.1 + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + text_encoder ([`MultilingualCLIP`]): + Frozen text-encoder. + tokenizer ([`XLMRobertaTokenizer`]): + Tokenizer of class + scheduler ([`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ image encoder and decoder + """ + + model_cpu_offload_seq = "text_encoder->unet->movq" + + def __init__( + self, + text_encoder: MultilingualCLIP, + movq: VQModel, + tokenizer: XLMRobertaTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + movq=movq, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + ) + self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) + self._warn_has_been_called = False + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=77, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_input_ids = text_input_ids.to(device) + text_mask = text_inputs.attention_mask.to(device) + + prompt_embeds, text_encoder_hidden_states = self.text_encoder( + input_ids=text_input_ids, attention_mask=text_mask + ) + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=77, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + uncond_text_input_ids = uncond_input.input_ids.to(device) + uncond_text_mask = uncond_input.attention_mask.to(device) + + negative_prompt_embeds, uncond_text_encoder_hidden_states = self.text_encoder( + input_ids=uncond_text_input_ids, attention_mask=uncond_text_mask + ) + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image: torch.Tensor | PIL.Image.Image, + mask_image: torch.Tensor | PIL.Image.Image | np.ndarray, + image_embeds: torch.Tensor, + negative_image_embeds: torch.Tensor, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`torch.Tensor`, `PIL.Image.Image` or `np.ndarray`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + mask_image (`PIL.Image.Image`,`torch.Tensor` or `np.ndarray`): + `Image`, or a tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. You can pass a pytorch tensor as mask only if the + image you passed is a pytorch tensor, and it should contain one color channel (L) instead of 3, so the + expected shape would be either `(B, 1, H, W,)`, `(B, H, W)`, `(1, H, W)` or `(H, W)` If image is an PIL + image or numpy array, mask should also be a either PIL image or numpy array. If it is a PIL image, it + will be converted to a single channel (luminance) before use. If it is a nummpy array, the expected + shape is `(H, W)`. + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + if not self._warn_has_been_called and version.parse(version.parse(__version__).base_version) < version.parse( + "0.23.0.dev0" + ): + logger.warning( + "Please note that the expected format of `mask_image` has recently been changed. " + "Before diffusers == 0.19.0, Kandinsky Inpainting pipelines repainted black pixels and preserved black pixels. " + "As of diffusers==0.19.0 this behavior has been inverted. Now white pixels are repainted and black pixels are preserved. " + "This way, Kandinsky's masking behavior is aligned with Stable Diffusion. " + "THIS means that you HAVE to invert the input mask to have the same behavior as before as explained in https://github.com/huggingface/diffusers/pull/4207. " + "This warning will be suppressed after the first inference call and will be removed in diffusers>0.23.0" + ) + self._warn_has_been_called = True + + # Define call parameters + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds, text_encoder_hidden_states, _ = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + + if do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=prompt_embeds.dtype, device=device + ) + + # preprocess image and mask + mask_image, image = prepare_mask_and_masked_image(image, mask_image, height, width) + + image = image.to(dtype=prompt_embeds.dtype, device=device) + image = self.movq.encode(image)["latents"] + + mask_image = mask_image.to(dtype=prompt_embeds.dtype, device=device) + + image_shape = tuple(image.shape[-2:]) + mask_image = F.interpolate( + mask_image, + image_shape, + mode="nearest", + ) + mask_image = prepare_mask(mask_image) + masked_image = image * mask_image + + mask_image = mask_image.repeat_interleave(num_images_per_prompt, dim=0) + masked_image = masked_image.repeat_interleave(num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + mask_image = mask_image.repeat(2, 1, 1, 1) + masked_image = masked_image.repeat(2, 1, 1, 1) + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps_tensor = self.scheduler.timesteps + + num_channels_latents = self.movq.config.latent_channels + + # get h, w for latents + sample_height, sample_width = get_new_h_w(height, width, self.movq_scale_factor) + + # create initial latent + latents = self.prepare_latents( + (batch_size, num_channels_latents, sample_height, sample_width), + text_encoder_hidden_states.dtype, + device, + generator, + latents, + self.scheduler, + ) + + # Check that sizes of mask, masked image and latents match with expected + num_channels_mask = mask_image.shape[1] + num_channels_masked_image = masked_image.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + + for i, t in enumerate(self.progress_bar(timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = torch.cat([latent_model_input, masked_image, mask_image], dim=1) + + added_cond_kwargs = {"text_embeds": prompt_embeds, "image_embeds": image_embeds} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_encoder_hidden_states, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + ).prev_sample + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # post-processing + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + + self.maybe_free_model_hooks() + + if output_type not in ["pt", "np", "pil"]: + raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}") + + if output_type in ["np", "pil"]: + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky/pipeline_kandinsky_prior.py b/diffusers/pipelines/kandinsky/pipeline_kandinsky_prior.py new file mode 100644 index 0000000000000000000000000000000000000000..5708286b7d5a51d08e87c7b963e23f40d9932de4 --- /dev/null +++ b/diffusers/pipelines/kandinsky/pipeline_kandinsky_prior.py @@ -0,0 +1,558 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...models import PriorTransformer +from ...schedulers import UnCLIPScheduler +from ...utils import ( + BaseOutput, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyPipeline, KandinskyPriorPipeline + >>> import torch + + >>> pipe_prior = KandinskyPriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-prior") + >>> pipe_prior.to("cuda") + + >>> prompt = "red cat, 4k photo" + >>> out = pipe_prior(prompt) + >>> image_emb = out.image_embeds + >>> negative_image_emb = out.negative_image_embeds + + >>> pipe = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1") + >>> pipe.to("cuda") + + >>> image = pipe( + ... prompt, + ... image_embeds=image_emb, + ... negative_image_embeds=negative_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=100, + ... ).images + + >>> image[0].save("cat.png") + ``` +""" + +EXAMPLE_INTERPOLATE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyPriorPipeline, KandinskyPipeline + >>> from diffusers.utils import load_image + >>> import PIL + + >>> import torch + >>> from torchvision import transforms + + >>> pipe_prior = KandinskyPriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + + >>> img1 = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ) + + >>> img2 = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/starry_night.jpeg" + ... ) + + >>> images_texts = ["a cat", img1, img2] + >>> weights = [0.3, 0.3, 0.4] + >>> image_emb, zero_image_emb = pipe_prior.interpolate(images_texts, weights) + + >>> pipe = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16) + >>> pipe.to("cuda") + + >>> image = pipe( + ... "", + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=150, + ... ).images[0] + + >>> image.save("starry_cat.png") + ``` +""" + + +@dataclass +class KandinskyPriorPipelineOutput(BaseOutput): + """ + Output class for KandinskyPriorPipeline. + + Args: + image_embeds (`torch.Tensor`) + clip image embeddings for text prompt + negative_image_embeds (`list[PIL.Image.Image]` or `np.ndarray`) + clip image embeddings for unconditional tokens + """ + + image_embeds: torch.Tensor | np.ndarray + negative_image_embeds: torch.Tensor | np.ndarray + + +class KandinskyPriorPipeline(DiffusionPipeline): + """ + Pipeline for generating image prior for Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + """ + + _exclude_from_cpu_offload = ["prior"] + model_cpu_offload_seq = "text_encoder->prior" + + def __init__( + self, + prior: PriorTransformer, + image_encoder: CLIPVisionModelWithProjection, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + scheduler: UnCLIPScheduler, + image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + prior=prior, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + image_encoder=image_encoder, + image_processor=image_processor, + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_INTERPOLATE_DOC_STRING) + def interpolate( + self, + images_and_prompts: list[str | PIL.Image.Image | torch.Tensor], + weights: list[float], + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + negative_prior_prompt: str | None = None, + negative_prompt: str = "", + guidance_scale: float = 4.0, + device=None, + ): + """ + Function invoked when using the prior pipeline for interpolation. + + Args: + images_and_prompts (`list[str | PIL.Image.Image | torch.Tensor]`): + list of prompts and images to guide the image generation. + weights: (`list[float]`): + list of weights for each condition in `images_and_prompts` + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + negative_prior_prompt (`str`, *optional*): + The prompt not to guide the prior diffusion process. Ignored when not using guidance (i.e., ignored if + `guidance_scale` is less than `1`). + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. Ignored when not using guidance (i.e., ignored if + `guidance_scale` is less than `1`). + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + + Examples: + + Returns: + [`KandinskyPriorPipelineOutput`] or `tuple` + """ + + device = device or self.device + + if len(images_and_prompts) != len(weights): + raise ValueError( + f"`images_and_prompts` contains {len(images_and_prompts)} items and `weights` contains {len(weights)} items - they should be lists of same length" + ) + + image_embeddings = [] + for cond, weight in zip(images_and_prompts, weights): + if isinstance(cond, str): + image_emb = self( + cond, + num_inference_steps=num_inference_steps, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + negative_prompt=negative_prior_prompt, + guidance_scale=guidance_scale, + ).image_embeds + + elif isinstance(cond, (PIL.Image.Image, torch.Tensor)): + if isinstance(cond, PIL.Image.Image): + cond = ( + self.image_processor(cond, return_tensors="pt") + .pixel_values[0] + .unsqueeze(0) + .to(dtype=self.image_encoder.dtype, device=device) + ) + + image_emb = self.image_encoder(cond)["image_embeds"] + + else: + raise ValueError( + f"`images_and_prompts` can only contains elements to be of type `str`, `PIL.Image.Image` or `torch.Tensor` but is {type(cond)}" + ) + + image_embeddings.append(image_emb * weight) + + image_emb = torch.cat(image_embeddings).sum(dim=0, keepdim=True) + + out_zero = self( + negative_prompt, + num_inference_steps=num_inference_steps, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + negative_prompt=negative_prior_prompt, + guidance_scale=guidance_scale, + ) + zero_image_emb = out_zero.negative_image_embeds if negative_prompt == "" else out_zero.image_embeds + + return KandinskyPriorPipelineOutput(image_embeds=image_emb, negative_image_embeds=zero_image_emb) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def get_zero_embed(self, batch_size=1, device=None): + device = device or self.device + zero_img = torch.zeros(1, 3, self.image_encoder.config.image_size, self.image_encoder.config.image_size).to( + device=device, dtype=self.image_encoder.dtype + ) + zero_image_emb = self.image_encoder(zero_img)["image_embeds"] + zero_image_emb = zero_image_emb.repeat(batch_size, 1) + return zero_image_emb + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + prompt_embeds = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds + uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + guidance_scale: float = 4.0, + output_type: str | None = "pt", + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + output_type (`str`, *optional*, defaults to `"pt"`): + The output format of the generate image. Choose between: `"np"` (`np.array`) or `"pt"` + (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`KandinskyPriorPipelineOutput`] or `tuple` + """ + + if isinstance(prompt, str): + prompt = [prompt] + elif not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + elif not isinstance(negative_prompt, list) and negative_prompt is not None: + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + # if the negative prompt is defined we double the batch size to + # directly retrieve the negative prompt embedding + if negative_prompt is not None: + prompt = prompt + negative_prompt + negative_prompt = 2 * negative_prompt + + device = self._execution_device + + batch_size = len(prompt) + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = guidance_scale > 1.0 + prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + # prior + self.scheduler.set_timesteps(num_inference_steps, device=device) + prior_timesteps_tensor = self.scheduler.timesteps + + embedding_dim = self.prior.config.embedding_dim + + latents = self.prepare_latents( + (batch_size, embedding_dim), + prompt_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + + for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=prompt_embeds, + encoder_hidden_states=text_encoder_hidden_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + if i + 1 == prior_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = prior_timesteps_tensor[i + 1] + + latents = self.scheduler.step( + predicted_image_embedding, + timestep=t, + sample=latents, + generator=generator, + prev_timestep=prev_timestep, + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + latents = self.prior.post_process_latents(latents) + + image_embeddings = latents + + # if negative prompt has been defined, we retrieve split the image embedding into two + if negative_prompt is None: + zero_embeds = self.get_zero_embed(latents.shape[0], device=latents.device) + + self.maybe_free_model_hooks() + else: + image_embeddings, zero_embeds = image_embeddings.chunk(2) + + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.prior_hook.offload() + + if output_type not in ["pt", "np"]: + raise ValueError(f"Only the output types `pt` and `np` are supported not output_type={output_type}") + + if output_type == "np": + image_embeddings = image_embeddings.cpu().numpy() + zero_embeds = zero_embeds.cpu().numpy() + + if not return_dict: + return (image_embeddings, zero_embeds) + + return KandinskyPriorPipelineOutput(image_embeds=image_embeddings, negative_image_embeds=zero_embeds) diff --git a/diffusers/pipelines/kandinsky/text_encoder.py b/diffusers/pipelines/kandinsky/text_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..58cc9ac4d3ed0fd7f17c074d835d96930b195d3b --- /dev/null +++ b/diffusers/pipelines/kandinsky/text_encoder.py @@ -0,0 +1,29 @@ +import torch +from transformers import PreTrainedModel, XLMRobertaConfig, XLMRobertaModel + + +class MCLIPConfig(XLMRobertaConfig): + model_type = "M-CLIP" + + def __init__(self, transformerDimSize=1024, imageDimSize=768, **kwargs): + self.transformerDimensions = transformerDimSize + self.numDims = imageDimSize + super().__init__(**kwargs) + + +class MultilingualCLIP(PreTrainedModel): + config_class = MCLIPConfig + + def __init__(self, config, *args, **kwargs): + super().__init__(config, *args, **kwargs) + self.transformer = XLMRobertaModel(config) + self.LinearTransformation = torch.nn.Linear( + in_features=config.transformerDimensions, out_features=config.numDims + ) + if hasattr(self, "post_init"): + self.post_init() + + def forward(self, input_ids, attention_mask): + embs = self.transformer(input_ids=input_ids, attention_mask=attention_mask)[0] + embs2 = (embs * attention_mask.unsqueeze(2)).sum(dim=1) / attention_mask.sum(dim=1)[:, None] + return self.LinearTransformation(embs2), embs diff --git a/diffusers/pipelines/kandinsky2_2/__init__.py b/diffusers/pipelines/kandinsky2_2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..67e97f161173ac8981dadf757fd8d6438307c973 --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/__init__.py @@ -0,0 +1,70 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_kandinsky2_2"] = ["KandinskyV22Pipeline"] + _import_structure["pipeline_kandinsky2_2_combined"] = [ + "KandinskyV22CombinedPipeline", + "KandinskyV22Img2ImgCombinedPipeline", + "KandinskyV22InpaintCombinedPipeline", + ] + _import_structure["pipeline_kandinsky2_2_controlnet"] = ["KandinskyV22ControlnetPipeline"] + _import_structure["pipeline_kandinsky2_2_controlnet_img2img"] = ["KandinskyV22ControlnetImg2ImgPipeline"] + _import_structure["pipeline_kandinsky2_2_img2img"] = ["KandinskyV22Img2ImgPipeline"] + _import_structure["pipeline_kandinsky2_2_inpainting"] = ["KandinskyV22InpaintPipeline"] + _import_structure["pipeline_kandinsky2_2_prior"] = ["KandinskyV22PriorPipeline"] + _import_structure["pipeline_kandinsky2_2_prior_emb2emb"] = ["KandinskyV22PriorEmb2EmbPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_kandinsky2_2 import KandinskyV22Pipeline + from .pipeline_kandinsky2_2_combined import ( + KandinskyV22CombinedPipeline, + KandinskyV22Img2ImgCombinedPipeline, + KandinskyV22InpaintCombinedPipeline, + ) + from .pipeline_kandinsky2_2_controlnet import KandinskyV22ControlnetPipeline + from .pipeline_kandinsky2_2_controlnet_img2img import KandinskyV22ControlnetImg2ImgPipeline + from .pipeline_kandinsky2_2_img2img import KandinskyV22Img2ImgPipeline + from .pipeline_kandinsky2_2_inpainting import KandinskyV22InpaintPipeline + from .pipeline_kandinsky2_2_prior import KandinskyV22PriorPipeline + from .pipeline_kandinsky2_2_prior_emb2emb import KandinskyV22PriorEmb2EmbPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py new file mode 100644 index 0000000000000000000000000000000000000000..668df4ac5980d80630be57eff10269e16f7ca04e --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py @@ -0,0 +1,331 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import torch + +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDPMScheduler +from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyV22Pipeline, KandinskyV22PriorPipeline + >>> import torch + + >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior") + >>> pipe_prior.to("cuda") + >>> prompt = "red cat, 4k photo" + >>> out = pipe_prior(prompt) + >>> image_emb = out.image_embeds + >>> zero_image_emb = out.negative_image_embeds + >>> pipe = KandinskyV22Pipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder") + >>> pipe.to("cuda") + >>> image = pipe( + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=50, + ... ).images + >>> image[0].save("cat.png") + ``` +""" + + +def downscale_height_and_width(height, width, scale_factor=8): + new_height = height // scale_factor**2 + if height % scale_factor**2 != 0: + new_height += 1 + new_width = width // scale_factor**2 + if width % scale_factor**2 != 0: + new_width += 1 + return new_height * scale_factor, new_width * scale_factor + + +class KandinskyV22Pipeline(DiffusionPipeline): + """ + Pipeline for text-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + """ + + model_cpu_offload_seq = "unet->movq" + _callback_tensor_inputs = ["latents", "image_embeds", "negative_image_embeds"] + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image_embeds: torch.Tensor | list[torch.Tensor], + negative_image_embeds: torch.Tensor | list[torch.Tensor], + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + device = self._execution_device + + self._guidance_scale = guidance_scale + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + batch_size = image_embeds.shape[0] * num_images_per_prompt + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + + if self.do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=self.unet.dtype, device=device + ) + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + num_channels_latents = self.unet.config.in_channels + + height, width = downscale_height_and_width(height, width, self.movq_scale_factor) + + # create initial latent + latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + image_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + + self._num_timesteps = len(timesteps) + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + added_cond_kwargs = {"image_embeds": image_embeds} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=None, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + image_embeds = callback_outputs.pop("image_embeds", image_embeds) + negative_image_embeds = callback_outputs.pop("negative_image_embeds", negative_image_embeds) + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type not in ["pt", "np", "pil", "latent"]: + raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}") + + if not output_type == "latent": + # post-processing + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + if output_type in ["np", "pil"]: + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + else: + image = latents + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py new file mode 100644 index 0000000000000000000000000000000000000000..d5616958a563ff64915fffeb1c4a4c586134a2c7 --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py @@ -0,0 +1,877 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...models import PriorTransformer, UNet2DConditionModel, VQModel +from ...schedulers import DDPMScheduler, UnCLIPScheduler +from ...utils import deprecate, logging, replace_example_docstring +from ..pipeline_utils import DiffusionPipeline +from .pipeline_kandinsky2_2 import KandinskyV22Pipeline +from .pipeline_kandinsky2_2_img2img import KandinskyV22Img2ImgPipeline +from .pipeline_kandinsky2_2_inpainting import KandinskyV22InpaintPipeline +from .pipeline_kandinsky2_2_prior import KandinskyV22PriorPipeline + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +TEXT2IMAGE_EXAMPLE_DOC_STRING = """ + Examples: + ```py + from diffusers import AutoPipelineForText2Image + import torch + + pipe = AutoPipelineForText2Image.from_pretrained( + "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16 + ) + pipe.enable_model_cpu_offload() + + prompt = "A lion in galaxies, spirals, nebulae, stars, smoke, iridescent, intricate detail, octane render, 8k" + + image = pipe(prompt=prompt, num_inference_steps=25).images[0] + ``` +""" + +IMAGE2IMAGE_EXAMPLE_DOC_STRING = """ + Examples: + ```py + from diffusers import AutoPipelineForImage2Image + import torch + import requests + from io import BytesIO + from PIL import Image + import os + + pipe = AutoPipelineForImage2Image.from_pretrained( + "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16 + ) + pipe.enable_model_cpu_offload() + + prompt = "A fantasy landscape, Cinematic lighting" + negative_prompt = "low quality, bad quality" + + url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + + response = requests.get(url) + image = Image.open(BytesIO(response.content)).convert("RGB") + image.thumbnail((768, 768)) + + image = pipe(prompt=prompt, image=original_image, num_inference_steps=25).images[0] + ``` +""" + +INPAINT_EXAMPLE_DOC_STRING = """ + Examples: + ```py + from diffusers import AutoPipelineForInpainting + from diffusers.utils import load_image + import torch + import numpy as np + + pipe = AutoPipelineForInpainting.from_pretrained( + "kandinsky-community/kandinsky-2-2-decoder-inpaint", torch_dtype=torch.float16 + ) + pipe.enable_model_cpu_offload() + + prompt = "A fantasy landscape, Cinematic lighting" + negative_prompt = "low quality, bad quality" + + original_image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" "/kandinsky/cat.png" + ) + + mask = np.zeros((768, 768), dtype=np.float32) + # Let's mask out an area above the cat's head + mask[:250, 250:-250] = 1 + + image = pipe(prompt=prompt, image=original_image, mask_image=mask, num_inference_steps=25).images[0] + ``` +""" + + +class KandinskyV22CombinedPipeline(DiffusionPipeline): + """ + Combined Pipeline for text-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + prior_prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + prior_tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + prior_scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + prior_image_processor ([`CLIPImageProcessor`]): + A image_processor to be used to preprocess image from clip. + """ + + model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->unet->movq" + _load_connected_pipes = True + _exclude_from_cpu_offload = ["prior_prior"] + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + prior_prior: PriorTransformer, + prior_image_encoder: CLIPVisionModelWithProjection, + prior_text_encoder: CLIPTextModelWithProjection, + prior_tokenizer: CLIPTokenizer, + prior_scheduler: UnCLIPScheduler, + prior_image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + prior_prior=prior_prior, + prior_image_encoder=prior_image_encoder, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + prior_image_processor=prior_image_processor, + ) + self.prior_pipe = KandinskyV22PriorPipeline( + prior=prior_prior, + image_encoder=prior_image_encoder, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + image_processor=prior_image_processor, + ) + self.decoder_pipe = KandinskyV22Pipeline( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + Note that offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.enable_model_cpu_offload() + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + height: int = 512, + width: int = 512, + prior_guidance_scale: float = 4.0, + prior_num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + prior_callback_on_step_end: Callable[[int, int], None] | None = None, + prior_callback_on_step_end_tensor_inputs: list[str] = ["latents"], + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + prior_callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference of the prior pipeline. + The function is called with the following arguments: `prior_callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. + prior_callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the + list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in + the `._callback_tensor_inputs` attribute of your prior pipeline class. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference of the decoder pipeline. + The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline, + step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors + as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + prior_outputs = self.prior_pipe( + prompt=prompt, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + num_inference_steps=prior_num_inference_steps, + generator=generator, + latents=latents, + guidance_scale=prior_guidance_scale, + output_type="pt", + return_dict=False, + callback_on_step_end=prior_callback_on_step_end, + callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs, + ) + image_embeds = prior_outputs[0] + negative_image_embeds = prior_outputs[1] + + prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt + + if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0: + prompt = (image_embeds.shape[0] // len(prompt)) * prompt + + outputs = self.decoder_pipe( + image_embeds=image_embeds, + negative_image_embeds=negative_image_embeds, + width=width, + height=height, + num_inference_steps=num_inference_steps, + generator=generator, + guidance_scale=guidance_scale, + output_type=output_type, + callback=callback, + callback_steps=callback_steps, + return_dict=return_dict, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + self.maybe_free_model_hooks() + + return outputs + + +class KandinskyV22Img2ImgCombinedPipeline(DiffusionPipeline): + """ + Combined Pipeline for image-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + prior_prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + prior_tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + prior_scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + prior_image_processor ([`CLIPImageProcessor`]): + A image_processor to be used to preprocess image from clip. + """ + + model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->unet->movq" + _load_connected_pipes = True + _exclude_from_cpu_offload = ["prior_prior"] + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + prior_prior: PriorTransformer, + prior_image_encoder: CLIPVisionModelWithProjection, + prior_text_encoder: CLIPTextModelWithProjection, + prior_tokenizer: CLIPTokenizer, + prior_scheduler: UnCLIPScheduler, + prior_image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + prior_prior=prior_prior, + prior_image_encoder=prior_image_encoder, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + prior_image_processor=prior_image_processor, + ) + self.prior_pipe = KandinskyV22PriorPipeline( + prior=prior_prior, + image_encoder=prior_image_encoder, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + image_processor=prior_image_processor, + ) + self.decoder_pipe = KandinskyV22Img2ImgPipeline( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_model_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + """ + self.prior_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + Note that offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.enable_model_cpu_offload() + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(IMAGE2IMAGE_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + negative_prompt: str | list[str] | None = None, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + strength: float = 0.3, + num_images_per_prompt: int = 1, + height: int = 512, + width: int = 512, + prior_guidance_scale: float = 4.0, + prior_num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + prior_callback_on_step_end: Callable[[int, int], None] | None = None, + prior_callback_on_step_end_tensor_inputs: list[str] = ["latents"], + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded + again. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + strength (`float`, *optional*, defaults to 0.3): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + prior_callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference of the prior pipeline. + The function is called with the following arguments: `prior_callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. + prior_callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the + list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in + the `._callback_tensor_inputs` attribute of your prior pipeline class. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference of the decoder pipeline. + The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline, + step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors + as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + prior_outputs = self.prior_pipe( + prompt=prompt, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + num_inference_steps=prior_num_inference_steps, + generator=generator, + latents=latents, + guidance_scale=prior_guidance_scale, + output_type="pt", + return_dict=False, + callback_on_step_end=prior_callback_on_step_end, + callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs, + ) + image_embeds = prior_outputs[0] + negative_image_embeds = prior_outputs[1] + + prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt + image = [image] if isinstance(image, PIL.Image.Image) else image + + if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0: + prompt = (image_embeds.shape[0] // len(prompt)) * prompt + + if ( + isinstance(image, (list, tuple)) + and len(image) < image_embeds.shape[0] + and image_embeds.shape[0] % len(image) == 0 + ): + image = (image_embeds.shape[0] // len(image)) * image + + outputs = self.decoder_pipe( + image=image, + image_embeds=image_embeds, + negative_image_embeds=negative_image_embeds, + width=width, + height=height, + strength=strength, + num_inference_steps=num_inference_steps, + generator=generator, + guidance_scale=guidance_scale, + output_type=output_type, + callback=callback, + callback_steps=callback_steps, + return_dict=return_dict, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self.maybe_free_model_hooks() + return outputs + + +class KandinskyV22InpaintCombinedPipeline(DiffusionPipeline): + """ + Combined Pipeline for inpainting generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler (`DDIMScheduler` | `DDPMScheduler`): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + prior_prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + prior_tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + prior_scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + prior_image_processor ([`CLIPImageProcessor`]): + A image_processor to be used to preprocess image from clip. + """ + + model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->unet->movq" + _load_connected_pipes = True + _exclude_from_cpu_offload = ["prior_prior"] + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + prior_prior: PriorTransformer, + prior_image_encoder: CLIPVisionModelWithProjection, + prior_text_encoder: CLIPTextModelWithProjection, + prior_tokenizer: CLIPTokenizer, + prior_scheduler: UnCLIPScheduler, + prior_image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + prior_prior=prior_prior, + prior_image_encoder=prior_image_encoder, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + prior_image_processor=prior_image_processor, + ) + self.prior_pipe = KandinskyV22PriorPipeline( + prior=prior_prior, + image_encoder=prior_image_encoder, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + image_processor=prior_image_processor, + ) + self.decoder_pipe = KandinskyV22InpaintPipeline( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + Note that offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.enable_model_cpu_offload() + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(INPAINT_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + mask_image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + negative_prompt: str | list[str] | None = None, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + height: int = 512, + width: int = 512, + prior_guidance_scale: float = 4.0, + prior_num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + prior_callback_on_step_end: Callable[[int, int], None] | None = None, + prior_callback_on_step_end_tensor_inputs: list[str] = ["latents"], + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded + again. + mask_image (`np.array`): + Tensor representing an image batch, to mask `image`. White pixels in the mask will be repainted, while + black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single + channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, + so the expected shape would be `(B, H, W, 1)`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + prior_callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `prior_callback_on_step_end(self: DiffusionPipeline, step: int, timestep: + int, callback_kwargs: Dict)`. + prior_callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the + list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in + the `._callback_tensor_inputs` attribute of your pipeline class. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + prior_kwargs = {} + if kwargs.get("prior_callback", None) is not None: + prior_kwargs["callback"] = kwargs.pop("prior_callback") + deprecate( + "prior_callback", + "1.0.0", + "Passing `prior_callback` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`", + ) + if kwargs.get("prior_callback_steps", None) is not None: + deprecate( + "prior_callback_steps", + "1.0.0", + "Passing `prior_callback_steps` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`", + ) + prior_kwargs["callback_steps"] = kwargs.pop("prior_callback_steps") + + prior_outputs = self.prior_pipe( + prompt=prompt, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + num_inference_steps=prior_num_inference_steps, + generator=generator, + latents=latents, + guidance_scale=prior_guidance_scale, + output_type="pt", + return_dict=False, + callback_on_step_end=prior_callback_on_step_end, + callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs, + **prior_kwargs, + ) + image_embeds = prior_outputs[0] + negative_image_embeds = prior_outputs[1] + + prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt + image = [image] if isinstance(image, PIL.Image.Image) else image + mask_image = [mask_image] if isinstance(mask_image, PIL.Image.Image) else mask_image + + if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0: + prompt = (image_embeds.shape[0] // len(prompt)) * prompt + + if ( + isinstance(image, (list, tuple)) + and len(image) < image_embeds.shape[0] + and image_embeds.shape[0] % len(image) == 0 + ): + image = (image_embeds.shape[0] // len(image)) * image + + if ( + isinstance(mask_image, (list, tuple)) + and len(mask_image) < image_embeds.shape[0] + and image_embeds.shape[0] % len(mask_image) == 0 + ): + mask_image = (image_embeds.shape[0] // len(mask_image)) * mask_image + + outputs = self.decoder_pipe( + image=image, + mask_image=mask_image, + image_embeds=image_embeds, + negative_image_embeds=negative_image_embeds, + width=width, + height=height, + num_inference_steps=num_inference_steps, + generator=generator, + guidance_scale=guidance_scale, + output_type=output_type, + return_dict=return_dict, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + **kwargs, + ) + self.maybe_free_model_hooks() + + return outputs diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py new file mode 100644 index 0000000000000000000000000000000000000000..a9b7be516114037f2f29d870ba49b8df20ec1308 --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py @@ -0,0 +1,328 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import torch + +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDPMScheduler +from ...utils import ( + is_torch_xla_available, + logging, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> import numpy as np + + >>> from diffusers import KandinskyV22PriorPipeline, KandinskyV22ControlnetPipeline + >>> from transformers import pipeline + >>> from diffusers.utils import load_image + + + >>> def make_hint(image, depth_estimator): + ... image = depth_estimator(image)["depth"] + ... image = np.array(image) + ... image = image[:, :, None] + ... image = np.concatenate([image, image, image], axis=2) + ... detected_map = torch.from_numpy(image).float() / 255.0 + ... hint = detected_map.permute(2, 0, 1) + ... return hint + + + >>> depth_estimator = pipeline("depth-estimation") + + >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior = pipe_prior.to("cuda") + + >>> pipe = KandinskyV22ControlnetPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + + >>> img = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ).resize((768, 768)) + + >>> hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda") + + >>> prompt = "A robot, 4k photo" + >>> negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature" + + >>> generator = torch.Generator(device="cuda").manual_seed(43) + + >>> image_emb, zero_image_emb = pipe_prior( + ... prompt=prompt, negative_prompt=negative_prior_prompt, generator=generator + ... ).to_tuple() + + >>> images = pipe( + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... hint=hint, + ... num_inference_steps=50, + ... generator=generator, + ... height=768, + ... width=768, + ... ).images + + >>> images[0].save("robot_cat.png") + ``` +""" + + +# Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.downscale_height_and_width +def downscale_height_and_width(height, width, scale_factor=8): + new_height = height // scale_factor**2 + if height % scale_factor**2 != 0: + new_height += 1 + new_width = width // scale_factor**2 + if width % scale_factor**2 != 0: + new_width += 1 + return new_height * scale_factor, new_width * scale_factor + + +class KandinskyV22ControlnetPipeline(DiffusionPipeline): + """ + Pipeline for text-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler ([`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + """ + + model_cpu_offload_seq = "unet->movq" + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + image_embeds: torch.Tensor | list[torch.Tensor], + negative_image_embeds: torch.Tensor | list[torch.Tensor], + hint: torch.Tensor, + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + hint (`torch.Tensor`): + The controlnet condition. + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + device = self._execution_device + + do_classifier_free_guidance = guidance_scale > 1.0 + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + if isinstance(hint, list): + hint = torch.cat(hint, dim=0) + + batch_size = image_embeds.shape[0] * num_images_per_prompt + + if do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + hint = hint.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=self.unet.dtype, device=device + ) + hint = torch.cat([hint, hint], dim=0).to(dtype=self.unet.dtype, device=device) + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps_tensor = self.scheduler.timesteps + + num_channels_latents = self.movq.config.latent_channels + + height, width = downscale_height_and_width(height, width, self.movq_scale_factor) + + # create initial latent + latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + image_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + + for i, t in enumerate(self.progress_bar(timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + added_cond_kwargs = {"image_embeds": image_embeds, "hint": hint} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=None, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + )[0] + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # post-processing + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + + # Offload all models + self.maybe_free_model_hooks() + + if output_type not in ["pt", "np", "pil"]: + raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}") + + if output_type in ["np", "pil"]: + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..f77a40595194e778236a4ada50d93c9734ee18c6 --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py @@ -0,0 +1,371 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import PIL.Image +import torch + +from ...image_processor import VaeImageProcessor +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDPMScheduler +from ...utils import ( + is_torch_xla_available, + logging, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> import numpy as np + + >>> from diffusers import KandinskyV22PriorEmb2EmbPipeline, KandinskyV22ControlnetImg2ImgPipeline + >>> from transformers import pipeline + >>> from diffusers.utils import load_image + + + >>> def make_hint(image, depth_estimator): + ... image = depth_estimator(image)["depth"] + ... image = np.array(image) + ... image = image[:, :, None] + ... image = np.concatenate([image, image, image], axis=2) + ... detected_map = torch.from_numpy(image).float() / 255.0 + ... hint = detected_map.permute(2, 0, 1) + ... return hint + + + >>> depth_estimator = pipeline("depth-estimation") + + >>> pipe_prior = KandinskyV22PriorEmb2EmbPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior = pipe_prior.to("cuda") + + >>> pipe = KandinskyV22ControlnetImg2ImgPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> img = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ).resize((768, 768)) + + + >>> hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda") + + >>> prompt = "A robot, 4k photo" + >>> negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature" + + >>> generator = torch.Generator(device="cuda").manual_seed(43) + + >>> img_emb = pipe_prior(prompt=prompt, image=img, strength=0.85, generator=generator) + >>> negative_emb = pipe_prior(prompt=negative_prior_prompt, image=img, strength=1, generator=generator) + + >>> images = pipe( + ... image=img, + ... strength=0.5, + ... image_embeds=img_emb.image_embeds, + ... negative_image_embeds=negative_emb.image_embeds, + ... hint=hint, + ... num_inference_steps=50, + ... generator=generator, + ... height=768, + ... width=768, + ... ).images + + >>> images[0].save("robot_cat.png") + ``` +""" + + +class KandinskyV22ControlnetImg2ImgPipeline(DiffusionPipeline): + """ + Pipeline for image-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler ([`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + """ + + model_cpu_offload_seq = "unet->movq" + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) if getattr(self, "movq", None) else 8 + movq_latent_channels = self.movq.config.latent_channels if getattr(self, "movq", None) else 4 + self.image_processor = VaeImageProcessor( + vae_scale_factor=movq_scale_factor, + vae_latent_channels=movq_latent_channels, + resample="bicubic", + reducing_gap=1, + ) + + # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_img2img.KandinskyImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start:] + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2_img2img.KandinskyV22Img2ImgPipeline.prepare_latents + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + self.movq.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = self.movq.encode(image).latent_dist.sample(generator) + + init_latents = self.movq.config.scaling_factor * init_latents + + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + @torch.no_grad() + def __call__( + self, + image_embeds: torch.Tensor | list[torch.Tensor], + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + negative_image_embeds: torch.Tensor | list[torch.Tensor], + hint: torch.Tensor, + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + strength: float = 0.3, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded + again. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + hint (`torch.Tensor`): + The controlnet condition. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + device = self._execution_device + + do_classifier_free_guidance = guidance_scale > 1.0 + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + if isinstance(hint, list): + hint = torch.cat(hint, dim=0) + + batch_size = image_embeds.shape[0] + + if do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + hint = hint.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=self.unet.dtype, device=device + ) + hint = torch.cat([hint, hint], dim=0).to(dtype=self.unet.dtype, device=device) + + if not isinstance(image, list): + image = [image] + if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image): + raise ValueError( + f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor" + ) + + image = torch.cat([self.image_processor.preprocess(i, width, height) for i in image], dim=0) + image = image.to(dtype=image_embeds.dtype, device=device) + + latents = self.movq.encode(image)["latents"] + latents = latents.repeat_interleave(num_images_per_prompt, dim=0) + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + latents = self.prepare_latents( + latents, latent_timestep, batch_size, num_images_per_prompt, image_embeds.dtype, device, generator + ) + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + added_cond_kwargs = {"image_embeds": image_embeds, "hint": hint} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=None, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + )[0] + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # post-processing + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + + # Offload all models + self.maybe_free_model_hooks() + + if output_type not in ["pt", "np", "pil"]: + raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}") + + image = self.image_processor.postprocess(image, output_type) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..dc17f49bbfe0ba4655af97c3fa93a52d414e7aa1 --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py @@ -0,0 +1,382 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import PIL.Image +import torch + +from ...image_processor import VaeImageProcessor +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDPMScheduler +from ...utils import deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyV22Img2ImgPipeline, KandinskyV22PriorPipeline + >>> from diffusers.utils import load_image + >>> import torch + + >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + + >>> prompt = "A red cartoon frog, 4k" + >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False) + + >>> pipe = KandinskyV22Img2ImgPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + + >>> init_image = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/frog.png" + ... ) + + >>> image = pipe( + ... image=init_image, + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=100, + ... strength=0.2, + ... ).images + + >>> image[0].save("red_frog.png") + ``` +""" + + +class KandinskyV22Img2ImgPipeline(DiffusionPipeline): + """ + Pipeline for image-to-image generation using Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler ([`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + """ + + model_cpu_offload_seq = "unet->movq" + _callback_tensor_inputs = ["latents", "image_embeds", "negative_image_embeds"] + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) if getattr(self, "movq", None) else 8 + movq_latent_channels = self.movq.config.latent_channels if getattr(self, "movq", None) else 4 + self.image_processor = VaeImageProcessor( + vae_scale_factor=movq_scale_factor, + vae_latent_channels=movq_latent_channels, + resample="bicubic", + reducing_gap=1, + ) + + # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_img2img.KandinskyImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start:] + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + self.movq.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = self.movq.encode(image).latent_dist.sample(generator) + + init_latents = self.movq.config.scaling_factor * init_latents + + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + def __call__( + self, + image_embeds: torch.Tensor | list[torch.Tensor], + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image], + negative_image_embeds: torch.Tensor | list[torch.Tensor], + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + strength: float = 0.3, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded + again. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + device = self._execution_device + + self._guidance_scale = guidance_scale + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + batch_size = image_embeds.shape[0] + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + + if self.do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=self.unet.dtype, device=device + ) + + if not isinstance(image, list): + image = [image] + if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image): + raise ValueError( + f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor" + ) + + image = torch.cat([self.image_processor.preprocess(i, width, height) for i in image], dim=0) + image = image.to(dtype=image_embeds.dtype, device=device) + + latents = self.movq.encode(image)["latents"] + latents = latents.repeat_interleave(num_images_per_prompt, dim=0) + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + latents = self.prepare_latents( + latents, latent_timestep, batch_size, num_images_per_prompt, image_embeds.dtype, device, generator + ) + self._num_timesteps = len(timesteps) + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + added_cond_kwargs = {"image_embeds": image_embeds} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=None, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + image_embeds = callback_outputs.pop("image_embeds", image_embeds) + negative_image_embeds = callback_outputs.pop("negative_image_embeds", negative_image_embeds) + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + image = self.image_processor.postprocess(image, output_type) + else: + image = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py new file mode 100644 index 0000000000000000000000000000000000000000..f258dfc0709433a1fc425edc3a1553329880c35c --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py @@ -0,0 +1,567 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from copy import deepcopy +from typing import Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from packaging import version +from PIL import Image + +from ... import __version__ +from ...models import UNet2DConditionModel, VQModel +from ...schedulers import DDPMScheduler +from ...utils import deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyV22InpaintPipeline, KandinskyV22PriorPipeline + >>> from diffusers.utils import load_image + >>> import torch + >>> import numpy as np + + >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + + >>> prompt = "a hat" + >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False) + + >>> pipe = KandinskyV22InpaintPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-decoder-inpaint", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + + >>> init_image = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ) + + >>> mask = np.zeros((768, 768), dtype=np.float32) + >>> mask[:250, 250:-250] = 1 + + >>> out = pipe( + ... image=init_image, + ... mask_image=mask, + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=50, + ... ) + + >>> image = out.images[0] + >>> image.save("cat_with_hat.png") + ``` +""" + + +# Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.downscale_height_and_width +def downscale_height_and_width(height, width, scale_factor=8): + new_height = height // scale_factor**2 + if height % scale_factor**2 != 0: + new_height += 1 + new_width = width // scale_factor**2 + if width % scale_factor**2 != 0: + new_width += 1 + return new_height * scale_factor, new_width * scale_factor + + +# Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_inpaint.prepare_mask +def prepare_mask(masks): + prepared_masks = [] + for mask in masks: + old_mask = deepcopy(mask) + for i in range(mask.shape[1]): + for j in range(mask.shape[2]): + if old_mask[0][i][j] == 1: + continue + if i != 0: + mask[:, i - 1, j] = 0 + if j != 0: + mask[:, i, j - 1] = 0 + if i != 0 and j != 0: + mask[:, i - 1, j - 1] = 0 + if i != mask.shape[1] - 1: + mask[:, i + 1, j] = 0 + if j != mask.shape[2] - 1: + mask[:, i, j + 1] = 0 + if i != mask.shape[1] - 1 and j != mask.shape[2] - 1: + mask[:, i + 1, j + 1] = 0 + prepared_masks.append(mask) + return torch.stack(prepared_masks, dim=0) + + +# Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_inpaint.prepare_mask_and_masked_image +def prepare_mask_and_masked_image(image, mask, height, width): + r""" + Prepares a pair (mask, image) to be consumed by the Kandinsky inpaint pipeline. This means that those inputs will + be converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for + the ``image`` and ``1`` for the ``mask``. + + The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be + binarized (``mask > 0.5``) and cast to ``torch.float32`` too. + + Args: + image (np.array | PIL.Image | torch.Tensor): The image to inpaint. + It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width`` + ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``. + mask (_type_): The mask to apply to the image, i.e. regions to inpaint. + It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width`` + ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + + + Raises: + ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask + should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions. + TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not + (ot the other way around). + + Returns: + tuple[torch.Tensor]: The pair (mask, image) as ``torch.Tensor`` with 4 + dimensions: ``batch x channels x height x width``. + """ + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + if mask is None: + raise ValueError("`mask_image` input cannot be undefined.") + + if isinstance(image, torch.Tensor): + if not isinstance(mask, torch.Tensor): + raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not") + + # Batch single image + if image.ndim == 3: + assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)" + image = image.unsqueeze(0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions" + assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions" + assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size" + + # Check image is in [-1, 1] + if image.min() < -1 or image.max() > 1: + raise ValueError("Image should be in [-1, 1] range") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("Mask should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + # Image as float32 + image = image.to(dtype=torch.float32) + elif isinstance(mask, torch.Tensor): + raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not") + else: + # preprocess image + if isinstance(image, (PIL.Image.Image, np.ndarray)): + image = [image] + + if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): + # resize all images w.r.t passed height an width + image = [i.resize((width, height), resample=Image.BICUBIC, reducing_gap=1) for i in image] + image = [np.array(i.convert("RGB"))[None, :] for i in image] + image = np.concatenate(image, axis=0) + elif isinstance(image, list) and isinstance(image[0], np.ndarray): + image = np.concatenate([i[None, :] for i in image], axis=0) + + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): + mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask] + mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) + mask = mask.astype(np.float32) / 255.0 + elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): + mask = np.concatenate([m[None, None, :] for m in mask], axis=0) + + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + mask = 1 - mask + + return mask, image + + +class KandinskyV22InpaintPipeline(DiffusionPipeline): + """ + Pipeline for text-guided image inpainting using Kandinsky2.1 + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + scheduler ([`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to generate image latents. + unet ([`UNet2DConditionModel`]): + Conditional U-Net architecture to denoise the image embedding. + movq ([`VQModel`]): + MoVQ Decoder to generate the image from the latents. + """ + + model_cpu_offload_seq = "unet->movq" + _callback_tensor_inputs = ["latents", "image_embeds", "negative_image_embeds", "masked_image", "mask_image"] + + def __init__( + self, + unet: UNet2DConditionModel, + scheduler: DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + unet=unet, + scheduler=scheduler, + movq=movq, + ) + self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) + self._warn_has_been_called = False + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + def __call__( + self, + image_embeds: torch.Tensor | list[torch.Tensor], + image: torch.Tensor | PIL.Image.Image, + mask_image: torch.Tensor | PIL.Image.Image | np.ndarray, + negative_image_embeds: torch.Tensor | list[torch.Tensor], + height: int = 512, + width: int = 512, + num_inference_steps: int = 100, + guidance_scale: float = 4.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for text prompt, that will be used to condition the image generation. + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`np.array`): + Tensor representing an image batch, to mask `image`. White pixels in the mask will be repainted, while + black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single + channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, + so the expected shape would be `(B, H, W, 1)`. + negative_image_embeds (`torch.Tensor` or `list[torch.Tensor]`): + The clip image embeddings for negative text prompt, will be used to condition the image generation. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + """ + if not self._warn_has_been_called and version.parse(version.parse(__version__).base_version) < version.parse( + "0.23.0.dev0" + ): + logger.warning( + "Please note that the expected format of `mask_image` has recently been changed. " + "Before diffusers == 0.19.0, Kandinsky Inpainting pipelines repainted black pixels and preserved black pixels. " + "As of diffusers==0.19.0 this behavior has been inverted. Now white pixels are repainted and black pixels are preserved. " + "This way, Kandinsky's masking behavior is aligned with Stable Diffusion. " + "THIS means that you HAVE to invert the input mask to have the same behavior as before as explained in https://github.com/huggingface/diffusers/pull/4207. " + "This warning will be suppressed after the first inference call and will be removed in diffusers>0.23.0" + ) + self._warn_has_been_called = True + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + self._guidance_scale = guidance_scale + + device = self._execution_device + + if isinstance(image_embeds, list): + image_embeds = torch.cat(image_embeds, dim=0) + batch_size = image_embeds.shape[0] * num_images_per_prompt + if isinstance(negative_image_embeds, list): + negative_image_embeds = torch.cat(negative_image_embeds, dim=0) + + if self.do_classifier_free_guidance: + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to( + dtype=self.unet.dtype, device=device + ) + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # preprocess image and mask + mask_image, image = prepare_mask_and_masked_image(image, mask_image, height, width) + + image = image.to(dtype=image_embeds.dtype, device=device) + image = self.movq.encode(image)["latents"] + + mask_image = mask_image.to(dtype=image_embeds.dtype, device=device) + + image_shape = tuple(image.shape[-2:]) + mask_image = F.interpolate( + mask_image, + image_shape, + mode="nearest", + ) + mask_image = prepare_mask(mask_image) + masked_image = image * mask_image + + mask_image = mask_image.repeat_interleave(num_images_per_prompt, dim=0) + masked_image = masked_image.repeat_interleave(num_images_per_prompt, dim=0) + if self.do_classifier_free_guidance: + mask_image = mask_image.repeat(2, 1, 1, 1) + masked_image = masked_image.repeat(2, 1, 1, 1) + + num_channels_latents = self.movq.config.latent_channels + + height, width = downscale_height_and_width(height, width, self.movq_scale_factor) + + # create initial latent + latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + image_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + noise = torch.clone(latents) + + self._num_timesteps = len(timesteps) + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = torch.cat([latent_model_input, masked_image, mask_image], dim=1) + + added_cond_kwargs = {"image_embeds": image_embeds} + noise_pred = self.unet( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=None, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1) + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + _, variance_pred_text = variance_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1) + + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred, _ = noise_pred.split(latents.shape[1], dim=1) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + )[0] + init_latents_proper = image[:1] + init_mask = mask_image[:1] + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = init_mask * init_latents_proper + (1 - init_mask) * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + image_embeds = callback_outputs.pop("image_embeds", image_embeds) + negative_image_embeds = callback_outputs.pop("negative_image_embeds", negative_image_embeds) + masked_image = callback_outputs.pop("masked_image", masked_image) + mask_image = callback_outputs.pop("mask_image", mask_image) + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # post-processing + latents = mask_image[:1] * image[:1] + (1 - mask_image[:1]) * latents + + if output_type not in ["pt", "np", "pil", "latent"]: + raise ValueError( + f"Only the output types `pt`, `pil`, `np` and `latent` are supported not output_type={output_type}" + ) + + if not output_type == "latent": + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + + if output_type in ["np", "pil"]: + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + else: + image = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py new file mode 100644 index 0000000000000000000000000000000000000000..8095f79280d4720fe8edd4efaa2814d1db60ab28 --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py @@ -0,0 +1,557 @@ +from typing import Callable + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...models import PriorTransformer +from ...schedulers import UnCLIPScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..kandinsky import KandinskyPriorPipelineOutput +from ..pipeline_utils import DiffusionPipeline + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyV22Pipeline, KandinskyV22PriorPipeline + >>> import torch + + >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior") + >>> pipe_prior.to("cuda") + >>> prompt = "red cat, 4k photo" + >>> image_emb, negative_image_emb = pipe_prior(prompt).to_tuple() + + >>> pipe = KandinskyV22Pipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder") + >>> pipe.to("cuda") + >>> image = pipe( + ... image_embeds=image_emb, + ... negative_image_embeds=negative_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=50, + ... ).images + >>> image[0].save("cat.png") + ``` +""" + +EXAMPLE_INTERPOLATE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyV22PriorPipeline, KandinskyV22Pipeline + >>> from diffusers.utils import load_image + >>> import PIL + >>> import torch + >>> from torchvision import transforms + + >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + >>> img1 = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ) + >>> img2 = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/starry_night.jpeg" + ... ) + >>> images_texts = ["a cat", img1, img2] + >>> weights = [0.3, 0.3, 0.4] + >>> out = pipe_prior.interpolate(images_texts, weights) + >>> pipe = KandinskyV22Pipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + >>> image = pipe( + ... image_embeds=out.image_embeds, + ... negative_image_embeds=out.negative_image_embeds, + ... height=768, + ... width=768, + ... num_inference_steps=50, + ... ).images[0] + >>> image.save("starry_cat.png") + ``` +""" + + +class KandinskyV22PriorPipeline(DiffusionPipeline): + """ + Pipeline for generating image prior for Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + image_processor ([`CLIPImageProcessor`]): + A image_processor to be used to preprocess image from clip. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->prior" + _exclude_from_cpu_offload = ["prior"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "text_encoder_hidden_states", "text_mask"] + + def __init__( + self, + prior: PriorTransformer, + image_encoder: CLIPVisionModelWithProjection, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + scheduler: UnCLIPScheduler, + image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + prior=prior, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + image_encoder=image_encoder, + image_processor=image_processor, + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_INTERPOLATE_DOC_STRING) + def interpolate( + self, + images_and_prompts: list[str | PIL.Image.Image | torch.Tensor], + weights: list[float], + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + negative_prior_prompt: str | None = None, + negative_prompt: str = "", + guidance_scale: float = 4.0, + device=None, + ): + """ + Function invoked when using the prior pipeline for interpolation. + + Args: + images_and_prompts (`list[str | PIL.Image.Image | torch.Tensor]`): + list of prompts and images to guide the image generation. + weights: (`list[float]`): + list of weights for each condition in `images_and_prompts` + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + negative_prior_prompt (`str`, *optional*): + The prompt not to guide the prior diffusion process. Ignored when not using guidance (i.e., ignored if + `guidance_scale` is less than `1`). + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. Ignored when not using guidance (i.e., ignored if + `guidance_scale` is less than `1`). + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + + Examples: + + Returns: + [`KandinskyPriorPipelineOutput`] or `tuple` + """ + + device = device or self.device + + if len(images_and_prompts) != len(weights): + raise ValueError( + f"`images_and_prompts` contains {len(images_and_prompts)} items and `weights` contains {len(weights)} items - they should be lists of same length" + ) + + image_embeddings = [] + for cond, weight in zip(images_and_prompts, weights): + if isinstance(cond, str): + image_emb = self( + cond, + num_inference_steps=num_inference_steps, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + negative_prompt=negative_prior_prompt, + guidance_scale=guidance_scale, + ).image_embeds.unsqueeze(0) + + elif isinstance(cond, (PIL.Image.Image, torch.Tensor)): + if isinstance(cond, PIL.Image.Image): + cond = ( + self.image_processor(cond, return_tensors="pt") + .pixel_values[0] + .unsqueeze(0) + .to(dtype=self.image_encoder.dtype, device=device) + ) + + image_emb = self.image_encoder(cond)["image_embeds"].repeat(num_images_per_prompt, 1).unsqueeze(0) + + else: + raise ValueError( + f"`images_and_prompts` can only contains elements to be of type `str`, `PIL.Image.Image` or `torch.Tensor` but is {type(cond)}" + ) + + image_embeddings.append(image_emb * weight) + + image_emb = torch.cat(image_embeddings).sum(dim=0) + + out_zero = self( + negative_prompt, + num_inference_steps=num_inference_steps, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + negative_prompt=negative_prior_prompt, + guidance_scale=guidance_scale, + ) + zero_image_emb = out_zero.negative_image_embeds if negative_prompt == "" else out_zero.image_embeds + + return KandinskyPriorPipelineOutput(image_embeds=image_emb, negative_image_embeds=zero_image_emb) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline.get_zero_embed + def get_zero_embed(self, batch_size=1, device=None): + device = device or self.device + zero_img = torch.zeros(1, 3, self.image_encoder.config.image_size, self.image_encoder.config.image_size).to( + device=device, dtype=self.image_encoder.dtype + ) + zero_image_emb = self.image_encoder(zero_img)["image_embeds"] + zero_image_emb = zero_image_emb.repeat(batch_size, 1) + return zero_image_emb + + # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + prompt_embeds = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds + uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + guidance_scale: float = 4.0, + output_type: str | None = "pt", # pt only + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + output_type (`str`, *optional*, defaults to `"pt"`): + The output format of the generate image. Choose between: `"np"` (`np.array`) or `"pt"` + (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`KandinskyPriorPipelineOutput`] or `tuple` + """ + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if isinstance(prompt, str): + prompt = [prompt] + elif not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + elif not isinstance(negative_prompt, list) and negative_prompt is not None: + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + # if the negative prompt is defined we double the batch size to + # directly retrieve the negative prompt embedding + if negative_prompt is not None: + prompt = prompt + negative_prompt + negative_prompt = 2 * negative_prompt + + device = self._execution_device + + batch_size = len(prompt) + batch_size = batch_size * num_images_per_prompt + + self._guidance_scale = guidance_scale + + prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, self.do_classifier_free_guidance, negative_prompt + ) + + # prior + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + embedding_dim = self.prior.config.embedding_dim + + latents = self.prepare_latents( + (batch_size, embedding_dim), + prompt_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + self._num_timesteps = len(timesteps) + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=prompt_embeds, + encoder_hidden_states=text_encoder_hidden_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if self.do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + self.guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + if i + 1 == timesteps.shape[0]: + prev_timestep = None + else: + prev_timestep = timesteps[i + 1] + + latents = self.scheduler.step( + predicted_image_embedding, + timestep=t, + sample=latents, + generator=generator, + prev_timestep=prev_timestep, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + text_encoder_hidden_states = callback_outputs.pop( + "text_encoder_hidden_states", text_encoder_hidden_states + ) + text_mask = callback_outputs.pop("text_mask", text_mask) + + if XLA_AVAILABLE: + xm.mark_step() + + latents = self.prior.post_process_latents(latents) + + image_embeddings = latents + + # if negative prompt has been defined, we retrieve split the image embedding into two + if negative_prompt is None: + zero_embeds = self.get_zero_embed(latents.shape[0], device=latents.device) + else: + image_embeddings, zero_embeds = image_embeddings.chunk(2) + + self.maybe_free_model_hooks() + + if output_type not in ["pt", "np"]: + raise ValueError(f"Only the output types `pt` and `np` are supported not output_type={output_type}") + + if output_type == "np": + image_embeddings = image_embeddings.cpu().numpy() + zero_embeds = zero_embeds.cpu().numpy() + + if not return_dict: + return (image_embeddings, zero_embeds) + + return KandinskyPriorPipelineOutput(image_embeds=image_embeddings, negative_image_embeds=zero_embeds) diff --git a/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py new file mode 100644 index 0000000000000000000000000000000000000000..72f1d8556ec5614129dbad3a60f43763c8ea561c --- /dev/null +++ b/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py @@ -0,0 +1,571 @@ +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...models import PriorTransformer +from ...schedulers import UnCLIPScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..kandinsky import KandinskyPriorPipelineOutput +from ..pipeline_utils import DiffusionPipeline + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyV22Pipeline, KandinskyV22PriorEmb2EmbPipeline + >>> import torch + + >>> pipe_prior = KandinskyPriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + + >>> prompt = "red cat, 4k photo" + >>> img = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ) + >>> image_emb, nagative_image_emb = pipe_prior(prompt, image=img, strength=0.2).to_tuple() + + >>> pipe = KandinskyPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-decoder, torch_dtype=torch.float16" + ... ) + >>> pipe.to("cuda") + + >>> image = pipe( + ... image_embeds=image_emb, + ... negative_image_embeds=negative_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=100, + ... ).images + + >>> image[0].save("cat.png") + ``` +""" + +EXAMPLE_INTERPOLATE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import KandinskyV22PriorEmb2EmbPipeline, KandinskyV22Pipeline + >>> from diffusers.utils import load_image + >>> import PIL + + >>> import torch + >>> from torchvision import transforms + + >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16 + ... ) + >>> pipe_prior.to("cuda") + + >>> img1 = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ) + + >>> img2 = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/starry_night.jpeg" + ... ) + + >>> images_texts = ["a cat", img1, img2] + >>> weights = [0.3, 0.3, 0.4] + >>> image_emb, zero_image_emb = pipe_prior.interpolate(images_texts, weights) + + >>> pipe = KandinskyV22Pipeline.from_pretrained( + ... "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + + >>> image = pipe( + ... image_embeds=image_emb, + ... negative_image_embeds=zero_image_emb, + ... height=768, + ... width=768, + ... num_inference_steps=150, + ... ).images[0] + + >>> image.save("starry_cat.png") + ``` +""" + + +class KandinskyV22PriorEmb2EmbPipeline(DiffusionPipeline): + """ + Pipeline for generating image prior for Kandinsky + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen image-encoder. + text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + scheduler ([`UnCLIPScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->prior" + _exclude_from_cpu_offload = ["prior"] + + def __init__( + self, + prior: PriorTransformer, + image_encoder: CLIPVisionModelWithProjection, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + scheduler: UnCLIPScheduler, + image_processor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + prior=prior, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + image_encoder=image_encoder, + image_processor=image_processor, + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start:] + + return timesteps, num_inference_steps - t_start + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_INTERPOLATE_DOC_STRING) + def interpolate( + self, + images_and_prompts: list[str | PIL.Image.Image | torch.Tensor], + weights: list[float], + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + negative_prior_prompt: str | None = None, + negative_prompt: str = "", + guidance_scale: float = 4.0, + device=None, + ): + """ + Function invoked when using the prior pipeline for interpolation. + + Args: + images_and_prompts (`list[str | PIL.Image.Image | torch.Tensor]`): + list of prompts and images to guide the image generation. + weights: (`list[float]`): + list of weights for each condition in `images_and_prompts` + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + negative_prior_prompt (`str`, *optional*): + The prompt not to guide the prior diffusion process. Ignored when not using guidance (i.e., ignored if + `guidance_scale` is less than `1`). + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. Ignored when not using guidance (i.e., ignored if + `guidance_scale` is less than `1`). + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + + Examples: + + Returns: + [`KandinskyPriorPipelineOutput`] or `tuple` + """ + + device = device or self.device + + if len(images_and_prompts) != len(weights): + raise ValueError( + f"`images_and_prompts` contains {len(images_and_prompts)} items and `weights` contains {len(weights)} items - they should be lists of same length" + ) + + image_embeddings = [] + for cond, weight in zip(images_and_prompts, weights): + if isinstance(cond, str): + image_emb = self( + cond, + num_inference_steps=num_inference_steps, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + negative_prompt=negative_prior_prompt, + guidance_scale=guidance_scale, + ).image_embeds.unsqueeze(0) + + elif isinstance(cond, (PIL.Image.Image, torch.Tensor)): + image_emb = self._encode_image( + cond, device=device, num_images_per_prompt=num_images_per_prompt + ).unsqueeze(0) + + else: + raise ValueError( + f"`images_and_prompts` can only contains elements to be of type `str`, `PIL.Image.Image` or `torch.Tensor` but is {type(cond)}" + ) + + image_embeddings.append(image_emb * weight) + + image_emb = torch.cat(image_embeddings).sum(dim=0) + + return KandinskyPriorPipelineOutput(image_embeds=image_emb, negative_image_embeds=torch.randn_like(image_emb)) + + def _encode_image( + self, + image: torch.Tensor | list[PIL.Image.Image], + device, + num_images_per_prompt, + ): + if not isinstance(image, torch.Tensor): + image = self.image_processor(image, return_tensors="pt").pixel_values.to( + dtype=self.image_encoder.dtype, device=device + ) + + image_emb = self.image_encoder(image)["image_embeds"] # B, D + image_emb = image_emb.repeat_interleave(num_images_per_prompt, dim=0) + image_emb.to(device=device) + + return image_emb + + def prepare_latents(self, emb, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + emb = emb.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + init_latents = emb + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline.get_zero_embed + def get_zero_embed(self, batch_size=1, device=None): + device = device or self.device + zero_img = torch.zeros(1, 3, self.image_encoder.config.image_size, self.image_encoder.config.image_size).to( + device=device, dtype=self.image_encoder.dtype + ) + zero_image_emb = self.image_encoder(zero_img)["image_embeds"] + zero_image_emb = zero_image_emb.repeat(batch_size, 1) + return zero_image_emb + + # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + prompt_embeds = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds + uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + image: torch.Tensor | list[torch.Tensor] | PIL.Image.Image | list[PIL.Image.Image], + strength: float = 0.3, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + guidance_scale: float = 4.0, + output_type: str | None = "pt", # pt only + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`torch.Tensor`, `PIL.Image.Image`, `list[torch.Tensor]` or `list[PIL.Image.Image]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the image + embedding. Can also accept image latents as `image`, if passing latents directly, it will not be + encoded again. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + output_type (`str`, *optional*, defaults to `"pt"`): + The output format of the generate image. Choose between: `"np"` (`np.array`) or `"pt"` + (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`KandinskyPriorPipelineOutput`] or `tuple` + """ + + if isinstance(prompt, str): + prompt = [prompt] + elif not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + elif not isinstance(negative_prompt, list) and negative_prompt is not None: + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + # if the negative prompt is defined we double the batch size to + # directly retrieve the negative prompt embedding + if negative_prompt is not None: + prompt = prompt + negative_prompt + negative_prompt = 2 * negative_prompt + + device = self._execution_device + + batch_size = len(prompt) + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = guidance_scale > 1.0 + prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt + ) + + if not isinstance(image, list): + image = [image] + + if isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + + if isinstance(image, torch.Tensor) and image.ndim == 2: + # allow user to pass image_embeds directly + image_embeds = image.repeat_interleave(num_images_per_prompt, dim=0) + elif isinstance(image, torch.Tensor) and image.ndim != 4: + raise ValueError( + f" if pass `image` as pytorch tensor, or a list of pytorch tensor, please make sure each tensor has shape [batch_size, channels, height, width], currently {image[0].unsqueeze(0).shape}" + ) + else: + image_embeds = self._encode_image(image, device, num_images_per_prompt) + + # prior + self.scheduler.set_timesteps(num_inference_steps, device=device) + + latents = image_embeds + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size) + latents = self.prepare_latents( + latents, + latent_timestep, + batch_size // num_images_per_prompt, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=prompt_embeds, + encoder_hidden_states=text_encoder_hidden_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + if i + 1 == timesteps.shape[0]: + prev_timestep = None + else: + prev_timestep = timesteps[i + 1] + + latents = self.scheduler.step( + predicted_image_embedding, + timestep=t, + sample=latents, + generator=generator, + prev_timestep=prev_timestep, + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + latents = self.prior.post_process_latents(latents) + + image_embeddings = latents + + # if negative prompt has been defined, we retrieve split the image embedding into two + if negative_prompt is None: + zero_embeds = self.get_zero_embed(latents.shape[0], device=latents.device) + else: + image_embeddings, zero_embeds = image_embeddings.chunk(2) + + self.maybe_free_model_hooks() + + if output_type not in ["pt", "np"]: + raise ValueError(f"Only the output types `pt` and `np` are supported not output_type={output_type}") + + if output_type == "np": + image_embeddings = image_embeddings.cpu().numpy() + zero_embeds = zero_embeds.cpu().numpy() + + if not return_dict: + return (image_embeddings, zero_embeds) + + return KandinskyPriorPipelineOutput(image_embeds=image_embeddings, negative_image_embeds=zero_embeds) diff --git a/diffusers/pipelines/kandinsky3/__init__.py b/diffusers/pipelines/kandinsky3/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e8a3063141b5e62682ce3ad8e8e0d02473402665 --- /dev/null +++ b/diffusers/pipelines/kandinsky3/__init__.py @@ -0,0 +1,49 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_kandinsky3"] = ["Kandinsky3Pipeline"] + _import_structure["pipeline_kandinsky3_img2img"] = ["Kandinsky3Img2ImgPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_kandinsky3 import Kandinsky3Pipeline + from .pipeline_kandinsky3_img2img import Kandinsky3Img2ImgPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/kandinsky3/convert_kandinsky3_unet.py b/diffusers/pipelines/kandinsky3/convert_kandinsky3_unet.py new file mode 100644 index 0000000000000000000000000000000000000000..5360632275b41b035f94d84d6156a6bafe84c613 --- /dev/null +++ b/diffusers/pipelines/kandinsky3/convert_kandinsky3_unet.py @@ -0,0 +1,98 @@ +#!/usr/bin/env python3 +import argparse +import fnmatch + +from safetensors.torch import load_file + +from diffusers import Kandinsky3UNet + + +MAPPING = { + "to_time_embed.1": "time_embedding.linear_1", + "to_time_embed.3": "time_embedding.linear_2", + "in_layer": "conv_in", + "out_layer.0": "conv_norm_out", + "out_layer.2": "conv_out", + "down_samples": "down_blocks", + "up_samples": "up_blocks", + "projection_lin": "encoder_hid_proj.projection_linear", + "projection_ln": "encoder_hid_proj.projection_norm", + "feature_pooling": "add_time_condition", + "to_query": "to_q", + "to_key": "to_k", + "to_value": "to_v", + "output_layer": "to_out.0", + "self_attention_block": "attentions.0", +} + +DYNAMIC_MAP = { + "resnet_attn_blocks.*.0": "resnets_in.*", + "resnet_attn_blocks.*.1": ("attentions.*", 1), + "resnet_attn_blocks.*.2": "resnets_out.*", +} +# MAPPING = {} + + +def convert_state_dict(unet_state_dict): + """ + Args: + Convert the state dict of a U-Net model to match the key format expected by Kandinsky3UNet model. + unet_model (torch.nn.Module): The original U-Net model. unet_kandi3_model (torch.nn.Module): The Kandinsky3UNet + model to match keys with. + + Returns: + OrderedDict: The converted state dictionary. + """ + # Example of renaming logic (this will vary based on your model's architecture) + converted_state_dict = {} + for key in unet_state_dict: + new_key = key + for pattern, new_pattern in MAPPING.items(): + new_key = new_key.replace(pattern, new_pattern) + + for dyn_pattern, dyn_new_pattern in DYNAMIC_MAP.items(): + has_matched = False + if fnmatch.fnmatch(new_key, f"*.{dyn_pattern}.*") and not has_matched: + star = int(new_key.split(dyn_pattern.split(".")[0])[-1].split(".")[1]) + + if isinstance(dyn_new_pattern, tuple): + new_star = star + dyn_new_pattern[-1] + dyn_new_pattern = dyn_new_pattern[0] + else: + new_star = star + + pattern = dyn_pattern.replace("*", str(star)) + new_pattern = dyn_new_pattern.replace("*", str(new_star)) + + new_key = new_key.replace(pattern, new_pattern) + has_matched = True + + converted_state_dict[new_key] = unet_state_dict[key] + + return converted_state_dict + + +def main(model_path, output_path): + # Load your original U-Net model + unet_state_dict = load_file(model_path) + + # Initialize your Kandinsky3UNet model + config = {} + + # Convert the state dict + converted_state_dict = convert_state_dict(unet_state_dict) + + unet = Kandinsky3UNet(config) + unet.load_state_dict(converted_state_dict) + + unet.save_pretrained(output_path) + print(f"Converted model saved to {output_path}") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser(description="Convert U-Net PyTorch model to Kandinsky3UNet format") + parser.add_argument("--model_path", type=str, required=True, help="Path to the original U-Net PyTorch model") + parser.add_argument("--output_path", type=str, required=True, help="Path to save the converted model") + + args = parser.parse_args() + main(args.model_path, args.output_path) diff --git a/diffusers/pipelines/kandinsky3/pipeline_kandinsky3.py b/diffusers/pipelines/kandinsky3/pipeline_kandinsky3.py new file mode 100644 index 0000000000000000000000000000000000000000..ca8f124c74cf4cd0a366330ae21a70d2c954f5fc --- /dev/null +++ b/diffusers/pipelines/kandinsky3/pipeline_kandinsky3.py @@ -0,0 +1,581 @@ +from typing import Callable + +import torch +from transformers import T5EncoderModel, T5Tokenizer + +from ...loaders import StableDiffusionLoraLoaderMixin +from ...models import Kandinsky3UNet, VQModel +from ...schedulers import DDPMScheduler +from ...utils import ( + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import AutoPipelineForText2Image + >>> import torch + + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "kandinsky-community/kandinsky-3", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe.enable_model_cpu_offload() + + >>> prompt = "A photograph of the inside of a subway train. There are raccoons sitting on the seats. One of them is reading a newspaper. The window shows the city in the background." + + >>> generator = torch.Generator(device="cpu").manual_seed(0) + >>> image = pipe(prompt, num_inference_steps=25, generator=generator).images[0] + ``` + +""" + + +def downscale_height_and_width(height, width, scale_factor=8): + new_height = height // scale_factor**2 + if height % scale_factor**2 != 0: + new_height += 1 + new_width = width // scale_factor**2 + if width % scale_factor**2 != 0: + new_width += 1 + return new_height * scale_factor, new_width * scale_factor + + +class Kandinsky3Pipeline(DiffusionPipeline, StableDiffusionLoraLoaderMixin): + model_cpu_offload_seq = "text_encoder->unet->movq" + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "negative_attention_mask", + "attention_mask", + ] + + def __init__( + self, + tokenizer: T5Tokenizer, + text_encoder: T5EncoderModel, + unet: Kandinsky3UNet, + scheduler: DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, unet=unet, scheduler=scheduler, movq=movq + ) + + def process_embeds(self, embeddings, attention_mask, cut_context): + if cut_context: + embeddings[attention_mask == 0] = torch.zeros_like(embeddings[attention_mask == 0]) + max_seq_length = attention_mask.sum(-1).max() + 1 + embeddings = embeddings[:, :max_seq_length] + attention_mask = attention_mask[:, :max_seq_length] + return embeddings, attention_mask + + @torch.no_grad() + def encode_prompt( + self, + prompt, + do_classifier_free_guidance=True, + num_images_per_prompt=1, + device=None, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + _cut_context=False, + attention_mask: torch.Tensor | None = None, + negative_attention_mask: torch.Tensor | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead. + Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask. Must provide if passing `prompt_embeds` directly. + negative_attention_mask (`torch.Tensor`, *optional*): + Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly. + """ + if prompt is not None and negative_prompt is not None: + if type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + + if device is None: + device = self._execution_device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + max_length = 128 + + if prompt_embeds is None: + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids.to(device) + attention_mask = text_inputs.attention_mask.to(device) + prompt_embeds = self.text_encoder( + text_input_ids, + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + prompt_embeds, attention_mask = self.process_embeds(prompt_embeds, attention_mask, _cut_context) + prompt_embeds = prompt_embeds * attention_mask.unsqueeze(2) + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + attention_mask = attention_mask.repeat(num_images_per_prompt, 1) + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + if negative_prompt is not None: + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=128, + truncation=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids = uncond_input.input_ids.to(device) + negative_attention_mask = uncond_input.attention_mask.to(device) + + negative_prompt_embeds = self.text_encoder( + text_input_ids, + attention_mask=negative_attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds[:, : prompt_embeds.shape[1]] + negative_attention_mask = negative_attention_mask[:, : prompt_embeds.shape[1]] + negative_prompt_embeds = negative_prompt_embeds * negative_attention_mask.unsqueeze(2) + + else: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_attention_mask = torch.zeros_like(attention_mask) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + if negative_prompt_embeds.shape != prompt_embeds.shape: + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + negative_attention_mask = negative_attention_mask.repeat(num_images_per_prompt, 1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + else: + negative_prompt_embeds = None + negative_attention_mask = None + return prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask + + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def check_inputs( + self, + prompt, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + attention_mask=None, + negative_attention_mask=None, + ): + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if negative_prompt_embeds is not None and negative_attention_mask is None: + raise ValueError("Please provide `negative_attention_mask` along with `negative_prompt_embeds`") + + if negative_prompt_embeds is not None and negative_attention_mask is not None: + if negative_prompt_embeds.shape[:2] != negative_attention_mask.shape: + raise ValueError( + "`negative_prompt_embeds` and `negative_attention_mask` must have the same batch_size and token length when passed directly, but" + f" got: `negative_prompt_embeds` {negative_prompt_embeds.shape[:2]} != `negative_attention_mask`" + f" {negative_attention_mask.shape}." + ) + + if prompt_embeds is not None and attention_mask is None: + raise ValueError("Please provide `attention_mask` along with `prompt_embeds`") + + if prompt_embeds is not None and attention_mask is not None: + if prompt_embeds.shape[:2] != attention_mask.shape: + raise ValueError( + "`prompt_embeds` and `attention_mask` must have the same batch_size and token length when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape[:2]} != `attention_mask`" + f" {attention_mask.shape}." + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + num_inference_steps: int = 25, + guidance_scale: float = 3.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + height: int | None = 1024, + width: int | None = 1024, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + attention_mask: torch.Tensor | None = None, + negative_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + latents=None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 3.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask. Must provide if passing `prompt_embeds` directly. + negative_attention_mask (`torch.Tensor`, *optional*): + Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + cut_context = True + device = self._execution_device + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + attention_mask, + negative_attention_mask, + ) + + self._guidance_scale = guidance_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + num_images_per_prompt=num_images_per_prompt, + device=device, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + _cut_context=cut_context, + attention_mask=attention_mask, + negative_attention_mask=negative_attention_mask, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + attention_mask = torch.cat([negative_attention_mask, attention_mask]).bool() + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latents + height, width = downscale_height_and_width(height, width, 8) + + latents = self.prepare_latents( + (batch_size * num_images_per_prompt, 4, height, width), + prompt_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + + if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None: + self.text_encoder_offload_hook.offload() + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=attention_mask, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + + noise_pred = (guidance_scale + 1.0) * noise_pred_text - guidance_scale * noise_pred_uncond + # noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + attention_mask = callback_outputs.pop("attention_mask", attention_mask) + negative_attention_mask = callback_outputs.pop("negative_attention_mask", negative_attention_mask) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # post-processing + if output_type not in ["pt", "np", "pil", "latent"]: + raise ValueError( + f"Only the output types `pt`, `pil`, `np` and `latent` are supported not output_type={output_type}" + ) + + if not output_type == "latent": + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + + if output_type in ["np", "pil"]: + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + else: + image = latents + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky3/pipeline_kandinsky3_img2img.py b/diffusers/pipelines/kandinsky3/pipeline_kandinsky3_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..beb4caafb6d323f6eece7ff4d5a00d74a28f0c64 --- /dev/null +++ b/diffusers/pipelines/kandinsky3/pipeline_kandinsky3_img2img.py @@ -0,0 +1,634 @@ +import inspect +from typing import Callable + +import PIL +import PIL.Image +import torch +from transformers import T5EncoderModel, T5Tokenizer + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin +from ...models import Kandinsky3UNet, VQModel +from ...schedulers import DDPMScheduler +from ...utils import ( + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import AutoPipelineForImage2Image + >>> from diffusers.utils import load_image + >>> import torch + + >>> pipe = AutoPipelineForImage2Image.from_pretrained( + ... "kandinsky-community/kandinsky-3", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe.enable_model_cpu_offload() + + >>> prompt = "A painting of the inside of a subway train with tiny raccoons." + >>> image = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky3/t2i.png" + ... ) + + >>> generator = torch.Generator(device="cpu").manual_seed(0) + >>> image = pipe(prompt, image=image, strength=0.75, num_inference_steps=25, generator=generator).images[0] + ``` +""" + + +class Kandinsky3Img2ImgPipeline(DiffusionPipeline, StableDiffusionLoraLoaderMixin): + model_cpu_offload_seq = "text_encoder->movq->unet->movq" + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "negative_attention_mask", + "attention_mask", + ] + + def __init__( + self, + tokenizer: T5Tokenizer, + text_encoder: T5EncoderModel, + unet: Kandinsky3UNet, + scheduler: DDPMScheduler, + movq: VQModel, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, unet=unet, scheduler=scheduler, movq=movq + ) + movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1) if getattr(self, "movq", None) else 8 + movq_latent_channels = self.movq.config.latent_channels if getattr(self, "movq", None) else 4 + self.image_processor = VaeImageProcessor( + vae_scale_factor=movq_scale_factor, + vae_latent_channels=movq_latent_channels, + resample="bicubic", + reducing_gap=1, + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start:] + + return timesteps, num_inference_steps - t_start + + def _process_embeds(self, embeddings, attention_mask, cut_context): + # return embeddings, attention_mask + if cut_context: + embeddings[attention_mask == 0] = torch.zeros_like(embeddings[attention_mask == 0]) + max_seq_length = attention_mask.sum(-1).max() + 1 + embeddings = embeddings[:, :max_seq_length] + attention_mask = attention_mask[:, :max_seq_length] + return embeddings, attention_mask + + @torch.no_grad() + def encode_prompt( + self, + prompt, + do_classifier_free_guidance=True, + num_images_per_prompt=1, + device=None, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + _cut_context=True, + attention_mask: torch.Tensor | None = None, + negative_attention_mask: torch.Tensor | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead. + Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask. Must provide if passing `prompt_embeds` directly. + negative_attention_mask (`torch.Tensor`, *optional*): + Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly. + """ + if prompt is not None and negative_prompt is not None: + if type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + + if device is None: + device = self._execution_device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + max_length = 128 + + if prompt_embeds is None: + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids.to(device) + attention_mask = text_inputs.attention_mask.to(device) + prompt_embeds = self.text_encoder( + text_input_ids, + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + prompt_embeds, attention_mask = self._process_embeds(prompt_embeds, attention_mask, _cut_context) + prompt_embeds = prompt_embeds * attention_mask.unsqueeze(2) + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + attention_mask = attention_mask.repeat(num_images_per_prompt, 1) + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + if negative_prompt is not None: + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=128, + truncation=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids = uncond_input.input_ids.to(device) + negative_attention_mask = uncond_input.attention_mask.to(device) + + negative_prompt_embeds = self.text_encoder( + text_input_ids, + attention_mask=negative_attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds[:, : prompt_embeds.shape[1]] + negative_attention_mask = negative_attention_mask[:, : prompt_embeds.shape[1]] + negative_prompt_embeds = negative_prompt_embeds * negative_attention_mask.unsqueeze(2) + + else: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_attention_mask = torch.zeros_like(attention_mask) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + if negative_prompt_embeds.shape != prompt_embeds.shape: + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + negative_attention_mask = negative_attention_mask.repeat(num_images_per_prompt, 1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + else: + negative_prompt_embeds = None + negative_attention_mask = None + return prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask + + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + self.movq.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = self.movq.encode(image).latent_dist.sample(generator) + + init_latents = self.movq.config.scaling_factor * init_latents + + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + attention_mask=None, + negative_attention_mask=None, + ): + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if negative_prompt_embeds is not None and negative_attention_mask is None: + raise ValueError("Please provide `negative_attention_mask` along with `negative_prompt_embeds`") + + if negative_prompt_embeds is not None and negative_attention_mask is not None: + if negative_prompt_embeds.shape[:2] != negative_attention_mask.shape: + raise ValueError( + "`negative_prompt_embeds` and `negative_attention_mask` must have the same batch_size and token length when passed directly, but" + f" got: `negative_prompt_embeds` {negative_prompt_embeds.shape[:2]} != `negative_attention_mask`" + f" {negative_attention_mask.shape}." + ) + + if prompt_embeds is not None and attention_mask is None: + raise ValueError("Please provide `attention_mask` along with `prompt_embeds`") + + if prompt_embeds is not None and attention_mask is not None: + if prompt_embeds.shape[:2] != attention_mask.shape: + raise ValueError( + "`prompt_embeds` and `attention_mask` must have the same batch_size and token length when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape[:2]} != `attention_mask`" + f" {attention_mask.shape}." + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image] = None, + strength: float = 0.3, + num_inference_steps: int = 25, + guidance_scale: float = 3.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + attention_mask: torch.Tensor | None = None, + negative_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 3.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask. Must provide if passing `prompt_embeds` directly. + negative_attention_mask (`torch.Tensor`, *optional*): + Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` + + """ + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + cut_context = True + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + attention_mask, + negative_attention_mask, + ) + + self._guidance_scale = guidance_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + num_images_per_prompt=num_images_per_prompt, + device=device, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + _cut_context=cut_context, + attention_mask=attention_mask, + negative_attention_mask=negative_attention_mask, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + attention_mask = torch.cat([negative_attention_mask, attention_mask]).bool() + if not isinstance(image, list): + image = [image] + if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image): + raise ValueError( + f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor" + ) + + image = torch.cat([self.image_processor.preprocess(i) for i in image], dim=0) + image = image.to(dtype=prompt_embeds.dtype, device=device) + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + # 5. Prepare latents + latents = self.movq.encode(image)["latents"] + latents = latents.repeat_interleave(num_images_per_prompt, dim=0) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + latents = self.prepare_latents( + latents, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator + ) + if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None: + self.text_encoder_offload_hook.offload() + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=attention_mask, + )[0] + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + + noise_pred = (guidance_scale + 1.0) * noise_pred_text - guidance_scale * noise_pred_uncond + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + noise_pred, + t, + latents, + generator=generator, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + attention_mask = callback_outputs.pop("attention_mask", attention_mask) + negative_attention_mask = callback_outputs.pop("negative_attention_mask", negative_attention_mask) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # post-processing + if not output_type == "latent": + image = self.movq.decode(latents, force_not_quantize=True)["sample"] + image = self.image_processor.postprocess(image, output_type) + else: + image = latents + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/kandinsky5/__init__.py b/diffusers/pipelines/kandinsky5/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..d417ed932b92172d88d5512c9dca8525dc3a7852 --- /dev/null +++ b/diffusers/pipelines/kandinsky5/__init__.py @@ -0,0 +1,54 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_kandinsky"] = ["Kandinsky5T2VPipeline"] + _import_structure["pipeline_kandinsky_i2i"] = ["Kandinsky5I2IPipeline"] + _import_structure["pipeline_kandinsky_i2v"] = ["Kandinsky5I2VPipeline"] + _import_structure["pipeline_kandinsky_t2i"] = ["Kandinsky5T2IPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_kandinsky import Kandinsky5T2VPipeline + from .pipeline_kandinsky_i2i import Kandinsky5I2IPipeline + from .pipeline_kandinsky_i2v import Kandinsky5I2VPipeline + from .pipeline_kandinsky_t2i import Kandinsky5T2IPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/kandinsky5/pipeline_kandinsky.py b/diffusers/pipelines/kandinsky5/pipeline_kandinsky.py new file mode 100644 index 0000000000000000000000000000000000000000..1ce885b21f5be8597633e9c152a3c55c90ed87a6 --- /dev/null +++ b/diffusers/pipelines/kandinsky5/pipeline_kandinsky.py @@ -0,0 +1,977 @@ +# Copyright 2025 The Kandinsky Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Callable + +import regex as re +import torch +from torch.nn import functional as F +from transformers import CLIPTextModel, CLIPTokenizer, Qwen2_5_VLForConditionalGeneration, Qwen2VLProcessor + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import KandinskyLoraLoaderMixin +from ...models import AutoencoderKLHunyuanVideo +from ...models.transformers import Kandinsky5Transformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler + +# Add imports for offloading and tiling +from ...utils import ( + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import KandinskyPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + + ```python + >>> import torch + >>> from diffusers import Kandinsky5T2VPipeline + >>> from diffusers.utils import export_to_video + + >>> # Available models: + >>> # kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-nocfg-5s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-pretrain-5s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-sft-10s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-nocfg-10s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-10s-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2V-Lite-pretrain-10s-Diffusers + + >>> model_id = "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers" + >>> pipe = Kandinsky5T2VPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen." + >>> negative_prompt = "Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards" + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=512, + ... width=768, + ... num_frames=121, + ... num_inference_steps=50, + ... guidance_scale=5.0, + ... ).frames[0] + + >>> export_to_video(output, "output.mp4", fps=24, quality=9) + ``` +""" + + +def basic_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Clean text using ftfy if available and unescape HTML entities. + """ + if is_ftfy_available(): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Normalize whitespace in text by replacing multiple spaces with single space. + """ + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Apply both basic cleaning and whitespace normalization to prompts. + """ + text = whitespace_clean(basic_clean(text)) + return text + + +class Kandinsky5T2VPipeline(DiffusionPipeline, KandinskyLoraLoaderMixin): + r""" + Pipeline for text-to-video generation using Kandinsky 5.0. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + transformer ([`Kandinsky5Transformer3DModel`]): + Conditional Transformer to denoise the encoded video latents. + vae ([`AutoencoderKLHunyuanVideo`]): + Variational Auto-Encoder Model [hunyuanvideo-community/HunyuanVideo + (vae)](https://huggingface.co/hunyuanvideo-community/HunyuanVideo) to encode and decode videos to and from + latent representations. + text_encoder ([`Qwen2_5_VLForConditionalGeneration`]): + Frozen text-encoder [Qwen2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct). + tokenizer ([`AutoProcessor`]): + Tokenizer for Qwen2.5-VL. + text_encoder_2 ([`CLIPTextModel`]): + Frozen [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 ([`CLIPTokenizer`]): + Tokenizer for CLIP. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded video latents. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = [ + "latents", + "prompt_embeds_qwen", + "prompt_embeds_clip", + "negative_prompt_embeds_qwen", + "negative_prompt_embeds_clip", + ] + + def __init__( + self, + transformer: Kandinsky5Transformer3DModel, + vae: AutoencoderKLHunyuanVideo, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2VLProcessor, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + transformer=transformer, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + scheduler=scheduler, + ) + + self.prompt_template = "\n".join( + [ + "<|im_start|>system\nYou are a promt engineer. Describe the video in detail.", + "Describe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.", + "Describe the location of the video, main characters or objects and their action.", + "Describe the dynamism of the video and presented actions.", + "Name the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or scren content.", + "Describe the visual effects, postprocessing and transitions if they are presented in the video.", + "Pay attention to the order of key actions shown in the scene.<|im_end|>", + "<|im_start|>user\n{}<|im_end|>", + ] + ) + self.prompt_template_encode_start_idx = 129 + + self.vae_scale_factor_temporal = ( + self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4 + ) + self.vae_scale_factor_spatial = self.vae.config.spatial_compression_ratio if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + def _get_scale_factor(self, height: int, width: int) -> tuple: + """ + Calculate the scale factor based on resolution. + + Args: + height (int): Video height + width (int): Video width + + Returns: + tuple: Scale factor as (temporal_scale, height_scale, width_scale) + """ + + def between_480p(x): + return 480 <= x <= 854 + + if between_480p(height) and between_480p(width): + return (1, 2, 2) + else: + return (1, 3.16, 3.16) + + @staticmethod + def fast_sta_nabla(T: int, H: int, W: int, wT: int = 3, wH: int = 3, wW: int = 3, device="cuda") -> torch.Tensor: + """ + Create a sparse temporal attention (STA) mask for efficient video generation. + + This method generates a mask that limits attention to nearby frames and spatial positions, reducing + computational complexity for video generation. + + Args: + T (int): Number of temporal frames + H (int): Height in latent space + W (int): Width in latent space + wT (int): Temporal attention window size + wH (int): Height attention window size + wW (int): Width attention window size + device (str): Device to create tensor on + + Returns: + torch.Tensor: Sparse attention mask of shape (T*H*W, T*H*W) + """ + l = torch.Tensor([T, H, W]).amax() + r = torch.arange(0, l, 1, dtype=torch.int16, device=device) + mat = (r.unsqueeze(1) - r.unsqueeze(0)).abs() + sta_t, sta_h, sta_w = ( + mat[:T, :T].flatten(), + mat[:H, :H].flatten(), + mat[:W, :W].flatten(), + ) + sta_t = sta_t <= wT // 2 + sta_h = sta_h <= wH // 2 + sta_w = sta_w <= wW // 2 + sta_hw = (sta_h.unsqueeze(1) * sta_w.unsqueeze(0)).reshape(H, H, W, W).transpose(1, 2).flatten() + sta = (sta_t.unsqueeze(1) * sta_hw.unsqueeze(0)).reshape(T, T, H * W, H * W).transpose(1, 2) + return sta.reshape(T * H * W, T * H * W) + + def get_sparse_params(self, sample, device): + """ + Generate sparse attention parameters for the transformer based on sample dimensions. + + This method computes the sparse attention configuration needed for efficient video processing in the + transformer model. + + Args: + sample (torch.Tensor): Input sample tensor + device (torch.device): Device to place tensors on + + Returns: + Dict: Dictionary containing sparse attention parameters + """ + assert self.transformer.config.patch_size[0] == 1 + B, T, H, W, _ = sample.shape + T, H, W = ( + T // self.transformer.config.patch_size[0], + H // self.transformer.config.patch_size[1], + W // self.transformer.config.patch_size[2], + ) + if self.transformer.config.attention_type == "nabla": + sta_mask = self.fast_sta_nabla( + T, + H // 8, + W // 8, + self.transformer.config.attention_wT, + self.transformer.config.attention_wH, + self.transformer.config.attention_wW, + device=device, + ) + + sparse_params = { + "sta_mask": sta_mask.unsqueeze_(0).unsqueeze_(0), + "attention_type": self.transformer.config.attention_type, + "to_fractal": True, + "P": self.transformer.config.attention_P, + "wT": self.transformer.config.attention_wT, + "wW": self.transformer.config.attention_wW, + "wH": self.transformer.config.attention_wH, + "add_sta": self.transformer.config.attention_add_sta, + "visual_shape": (T, H, W), + "method": self.transformer.config.attention_method, + } + else: + sparse_params = None + + return sparse_params + + def _encode_prompt_qwen( + self, + prompt: str | list[str], + device: torch.device | None = None, + max_sequence_length: int = 256, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using Qwen2.5-VL text encoder. + + This method processes the input prompt through the Qwen2.5-VL model to generate text embeddings suitable for + video generation. + + Args: + prompt (str | list[str]): Input prompt or list of prompts + device (torch.device): Device to run encoding on + num_videos_per_prompt (int): Number of videos to generate per prompt + max_sequence_length (int): Maximum sequence length for tokenization + dtype (torch.dtype): Data type for embeddings + + Returns: + tuple[torch.Tensor, torch.Tensor]: Text embeddings and cumulative sequence lengths + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + full_texts = [self.prompt_template.format(p) for p in prompt] + max_allowed_len = self.prompt_template_encode_start_idx + max_sequence_length + + untruncated_ids = self.tokenizer( + text=full_texts, + images=None, + videos=None, + return_tensors="pt", + padding="longest", + )["input_ids"] + + if untruncated_ids.shape[-1] > max_allowed_len: + for i, text in enumerate(full_texts): + tokens = untruncated_ids[i][self.prompt_template_encode_start_idx : -2] + removed_text = self.tokenizer.decode(tokens[max_sequence_length - 2 :]) + if len(removed_text) > 0: + full_texts[i] = text[: -len(removed_text)] + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + inputs = self.tokenizer( + text=full_texts, + images=None, + videos=None, + max_length=max_allowed_len, + truncation=True, + return_tensors="pt", + padding=True, + ).to(device) + + embeds = self.text_encoder( + input_ids=inputs["input_ids"], + return_dict=True, + output_hidden_states=True, + )["hidden_states"][-1][:, self.prompt_template_encode_start_idx :] + + attention_mask = inputs["attention_mask"][:, self.prompt_template_encode_start_idx :] + cu_seqlens = torch.cumsum(attention_mask.sum(1), dim=0) + cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0).to(dtype=torch.int32) + + return embeds.to(dtype), cu_seqlens + + def _encode_prompt_clip( + self, + prompt: str | list[str], + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using CLIP text encoder. + + This method processes the input prompt through the CLIP model to generate pooled embeddings that capture + semantic information. + + Args: + prompt (str | list[str]): Input prompt or list of prompts + device (torch.device): Device to run encoding on + num_videos_per_prompt (int): Number of videos to generate per prompt + dtype (torch.dtype): Data type for embeddings + + Returns: + torch.Tensor: Pooled text embeddings from CLIP + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + inputs = self.tokenizer_2( + prompt, + max_length=77, + truncation=True, + add_special_tokens=True, + padding="max_length", + return_tensors="pt", + ).to(device) + + pooled_embed = self.text_encoder_2(**inputs)["pooler_output"] + + return pooled_embed.to(dtype) + + def encode_prompt( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes a single prompt (positive or negative) into text encoder hidden states. + + This method combines embeddings from both Qwen2.5-VL and CLIP text encoders to create comprehensive text + representations for video generation. + + Args: + prompt (`str` or `list[str]`): + Prompt to be encoded. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos to generate per prompt. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length for text encoding. + device (`torch.device`, *optional*): + Torch device. + dtype (`torch.dtype`, *optional*): + Torch dtype. + + Returns: + tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + - Qwen text embeddings of shape (batch_size * num_videos_per_prompt, sequence_length, embedding_dim) + - CLIP pooled embeddings of shape (batch_size * num_videos_per_prompt, clip_embedding_dim) + - Cumulative sequence lengths (`cu_seqlens`) for Qwen embeddings of shape (batch_size * + num_videos_per_prompt + 1,) + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + if not isinstance(prompt, list): + prompt = [prompt] + + batch_size = len(prompt) + + prompt = [prompt_clean(p) for p in prompt] + + # Encode with Qwen2.5-VL + prompt_embeds_qwen, prompt_cu_seqlens = self._encode_prompt_qwen( + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + dtype=dtype, + ) + # prompt_embeds_qwen shape: [batch_size, seq_len, embed_dim] + + # Encode with CLIP + prompt_embeds_clip = self._encode_prompt_clip( + prompt=prompt, + device=device, + dtype=dtype, + ) + # prompt_embeds_clip shape: [batch_size, clip_embed_dim] + + # Repeat embeddings for num_videos_per_prompt + # Qwen embeddings: repeat sequence for each video, then reshape + prompt_embeds_qwen = prompt_embeds_qwen.repeat( + 1, num_videos_per_prompt, 1 + ) # [batch_size, seq_len * num_videos_per_prompt, embed_dim] + # Reshape to [batch_size * num_videos_per_prompt, seq_len, embed_dim] + prompt_embeds_qwen = prompt_embeds_qwen.view( + batch_size * num_videos_per_prompt, -1, prompt_embeds_qwen.shape[-1] + ) + + # CLIP embeddings: repeat for each video + prompt_embeds_clip = prompt_embeds_clip.repeat( + 1, num_videos_per_prompt, 1 + ) # [batch_size, num_videos_per_prompt, clip_embed_dim] + # Reshape to [batch_size * num_videos_per_prompt, clip_embed_dim] + prompt_embeds_clip = prompt_embeds_clip.view(batch_size * num_videos_per_prompt, -1) + + # Repeat cumulative sequence lengths for num_videos_per_prompt + # Original cu_seqlens: [0, len1, len1+len2, ...] + # Need to repeat the differences and reconstruct for repeated prompts + # Original differences (lengths) for each prompt in the batch + original_lengths = prompt_cu_seqlens.diff() # [len1, len2, ...] + # Repeat the lengths for num_videos_per_prompt + repeated_lengths = original_lengths.repeat_interleave( + num_videos_per_prompt + ) # [len1, len1, ..., len2, len2, ...] + # Reconstruct the cumulative lengths + repeated_cu_seqlens = torch.cat( + [torch.tensor([0], device=device, dtype=torch.int32), repeated_lengths.cumsum(0)] + ) + + return prompt_embeds_qwen, prompt_embeds_clip, repeated_cu_seqlens + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds_qwen=None, + prompt_embeds_clip=None, + negative_prompt_embeds_qwen=None, + negative_prompt_embeds_clip=None, + prompt_cu_seqlens=None, + negative_prompt_cu_seqlens=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + """ + Validate input parameters for the pipeline. + + Args: + prompt: Input prompt + negative_prompt: Negative prompt for guidance + height: Video height + width: Video width + prompt_embeds_qwen: Pre-computed Qwen prompt embeddings + prompt_embeds_clip: Pre-computed CLIP prompt embeddings + negative_prompt_embeds_qwen: Pre-computed Qwen negative prompt embeddings + negative_prompt_embeds_clip: Pre-computed CLIP negative prompt embeddings + prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen positive prompt + negative_prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen negative prompt + callback_on_step_end_tensor_inputs: Callback tensor inputs + + Raises: + ValueError: If inputs are invalid + """ + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError("max_sequence_length must be less than 1024") + + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Check for consistency within positive prompt embeddings and sequence lengths + if prompt_embeds_qwen is not None or prompt_embeds_clip is not None or prompt_cu_seqlens is not None: + if prompt_embeds_qwen is None or prompt_embeds_clip is None or prompt_cu_seqlens is None: + raise ValueError( + "If any of `prompt_embeds_qwen`, `prompt_embeds_clip`, or `prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check for consistency within negative prompt embeddings and sequence lengths + if ( + negative_prompt_embeds_qwen is not None + or negative_prompt_embeds_clip is not None + or negative_prompt_cu_seqlens is not None + ): + if ( + negative_prompt_embeds_qwen is None + or negative_prompt_embeds_clip is None + or negative_prompt_cu_seqlens is None + ): + raise ValueError( + "If any of `negative_prompt_embeds_qwen`, `negative_prompt_embeds_clip`, or `negative_prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check if prompt or embeddings are provided (either prompt or all required embedding components for positive) + if prompt is None and prompt_embeds_qwen is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_qwen` (and corresponding `prompt_embeds_clip` and `prompt_cu_seqlens`). Cannot leave all undefined." + ) + + # Validate types for prompt and negative_prompt if provided + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + if negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + Prepare initial latent variables for video generation. + + This method creates random noise latents or uses provided latents as starting point for the denoising process. + + Args: + batch_size (int): Number of videos to generate + num_channels_latents (int): Number of channels in latent space + height (int): Height of generated video + width (int): Width of generated video + num_frames (int): Number of frames in video + dtype (torch.dtype): Data type for latents + device (torch.device): Device to create latents on + generator (torch.Generator): Random number generator + latents (torch.Tensor): Pre-existing latents to use + + Returns: + torch.Tensor: Prepared latent tensor + """ + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + num_channels_latents, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + if self.transformer.visual_cond: + # For visual conditioning, concatenate with zeros and mask + visual_cond = torch.zeros_like(latents) + visual_cond_mask = torch.zeros( + [ + batch_size, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + 1, + ], + dtype=latents.dtype, + device=latents.device, + ) + latents = torch.cat([latents, visual_cond, visual_cond_mask], dim=-1) + + return latents + + @property + def guidance_scale(self): + """Get the current guidance scale value.""" + return self._guidance_scale + + @property + def num_timesteps(self): + """Get the number of denoising timesteps.""" + return self._num_timesteps + + @property + def interrupt(self): + """Check if generation has been interrupted.""" + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds_qwen: torch.Tensor | None = None, + prompt_embeds_clip: torch.Tensor | None = None, + negative_prompt_embeds_qwen: torch.Tensor | None = None, + negative_prompt_embeds_clip: torch.Tensor | None = None, + prompt_cu_seqlens: torch.Tensor | None = None, + negative_prompt_cu_seqlens: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to avoid during video generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`, defaults to `512`): + The height in pixels of the generated video. + width (`int`, defaults to `768`): + The width in pixels of the generated video. + num_frames (`int`, defaults to `25`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in classifier-free guidance. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A torch generator to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents. + prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated text embeddings from the Qwen text encoder. + prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated text embeddings from the CLIP text encoder. + negative_prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings from the Qwen text encoder. + negative_prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings from the CLIP text encoder. + prompt_cu_seqlens (`torch.Tensor`, *optional*): + Cumulative sequence lengths for the Qwen prompt embeddings, used for variable-length attention. + negative_prompt_cu_seqlens (`torch.Tensor`, *optional*): + Cumulative sequence lengths for the Qwen negative prompt embeddings, used for variable-length + attention. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated video. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`KandinskyPipelineOutput`]. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function that is called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length for text encoding. + + Examples: + + Returns: + [`~KandinskyPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`KandinskyPipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images. + """ + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + negative_prompt=negative_prompt, + height=height, + width=width, + prompt_embeds_qwen=prompt_embeds_qwen, + prompt_embeds_clip=prompt_embeds_clip, + negative_prompt_embeds_qwen=negative_prompt_embeds_qwen, + negative_prompt_embeds_clip=negative_prompt_embeds_clip, + prompt_cu_seqlens=prompt_cu_seqlens, + negative_prompt_cu_seqlens=negative_prompt_cu_seqlens, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._interrupt = False + + device = self._execution_device + dtype = self.transformer.dtype + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + prompt = [prompt] + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds_qwen.shape[0] + + # 3. Encode input prompt + if prompt_embeds_qwen is None: + prompt_embeds_qwen, prompt_embeds_clip, prompt_cu_seqlens = self.encode_prompt( + prompt=prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if self.guidance_scale > 1.0: + if negative_prompt is None: + negative_prompt = "Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards" + + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * len(prompt) if prompt is not None else [negative_prompt] + elif len(negative_prompt) != len(prompt): + raise ValueError( + f"`negative_prompt` must have same length as `prompt`. Got {len(negative_prompt)} vs {len(prompt)}." + ) + + if negative_prompt_embeds_qwen is None: + negative_prompt_embeds_qwen, negative_prompt_embeds_clip, negative_prompt_cu_seqlens = ( + self.encode_prompt( + prompt=negative_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_visual_dim + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + dtype, + device, + generator, + latents, + ) + + # 6. Prepare rope positions for positional encoding + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + visual_rope_pos = [ + torch.arange(num_latent_frames, device=device), + torch.arange(height // self.vae_scale_factor_spatial // 2, device=device), + torch.arange(width // self.vae_scale_factor_spatial // 2, device=device), + ] + + text_rope_pos = torch.arange(prompt_cu_seqlens.diff().max().item(), device=device) + + negative_text_rope_pos = ( + torch.arange(negative_prompt_cu_seqlens.diff().max().item(), device=device) + if negative_prompt_cu_seqlens is not None + else None + ) + + # 7. Calculate dynamic scale factor based on resolution + scale_factor = self._get_scale_factor(height, width) + + # 8. Sparse Params for efficient attention + sparse_params = self.get_sparse_params(latents, device) + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + timestep = t.unsqueeze(0).repeat(batch_size * num_videos_per_prompt) + + # Predict noise residual + pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=prompt_embeds_qwen.to(dtype), + pooled_projections=prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + if self.guidance_scale > 1.0 and negative_prompt_embeds_qwen is not None: + uncond_pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=negative_prompt_embeds_qwen.to(dtype), + pooled_projections=negative_prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=negative_text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + pred_velocity = uncond_pred_velocity + guidance_scale * (pred_velocity - uncond_pred_velocity) + # Compute previous sample using the scheduler + latents[:, :, :, :, :num_channels_latents] = self.scheduler.step( + pred_velocity, t, latents[:, :, :, :, :num_channels_latents], return_dict=False + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds_qwen = callback_outputs.pop("prompt_embeds_qwen", prompt_embeds_qwen) + prompt_embeds_clip = callback_outputs.pop("prompt_embeds_clip", prompt_embeds_clip) + negative_prompt_embeds_qwen = callback_outputs.pop( + "negative_prompt_embeds_qwen", negative_prompt_embeds_qwen + ) + negative_prompt_embeds_clip = callback_outputs.pop( + "negative_prompt_embeds_clip", negative_prompt_embeds_clip + ) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 10. Post-processing - extract main latents + latents = latents[:, :, :, :, :num_channels_latents] + + # 11. Decode latents to video + if output_type != "latent": + latents = latents.to(self.vae.dtype) + # Reshape and normalize latents + video = latents.reshape( + batch_size, + num_videos_per_prompt, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + num_channels_latents, + ) + video = video.permute(0, 1, 5, 2, 3, 4) # [batch, num_videos, channels, frames, height, width] + video = video.reshape( + batch_size * num_videos_per_prompt, + num_channels_latents, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + ) + + # Normalize and decode through VAE + video = video / self.vae.config.scaling_factor + video = self.vae.decode(video).sample + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return KandinskyPipelineOutput(frames=video) diff --git a/diffusers/pipelines/kandinsky5/pipeline_kandinsky_i2i.py b/diffusers/pipelines/kandinsky5/pipeline_kandinsky_i2i.py new file mode 100644 index 0000000000000000000000000000000000000000..0a8382d6031fb05271c01e877131524c83a28a03 --- /dev/null +++ b/diffusers/pipelines/kandinsky5/pipeline_kandinsky_i2i.py @@ -0,0 +1,862 @@ +# Copyright 2025 The Kandinsky Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Callable + +import numpy as np +import regex as re +import torch +from torch.nn import functional as F +from transformers import CLIPTextModel, CLIPTokenizer, Qwen2_5_VLForConditionalGeneration, Qwen2VLProcessor + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import KandinskyLoraLoaderMixin +from ...models import AutoencoderKL +from ...models.transformers import Kandinsky5Transformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler + +# Add imports for offloading and tiling +from ...utils import ( + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import KandinskyImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + + ```python + >>> import torch + >>> from diffusers import Kandinsky5I2IPipeline + + >>> # Available models: + >>> # kandinskylab/Kandinsky-5.0-I2I-Lite-sft-Diffusers + >>> # kandinskylab/Kandinsky-5.0-I2I-Lite-pretrain-Diffusers + + >>> model_id = "kandinskylab/Kandinsky-5.0-I2I-Lite-sft-Diffusers" + >>> pipe = Kandinsky5I2IPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen." + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt="", + ... height=1024, + ... width=1024, + ... num_inference_steps=50, + ... guidance_scale=3.5, + ... ).frames[0] + ``` +""" + + +def basic_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Clean text using ftfy if available and unescape HTML entities. + """ + if is_ftfy_available(): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Normalize whitespace in text by replacing multiple spaces with single space. + """ + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Apply both basic cleaning and whitespace normalization to prompts. + """ + text = whitespace_clean(basic_clean(text)) + return text + + +class Kandinsky5I2IPipeline(DiffusionPipeline, KandinskyLoraLoaderMixin): + r""" + Pipeline for image-to-image generation using Kandinsky 5.0. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + transformer ([`Kandinsky5Transformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder Model [black-forest-labs/FLUX.1-dev + (vae)](https://huggingface.co/black-forest-labs/FLUX.1-dev) to encode and decode videos to and from latent + representations. + text_encoder ([`Qwen2_5_VLForConditionalGeneration`]): + Frozen text-encoder [Qwen2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct). + tokenizer ([`AutoProcessor`]): + Tokenizer for Qwen2.5-VL. + text_encoder_2 ([`CLIPTextModel`]): + Frozen [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 ([`CLIPTokenizer`]): + Tokenizer for CLIP. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = [ + "latents", + "prompt_embeds_qwen", + "prompt_embeds_clip", + "negative_prompt_embeds_qwen", + "negative_prompt_embeds_clip", + ] + + def __init__( + self, + transformer: Kandinsky5Transformer3DModel, + vae: AutoencoderKL, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2VLProcessor, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + transformer=transformer, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + scheduler=scheduler, + ) + self.prompt_template = "<|im_start|>system\nYou are a promt engineer. Based on the provided source image (first image) and target image (second image), create an interesting text prompt that can be used together with the source image to create the target image:<|im_end|><|im_start|>user{}<|vision_start|><|image_pad|><|vision_end|><|im_end|>" + self.prompt_template_encode_start_idx = 55 + + self.vae_scale_factor_spatial = 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.resolutions = [(1024, 1024), (640, 1408), (1408, 640), (768, 1280), (1280, 768), (896, 1152), (1152, 896)] + + def _encode_prompt_qwen( + self, + prompt: list[str], + image: PipelineImageInput | None = None, + device: torch.device | None = None, + max_sequence_length: int = 1024, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using Qwen2.5-VL text encoder. + + This method processes the input prompt through the Qwen2.5-VL model to generate text embeddings suitable for + image generation. + + Args: + prompt list[str]: Input list of prompts + image (PipelineImageInput): Input list of images to condition the generation on + device (torch.device): Device to run encoding on + max_sequence_length (int): Maximum sequence length for tokenization + dtype (torch.dtype): Data type for embeddings + + Returns: + tuple[torch.Tensor, torch.Tensor]: Text embeddings and cumulative sequence lengths + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + if not isinstance(image, list): + image = [image] + image = [i.resize((i.size[0] // 2, i.size[1] // 2)) for i in image] + full_texts = [self.prompt_template.format(p) for p in prompt] + max_allowed_len = self.prompt_template_encode_start_idx + max_sequence_length + + untruncated_ids = self.tokenizer( + text=full_texts, + images=image, + videos=None, + return_tensors="pt", + padding="longest", + )["input_ids"] + + if untruncated_ids.shape[-1] > max_allowed_len: + for i, text in enumerate(full_texts): + tokens = untruncated_ids[i] + num_image_tokens = (tokens == self.tokenizer.image_token_id).sum() + tokens = tokens[tokens != self.tokenizer.image_token_id][self.prompt_template_encode_start_idx : -3] + removed_text = self.tokenizer.decode(tokens[max_sequence_length - num_image_tokens - 3 :]) + if len(removed_text) > 0: + full_texts[i] = text[: -len(removed_text)] + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + inputs = self.tokenizer( + text=full_texts, + images=image, + videos=None, + max_length=max_allowed_len, + truncation=True, + return_tensors="pt", + padding=True, + ).to(device) + + embeds = self.text_encoder( + **inputs, + return_dict=True, + output_hidden_states=True, + )["hidden_states"][-1][:, self.prompt_template_encode_start_idx :] + + attention_mask = inputs["attention_mask"][:, self.prompt_template_encode_start_idx :] + cu_seqlens = torch.cumsum(attention_mask.sum(1), dim=0) + cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0).to(dtype=torch.int32) + + return embeds.to(dtype), cu_seqlens + + def _encode_prompt_clip( + self, + prompt: str | list[str], + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using CLIP text encoder. + + This method processes the input prompt through the CLIP model to generate pooled embeddings that capture + semantic information. + + Args: + prompt (str | list[str]): Input prompt or list of prompts + device (torch.device): Device to run encoding on + dtype (torch.dtype): Data type for embeddings + + Returns: + torch.Tensor: Pooled text embeddings from CLIP + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + inputs = self.tokenizer_2( + prompt, + max_length=77, + truncation=True, + add_special_tokens=True, + padding="max_length", + return_tensors="pt", + ).to(device) + + pooled_embed = self.text_encoder_2(**inputs)["pooler_output"] + + return pooled_embed.to(dtype) + + def encode_prompt( + self, + prompt: str | list[str], + image: torch.Tensor, + num_images_per_prompt: int = 1, + max_sequence_length: int = 1024, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes a single prompt (positive or negative) into text encoder hidden states. + + This method combines embeddings from both Qwen2.5-VL and CLIP text encoders to create comprehensive text + representations for image generation. + + Args: + prompt (`str` or `list[str]`): + Prompt to be encoded. + num_images_per_prompt (`int`, *optional*, defaults to 1): + Number of images to generate per prompt. + max_sequence_length (`int`, *optional*, defaults to 1024): + Maximum sequence length for text encoding. Must be less than 1024 + device (`torch.device`, *optional*): + Torch device. + dtype (`torch.dtype`, *optional*): + Torch dtype. + + Returns: + tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + - Qwen text embeddings of shape (batch_size * num_images_per_prompt, sequence_length, embedding_dim) + - CLIP pooled embeddings of shape (batch_size * num_images_per_prompt, clip_embedding_dim) + - Cumulative sequence lengths (`cu_seqlens`) for Qwen embeddings of shape (batch_size * + num_images_per_prompt + 1,) + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + if not isinstance(prompt, list): + prompt = [prompt] + + batch_size = len(prompt) + + prompt = [prompt_clean(p) for p in prompt] + + # Encode with Qwen2.5-VL + prompt_embeds_qwen, prompt_cu_seqlens = self._encode_prompt_qwen( + prompt=prompt, + image=image, + device=device, + max_sequence_length=max_sequence_length, + dtype=dtype, + ) + # prompt_embeds_qwen shape: [batch_size, seq_len, embed_dim] + + # Encode with CLIP + prompt_embeds_clip = self._encode_prompt_clip( + prompt=prompt, + device=device, + dtype=dtype, + ) + # prompt_embeds_clip shape: [batch_size, clip_embed_dim] + + # Repeat embeddings for num_images_per_prompt + # Qwen embeddings: repeat sequence for each image, then reshape + prompt_embeds_qwen = prompt_embeds_qwen.repeat( + 1, num_images_per_prompt, 1 + ) # [batch_size, seq_len * num_images_per_prompt, embed_dim] + # Reshape to [batch_size * num_images_per_prompt, seq_len, embed_dim] + prompt_embeds_qwen = prompt_embeds_qwen.view( + batch_size * num_images_per_prompt, -1, prompt_embeds_qwen.shape[-1] + ) + + # CLIP embeddings: repeat for each image + prompt_embeds_clip = prompt_embeds_clip.repeat( + 1, num_images_per_prompt, 1 + ) # [batch_size, num_images_per_prompt, clip_embed_dim] + # Reshape to [batch_size * num_images_per_prompt, clip_embed_dim] + prompt_embeds_clip = prompt_embeds_clip.view(batch_size * num_images_per_prompt, -1) + + # Repeat cumulative sequence lengths for num_images_per_prompt + # Original differences (lengths) for each prompt in the batch + original_lengths = prompt_cu_seqlens.diff() # [len1, len2, ...] + # Repeat the lengths for num_images_per_prompt + repeated_lengths = original_lengths.repeat_interleave( + num_images_per_prompt + ) # [len1, len1, ..., len2, len2, ...] + # Reconstruct the cumulative lengths + repeated_cu_seqlens = torch.cat( + [torch.tensor([0], device=device, dtype=torch.int32), repeated_lengths.cumsum(0)] + ) + + return prompt_embeds_qwen, prompt_embeds_clip, repeated_cu_seqlens + + def check_inputs( + self, + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds_qwen=None, + prompt_embeds_clip=None, + negative_prompt_embeds_qwen=None, + negative_prompt_embeds_clip=None, + prompt_cu_seqlens=None, + negative_prompt_cu_seqlens=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + """ + Validate input parameters for the pipeline. + + Args: + prompt: Input prompt + negative_prompt: Negative prompt for guidance + image: Input image for conditioning + height: Image height + width: Image width + prompt_embeds_qwen: Pre-computed Qwen prompt embeddings + prompt_embeds_clip: Pre-computed CLIP prompt embeddings + negative_prompt_embeds_qwen: Pre-computed Qwen negative prompt embeddings + negative_prompt_embeds_clip: Pre-computed CLIP negative prompt embeddings + prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen positive prompt + negative_prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen negative prompt + callback_on_step_end_tensor_inputs: Callback tensor inputs + + Raises: + ValueError: If inputs are invalid + """ + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError("max_sequence_length must be less than 1024") + + if image is None: + raise ValueError("`image` must be provided for image-to-image generation") + + if (width, height) not in self.resolutions: + resolutions_str = ",".join([f"({w},{h})" for w, h in self.resolutions]) + logger.warning( + f"`height` and `width` have to be one of {resolutions_str}, but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Check for consistency within positive prompt embeddings and sequence lengths + if prompt_embeds_qwen is not None or prompt_embeds_clip is not None or prompt_cu_seqlens is not None: + if prompt_embeds_qwen is None or prompt_embeds_clip is None or prompt_cu_seqlens is None: + raise ValueError( + "If any of `prompt_embeds_qwen`, `prompt_embeds_clip`, or `prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check for consistency within negative prompt embeddings and sequence lengths + if ( + negative_prompt_embeds_qwen is not None + or negative_prompt_embeds_clip is not None + or negative_prompt_cu_seqlens is not None + ): + if ( + negative_prompt_embeds_qwen is None + or negative_prompt_embeds_clip is None + or negative_prompt_cu_seqlens is None + ): + raise ValueError( + "If any of `negative_prompt_embeds_qwen`, `negative_prompt_embeds_clip`, or `negative_prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check if prompt or embeddings are provided (either prompt or all required embedding components for positive) + if prompt is None and prompt_embeds_qwen is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_qwen` (and corresponding `prompt_embeds_clip` and `prompt_cu_seqlens`). Cannot leave all undefined." + ) + + # Validate types for prompt and negative_prompt if provided + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + if negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + def prepare_latents( + self, + image: PipelineImageInput, + batch_size: int, + num_channels_latents: int = 16, + height: int = 1024, + width: int = 1024, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + Prepare initial latent variables for image-to-image generation. + + This method creates random noise latents with encoded image, + + Args: + image (PipelineImageInput): Input image to condition the generation on + batch_size (int): Number of images to generate + num_channels_latents (int): Number of channels in latent space + height (int): Height of generated image + width (int): Width of generated image + dtype (torch.dtype): Data type for latents + device (torch.device): Device to create latents on + generator (torch.Generator): Random number generator + latents (torch.Tensor): Pre-existing latents to use + + Returns: + torch.Tensor: Prepared latent tensor with encoded image + """ + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + 1, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + num_channels_latents, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Generate random noise for all frames + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # Encode the input image to use as first frame + # Preprocess image + image_tensor = self.image_processor.preprocess(image, height=height, width=width).to(device, dtype=dtype) + # Encode image to latents using VAE + with torch.no_grad(): + image_latents = self.vae.encode(image_tensor).latent_dist.sample(generator=generator) + image_latents = image_latents.unsqueeze(2) # Add temporal dimension + + # Normalize latents if needed + if hasattr(self.vae.config, "scaling_factor"): + image_latents = image_latents * self.vae.config.scaling_factor + + # Reshape to match latent dimensions [batch, 1, height, width, channels] + image_latents = image_latents.permute(0, 2, 3, 4, 1) # [batch, 1, H, W, C] + image_latents = image_latents.to(device=latents.device, dtype=latents.dtype) + latents = torch.cat([latents, image_latents, torch.ones_like(latents[..., :1])], -1) + + return latents + + @property + def guidance_scale(self): + """Get the current guidance scale value.""" + return self._guidance_scale + + @property + def num_timesteps(self): + """Get the number of denoising timesteps.""" + return self._num_timesteps + + @property + def interrupt(self): + """Check if generation has been interrupted.""" + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 3.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds_qwen: torch.Tensor | None = None, + prompt_embeds_clip: torch.Tensor | None = None, + negative_prompt_embeds_qwen: torch.Tensor | None = None, + negative_prompt_embeds_clip: torch.Tensor | None = None, + prompt_cu_seqlens: torch.Tensor | None = None, + negative_prompt_cu_seqlens: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int, None], PipelineCallback | MultiPipelineCallbacks] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 1024, + ): + r""" + The call function to the pipeline for image-to-image generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to avoid during image generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`): + The height in pixels of the generated image. + width (`int`): + The width in pixels of the generated image. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in classifier-free guidance. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A torch generator to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents. + prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated Qwen text embeddings. + prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated CLIP text embeddings. + negative_prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated Qwen negative text embeddings. + negative_prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated CLIP negative text embeddings. + prompt_cu_seqlens (`torch.Tensor`, *optional*): + Pre-generated cumulative sequence lengths for Qwen positive prompt. + negative_prompt_cu_seqlens (`torch.Tensor`, *optional*): + Pre-generated cumulative sequence lengths for Qwen negative prompt. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`KandinskyImagePipelineOutput`]. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function that is called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. + max_sequence_length (`int`, defaults to `1024`): + The maximum sequence length for text and image qwen encoding. Must be less than 1024 + + Examples: + + Returns: + [`~KandinskyImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`KandinskyImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + # 1. Check inputs. Raise error if not correct + if height is None and width is None: + width, height = image[0].size if isinstance(image, list) else image.size + self.check_inputs( + prompt=prompt, + negative_prompt=negative_prompt, + image=image, + height=height, + width=width, + prompt_embeds_qwen=prompt_embeds_qwen, + prompt_embeds_clip=prompt_embeds_clip, + negative_prompt_embeds_qwen=negative_prompt_embeds_qwen, + negative_prompt_embeds_clip=negative_prompt_embeds_clip, + prompt_cu_seqlens=prompt_cu_seqlens, + negative_prompt_cu_seqlens=negative_prompt_cu_seqlens, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + if (width, height) not in self.resolutions: + width, height = self.resolutions[ + np.argmin([abs((i[0] / i[1]) - (width / height)) for i in self.resolutions]) + ] + + self._guidance_scale = guidance_scale + self._interrupt = False + + device = self._execution_device + dtype = self.transformer.dtype + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + prompt = [prompt] + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds_qwen.shape[0] + + # 3. Encode input prompt + if prompt_embeds_qwen is None: + prompt_embeds_qwen, prompt_embeds_clip, prompt_cu_seqlens = self.encode_prompt( + prompt=prompt, + image=image, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if self.guidance_scale > 1.0: + if negative_prompt is None: + negative_prompt = "" + + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * len(prompt) if prompt is not None else [negative_prompt] + elif len(negative_prompt) != len(prompt): + raise ValueError( + f"`negative_prompt` must have same length as `prompt`. Got {len(negative_prompt)} vs {len(prompt)}." + ) + + if negative_prompt_embeds_qwen is None: + negative_prompt_embeds_qwen, negative_prompt_embeds_clip, negative_prompt_cu_seqlens = ( + self.encode_prompt( + prompt=negative_prompt, + image=image, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables with image conditioning + num_channels_latents = self.transformer.config.in_visual_dim + latents = self.prepare_latents( + image=image, + batch_size=batch_size * num_images_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + dtype=dtype, + device=device, + generator=generator, + latents=latents, + ) + + # 6. Prepare rope positions for positional encoding + visual_rope_pos = [ + torch.arange(1, device=device), + torch.arange(height // self.vae_scale_factor_spatial // 2, device=device), + torch.arange(width // self.vae_scale_factor_spatial // 2, device=device), + ] + + text_rope_pos = torch.arange(prompt_cu_seqlens.diff().max().item(), device=device) + + negative_text_rope_pos = ( + torch.arange(negative_prompt_cu_seqlens.diff().max().item(), device=device) + if negative_prompt_cu_seqlens is not None + else None + ) + + # 7. Calculate dynamic scale factor based on resolution + scale_factor = [1.0, 1.0, 1.0] + + # 8. Sparse Params for efficient attention + sparse_params = None + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + timestep = t.unsqueeze(0).repeat(batch_size * num_images_per_prompt) + + # Predict noise residual + pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=prompt_embeds_qwen.to(dtype), + pooled_projections=prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + if self.guidance_scale > 1.0 and negative_prompt_embeds_qwen is not None: + uncond_pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=negative_prompt_embeds_qwen.to(dtype), + pooled_projections=negative_prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=negative_text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + pred_velocity = uncond_pred_velocity + guidance_scale * (pred_velocity - uncond_pred_velocity) + + latents[:, :, :, :, :num_channels_latents] = self.scheduler.step( + pred_velocity[:, :], t, latents[:, :, :, :, :num_channels_latents], return_dict=False + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds_qwen = callback_outputs.pop("prompt_embeds_qwen", prompt_embeds_qwen) + prompt_embeds_clip = callback_outputs.pop("prompt_embeds_clip", prompt_embeds_clip) + negative_prompt_embeds_qwen = callback_outputs.pop( + "negative_prompt_embeds_qwen", negative_prompt_embeds_qwen + ) + negative_prompt_embeds_clip = callback_outputs.pop( + "negative_prompt_embeds_clip", negative_prompt_embeds_clip + ) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing - extract main latents + latents = latents[:, :, :, :, :num_channels_latents] + + # 10. Decode latents to image + if output_type != "latent": + latents = latents.to(self.vae.dtype) + # Reshape and normalize latents + latents = latents.reshape( + batch_size, + num_images_per_prompt, + 1, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + num_channels_latents, + ) + latents = latents.permute(0, 1, 5, 2, 3, 4) # [batch, num_images, channels, 1, height, width] + latents = latents.reshape( + batch_size * num_images_per_prompt, + num_channels_latents, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + ) + + # Normalize and decode through VAE + latents = latents / self.vae.config.scaling_factor + image = self.vae.decode(latents).sample + image = self.image_processor.postprocess(image, output_type=output_type) + else: + image = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return KandinskyImagePipelineOutput(image=image) diff --git a/diffusers/pipelines/kandinsky5/pipeline_kandinsky_i2v.py b/diffusers/pipelines/kandinsky5/pipeline_kandinsky_i2v.py new file mode 100644 index 0000000000000000000000000000000000000000..e82dc737f1a920999ed94f9d8e5656616cc54478 --- /dev/null +++ b/diffusers/pipelines/kandinsky5/pipeline_kandinsky_i2v.py @@ -0,0 +1,1053 @@ +# Copyright 2025 The Kandinsky Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Callable + +import regex as re +import torch +from torch.nn import functional as F +from transformers import CLIPTextModel, CLIPTokenizer, Qwen2_5_VLForConditionalGeneration, Qwen2VLProcessor + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import KandinskyLoraLoaderMixin +from ...models import AutoencoderKLHunyuanVideo +from ...models.transformers import Kandinsky5Transformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler + +# Add imports for offloading and tiling +from ...utils import ( + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import KandinskyPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + + ```python + >>> import torch + >>> from diffusers import Kandinsky5I2VPipeline + >>> from diffusers.utils import export_to_video, load_image + + >>> # Available models: + >>> # kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers + + >>> model_id = "kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers" + >>> pipe = Kandinsky5I2VPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) + >>> pipe = pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + ... ) + >>> prompt = "An astronaut floating in space with Earth in the background, cinematic shot" + >>> negative_prompt = "Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards" + + >>> output = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=512, + ... width=768, + ... num_frames=121, + ... num_inference_steps=50, + ... guidance_scale=5.0, + ... ).frames[0] + + >>> export_to_video(output, "output.mp4", fps=24, quality=9) + ``` +""" + + +def basic_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Clean text using ftfy if available and unescape HTML entities. + """ + if is_ftfy_available(): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Normalize whitespace in text by replacing multiple spaces with single space. + """ + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Apply both basic cleaning and whitespace normalization to prompts. + """ + text = whitespace_clean(basic_clean(text)) + return text + + +class Kandinsky5I2VPipeline(DiffusionPipeline, KandinskyLoraLoaderMixin): + r""" + Pipeline for image-to-video generation using Kandinsky 5.0. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + transformer ([`Kandinsky5Transformer3DModel`]): + Conditional Transformer to denoise the encoded video latents. + vae ([`AutoencoderKLHunyuanVideo`]): + Variational Auto-Encoder Model [hunyuanvideo-community/HunyuanVideo + (vae)](https://huggingface.co/hunyuanvideo-community/HunyuanVideo) to encode and decode videos to and from + latent representations. + text_encoder ([`Qwen2_5_VLForConditionalGeneration`]): + Frozen text-encoder [Qwen2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct). + tokenizer ([`AutoProcessor`]): + Tokenizer for Qwen2.5-VL. + text_encoder_2 ([`CLIPTextModel`]): + Frozen [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 ([`CLIPTokenizer`]): + Tokenizer for CLIP. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded video latents. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = [ + "latents", + "prompt_embeds_qwen", + "prompt_embeds_clip", + "negative_prompt_embeds_qwen", + "negative_prompt_embeds_clip", + ] + + def __init__( + self, + transformer: Kandinsky5Transformer3DModel, + vae: AutoencoderKLHunyuanVideo, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2VLProcessor, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + transformer=transformer, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + scheduler=scheduler, + ) + + self.prompt_template = "\n".join( + [ + "<|im_start|>system\nYou are a promt engineer. Describe the video in detail.", + "Describe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.", + "Describe the location of the video, main characters or objects and their action.", + "Describe the dynamism of the video and presented actions.", + "Name the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or scren content.", + "Describe the visual effects, postprocessing and transitions if they are presented in the video.", + "Pay attention to the order of key actions shown in the scene.<|im_end|>", + "<|im_start|>user\n{}<|im_end|>", + ] + ) + self.prompt_template_encode_start_idx = 129 + + self.vae_scale_factor_temporal = ( + self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4 + ) + self.vae_scale_factor_spatial = self.vae.config.spatial_compression_ratio if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + def _get_scale_factor(self, height: int, width: int) -> tuple: + """ + Calculate the scale factor based on resolution. + + Args: + height (int): Video height + width (int): Video width + + Returns: + tuple: Scale factor as (temporal_scale, height_scale, width_scale) + """ + + def between_480p(x): + return 480 <= x <= 854 + + if between_480p(height) and between_480p(width): + return (1, 2, 2) + else: + return (1, 3.16, 3.16) + + @staticmethod + def fast_sta_nabla(T: int, H: int, W: int, wT: int = 3, wH: int = 3, wW: int = 3, device="cuda") -> torch.Tensor: + """ + Create a sparse temporal attention (STA) mask for efficient video generation. + + This method generates a mask that limits attention to nearby frames and spatial positions, reducing + computational complexity for video generation. + + Args: + T (int): Number of temporal frames + H (int): Height in latent space + W (int): Width in latent space + wT (int): Temporal attention window size + wH (int): Height attention window size + wW (int): Width attention window size + device (str): Device to create tensor on + + Returns: + torch.Tensor: Sparse attention mask of shape (T*H*W, T*H*W) + """ + l = torch.Tensor([T, H, W]).amax() + r = torch.arange(0, l, 1, dtype=torch.int16, device=device) + mat = (r.unsqueeze(1) - r.unsqueeze(0)).abs() + sta_t, sta_h, sta_w = ( + mat[:T, :T].flatten(), + mat[:H, :H].flatten(), + mat[:W, :W].flatten(), + ) + sta_t = sta_t <= wT // 2 + sta_h = sta_h <= wH // 2 + sta_w = sta_w <= wW // 2 + sta_hw = (sta_h.unsqueeze(1) * sta_w.unsqueeze(0)).reshape(H, H, W, W).transpose(1, 2).flatten() + sta = (sta_t.unsqueeze(1) * sta_hw.unsqueeze(0)).reshape(T, T, H * W, H * W).transpose(1, 2) + return sta.reshape(T * H * W, T * H * W) + + def get_sparse_params(self, sample, device): + """ + Generate sparse attention parameters for the transformer based on sample dimensions. + + This method computes the sparse attention configuration needed for efficient video processing in the + transformer model. + + Args: + sample (torch.Tensor): Input sample tensor + device (torch.device): Device to place tensors on + + Returns: + Dict: Dictionary containing sparse attention parameters + """ + assert self.transformer.config.patch_size[0] == 1 + B, T, H, W, _ = sample.shape + T, H, W = ( + T // self.transformer.config.patch_size[0], + H // self.transformer.config.patch_size[1], + W // self.transformer.config.patch_size[2], + ) + if self.transformer.config.attention_type == "nabla": + sta_mask = self.fast_sta_nabla( + T, + H // 8, + W // 8, + self.transformer.config.attention_wT, + self.transformer.config.attention_wH, + self.transformer.config.attention_wW, + device=device, + ) + + sparse_params = { + "sta_mask": sta_mask.unsqueeze_(0).unsqueeze_(0), + "attention_type": self.transformer.config.attention_type, + "to_fractal": True, + "P": self.transformer.config.attention_P, + "wT": self.transformer.config.attention_wT, + "wW": self.transformer.config.attention_wW, + "wH": self.transformer.config.attention_wH, + "add_sta": self.transformer.config.attention_add_sta, + "visual_shape": (T, H, W), + "method": self.transformer.config.attention_method, + } + else: + sparse_params = None + + return sparse_params + + def _encode_prompt_qwen( + self, + prompt: str | list[str], + device: torch.device | None = None, + max_sequence_length: int = 256, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using Qwen2.5-VL text encoder. + + This method processes the input prompt through the Qwen2.5-VL model to generate text embeddings suitable for + video generation. + + Args: + prompt (str | list[str]): Input prompt or list of prompts + device (torch.device): Device to run encoding on + max_sequence_length (int): Maximum sequence length for tokenization + dtype (torch.dtype): Data type for embeddings + + Returns: + tuple[torch.Tensor, torch.Tensor]: Text embeddings and cumulative sequence lengths + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + full_texts = [self.prompt_template.format(p) for p in prompt] + max_allowed_len = self.prompt_template_encode_start_idx + max_sequence_length + + untruncated_ids = self.tokenizer( + text=full_texts, + images=None, + videos=None, + return_tensors="pt", + padding="longest", + )["input_ids"] + + if untruncated_ids.shape[-1] > max_allowed_len: + for i, text in enumerate(full_texts): + tokens = untruncated_ids[i][self.prompt_template_encode_start_idx : -2] + removed_text = self.tokenizer.decode(tokens[max_sequence_length - 2 :]) + if len(removed_text) > 0: + full_texts[i] = text[: -len(removed_text)] + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + inputs = self.tokenizer( + text=full_texts, + images=None, + videos=None, + max_length=max_allowed_len, + truncation=True, + return_tensors="pt", + padding=True, + ).to(device) + + embeds = self.text_encoder( + input_ids=inputs["input_ids"], + return_dict=True, + output_hidden_states=True, + )["hidden_states"][-1][:, self.prompt_template_encode_start_idx :] + + attention_mask = inputs["attention_mask"][:, self.prompt_template_encode_start_idx :] + cu_seqlens = torch.cumsum(attention_mask.sum(1), dim=0) + cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0).to(dtype=torch.int32) + + return embeds.to(dtype), cu_seqlens + + def _encode_prompt_clip( + self, + prompt: str | list[str], + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using CLIP text encoder. + + This method processes the input prompt through the CLIP model to generate pooled embeddings that capture + semantic information. + + Args: + prompt (str | list[str]): Input prompt or list of prompts + device (torch.device): Device to run encoding on + dtype (torch.dtype): Data type for embeddings + + Returns: + torch.Tensor: Pooled text embeddings from CLIP + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + inputs = self.tokenizer_2( + prompt, + max_length=77, + truncation=True, + add_special_tokens=True, + padding="max_length", + return_tensors="pt", + ).to(device) + + pooled_embed = self.text_encoder_2(**inputs)["pooler_output"] + + return pooled_embed.to(dtype) + + @staticmethod + def adaptive_mean_std_normalization(source, reference): + source_mean = source.mean(dim=(1, 2, 3, 4), keepdim=True) + source_std = source.std(dim=(1, 2, 3, 4), keepdim=True) + # magic constants - limit changes in latents + clump_mean_low = 0.05 + clump_mean_high = 0.1 + clump_std_low = 0.1 + clump_std_high = 0.25 + + reference_mean = torch.clamp(reference.mean(), source_mean - clump_mean_low, source_mean + clump_mean_high) + reference_std = torch.clamp(reference.std(), source_std - clump_std_low, source_std + clump_std_high) + + # normalization + normalized = (source - source_mean) / source_std + normalized = normalized * reference_std + reference_mean + + return normalized + + def normalize_first_frame(self, latents, reference_frames=5, clump_values=False): + latents_copy = latents.clone() + samples = latents_copy + + if samples.shape[1] <= 1: + return (latents, "Only one frame, no normalization needed") + + nFr = 4 + first_frames = samples.clone()[:, :nFr] + reference_frames_data = samples[:, nFr : nFr + min(reference_frames, samples.shape[1] - 1)] + + normalized_first = self.adaptive_mean_std_normalization(first_frames, reference_frames_data) + if clump_values: + min_val = reference_frames_data.min() + max_val = reference_frames_data.max() + normalized_first = torch.clamp(normalized_first, min_val, max_val) + + samples[:, :nFr] = normalized_first + + return samples + + def encode_prompt( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes a single prompt (positive or negative) into text encoder hidden states. + + This method combines embeddings from both Qwen2.5-VL and CLIP text encoders to create comprehensive text + representations for video generation. + + Args: + prompt (`str` or `list[str]`): + Prompt to be encoded. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos to generate per prompt. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length for text encoding. + device (`torch.device`, *optional*): + Torch device. + dtype (`torch.dtype`, *optional*): + Torch dtype. + + Returns: + tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + - Qwen text embeddings of shape (batch_size * num_videos_per_prompt, sequence_length, embedding_dim) + - CLIP pooled embeddings of shape (batch_size * num_videos_per_prompt, clip_embedding_dim) + - Cumulative sequence lengths (`cu_seqlens`) for Qwen embeddings of shape (batch_size * + num_videos_per_prompt + 1,) + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + if not isinstance(prompt, list): + prompt = [prompt] + + batch_size = len(prompt) + + prompt = [prompt_clean(p) for p in prompt] + + # Encode with Qwen2.5-VL + prompt_embeds_qwen, prompt_cu_seqlens = self._encode_prompt_qwen( + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + dtype=dtype, + ) + # prompt_embeds_qwen shape: [batch_size, seq_len, embed_dim] + + # Encode with CLIP + prompt_embeds_clip = self._encode_prompt_clip( + prompt=prompt, + device=device, + dtype=dtype, + ) + # prompt_embeds_clip shape: [batch_size, clip_embed_dim] + + # Repeat embeddings for num_videos_per_prompt + # Qwen embeddings: repeat sequence for each video, then reshape + prompt_embeds_qwen = prompt_embeds_qwen.repeat( + 1, num_videos_per_prompt, 1 + ) # [batch_size, seq_len * num_videos_per_prompt, embed_dim] + # Reshape to [batch_size * num_videos_per_prompt, seq_len, embed_dim] + prompt_embeds_qwen = prompt_embeds_qwen.view( + batch_size * num_videos_per_prompt, -1, prompt_embeds_qwen.shape[-1] + ) + + # CLIP embeddings: repeat for each video + prompt_embeds_clip = prompt_embeds_clip.repeat( + 1, num_videos_per_prompt, 1 + ) # [batch_size, num_videos_per_prompt, clip_embed_dim] + # Reshape to [batch_size * num_videos_per_prompt, clip_embed_dim] + prompt_embeds_clip = prompt_embeds_clip.view(batch_size * num_videos_per_prompt, -1) + + # Repeat cumulative sequence lengths for num_videos_per_prompt + # Original differences (lengths) for each prompt in the batch + original_lengths = prompt_cu_seqlens.diff() # [len1, len2, ...] + # Repeat the lengths for num_videos_per_prompt + repeated_lengths = original_lengths.repeat_interleave( + num_videos_per_prompt + ) # [len1, len1, ..., len2, len2, ...] + # Reconstruct the cumulative lengths + repeated_cu_seqlens = torch.cat( + [torch.tensor([0], device=device, dtype=torch.int32), repeated_lengths.cumsum(0)] + ) + + return prompt_embeds_qwen, prompt_embeds_clip, repeated_cu_seqlens + + def check_inputs( + self, + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds_qwen=None, + prompt_embeds_clip=None, + negative_prompt_embeds_qwen=None, + negative_prompt_embeds_clip=None, + prompt_cu_seqlens=None, + negative_prompt_cu_seqlens=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + """ + Validate input parameters for the pipeline. + + Args: + prompt: Input prompt + negative_prompt: Negative prompt for guidance + image: Input image for conditioning + height: Video height + width: Video width + prompt_embeds_qwen: Pre-computed Qwen prompt embeddings + prompt_embeds_clip: Pre-computed CLIP prompt embeddings + negative_prompt_embeds_qwen: Pre-computed Qwen negative prompt embeddings + negative_prompt_embeds_clip: Pre-computed CLIP negative prompt embeddings + prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen positive prompt + negative_prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen negative prompt + callback_on_step_end_tensor_inputs: Callback tensor inputs + + Raises: + ValueError: If inputs are invalid + """ + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError("max_sequence_length must be less than 1024") + + if image is None: + raise ValueError("`image` must be provided for image-to-video generation") + + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Check for consistency within positive prompt embeddings and sequence lengths + if prompt_embeds_qwen is not None or prompt_embeds_clip is not None or prompt_cu_seqlens is not None: + if prompt_embeds_qwen is None or prompt_embeds_clip is None or prompt_cu_seqlens is None: + raise ValueError( + "If any of `prompt_embeds_qwen`, `prompt_embeds_clip`, or `prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check for consistency within negative prompt embeddings and sequence lengths + if ( + negative_prompt_embeds_qwen is not None + or negative_prompt_embeds_clip is not None + or negative_prompt_cu_seqlens is not None + ): + if ( + negative_prompt_embeds_qwen is None + or negative_prompt_embeds_clip is None + or negative_prompt_cu_seqlens is None + ): + raise ValueError( + "If any of `negative_prompt_embeds_qwen`, `negative_prompt_embeds_clip`, or `negative_prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check if prompt or embeddings are provided (either prompt or all required embedding components for positive) + if prompt is None and prompt_embeds_qwen is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_qwen` (and corresponding `prompt_embeds_clip` and `prompt_cu_seqlens`). Cannot leave all undefined." + ) + + # Validate types for prompt and negative_prompt if provided + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + if negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + def prepare_latents( + self, + image: PipelineImageInput, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + Prepare initial latent variables for image-to-video generation. + + This method creates random noise latents for all frames except the first frame, which is replaced with the + encoded input image. + + Args: + image (PipelineImageInput): Input image to condition the generation on + batch_size (int): Number of videos to generate + num_channels_latents (int): Number of channels in latent space + height (int): Height of generated video + width (int): Width of generated video + num_frames (int): Number of frames in video + dtype (torch.dtype): Data type for latents + device (torch.device): Device to create latents on + generator (torch.Generator): Random number generator + latents (torch.Tensor): Pre-existing latents to use + + Returns: + torch.Tensor: Prepared latent tensor with first frame as encoded image + """ + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + num_channels_latents, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Generate random noise for all frames + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # Encode the input image to use as first frame + # Preprocess image + image_tensor = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=dtype) + + # Encode image to latents using VAE + with torch.no_grad(): + # Convert image to video format [batch, channels, 1, height, width] + image_video = image_tensor.unsqueeze(2) # Add temporal dimension + image_latents = self.vae.encode(image_video).latent_dist.sample(generator=generator) + + # Normalize latents if needed + if hasattr(self.vae.config, "scaling_factor"): + image_latents = image_latents * self.vae.config.scaling_factor + + # Reshape to match latent dimensions [batch, frames, height, width, channels] + image_latents = image_latents.permute(0, 2, 3, 4, 1) # [batch, 1, H, W, C] + image_latents = image_latents.to(device=latents.device, dtype=latents.dtype) + + # Replace first frame with encoded image + latents[:, 0:1] = image_latents + + if self.transformer.visual_cond: + # For visual conditioning, concatenate with zeros and mask + visual_cond = torch.zeros_like(latents) + visual_cond_mask = torch.zeros( + [ + batch_size, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + 1, + ], + dtype=latents.dtype, + device=latents.device, + ) + + visual_cond_mask[:, 0:1] = 1 + visual_cond[:, 0:1] = image_latents + + latents = torch.cat([latents, visual_cond, visual_cond_mask], dim=-1) + + return latents + + @property + def guidance_scale(self): + """Get the current guidance scale value.""" + return self._guidance_scale + + @property + def num_timesteps(self): + """Get the number of denoising timesteps.""" + return self._num_timesteps + + @property + def interrupt(self): + """Check if generation has been interrupted.""" + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds_qwen: torch.Tensor | None = None, + prompt_embeds_clip: torch.Tensor | None = None, + negative_prompt_embeds_qwen: torch.Tensor | None = None, + negative_prompt_embeds_clip: torch.Tensor | None = None, + prompt_cu_seqlens: torch.Tensor | None = None, + negative_prompt_cu_seqlens: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int, None], PipelineCallback | MultiPipelineCallbacks] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for image-to-video generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to avoid during video generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`, defaults to `512`): + The height in pixels of the generated video. + width (`int`, defaults to `768`): + The width in pixels of the generated video. + num_frames (`int`, defaults to `121`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in classifier-free guidance. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A torch generator to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents. + prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated Qwen text embeddings. + prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated CLIP text embeddings. + negative_prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated Qwen negative text embeddings. + negative_prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated CLIP negative text embeddings. + prompt_cu_seqlens (`torch.Tensor`, *optional*): + Pre-generated cumulative sequence lengths for Qwen positive prompt. + negative_prompt_cu_seqlens (`torch.Tensor`, *optional*): + Pre-generated cumulative sequence lengths for Qwen negative prompt. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated video. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`KandinskyPipelineOutput`]. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function that is called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length for text encoding. + + Examples: + + Returns: + [`~KandinskyPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`KandinskyPipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated videos. + """ + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + negative_prompt=negative_prompt, + image=image, + height=height, + width=width, + prompt_embeds_qwen=prompt_embeds_qwen, + prompt_embeds_clip=prompt_embeds_clip, + negative_prompt_embeds_qwen=negative_prompt_embeds_qwen, + negative_prompt_embeds_clip=negative_prompt_embeds_clip, + prompt_cu_seqlens=prompt_cu_seqlens, + negative_prompt_cu_seqlens=negative_prompt_cu_seqlens, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._interrupt = False + + device = self._execution_device + dtype = self.transformer.dtype + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + prompt = [prompt] + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds_qwen.shape[0] + + # 3. Encode input prompt + if prompt_embeds_qwen is None: + prompt_embeds_qwen, prompt_embeds_clip, prompt_cu_seqlens = self.encode_prompt( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if self.guidance_scale > 1.0: + if negative_prompt is None: + negative_prompt = "Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards" + + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * len(prompt) if prompt is not None else [negative_prompt] + elif len(negative_prompt) != len(prompt): + raise ValueError( + f"`negative_prompt` must have same length as `prompt`. Got {len(negative_prompt)} vs {len(prompt)}." + ) + + if negative_prompt_embeds_qwen is None: + negative_prompt_embeds_qwen, negative_prompt_embeds_clip, negative_prompt_cu_seqlens = ( + self.encode_prompt( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables with image conditioning + num_channels_latents = self.transformer.config.in_visual_dim + latents = self.prepare_latents( + image=image, + batch_size=batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + num_frames=num_frames, + dtype=dtype, + device=device, + generator=generator, + latents=latents, + ) + + # 6. Prepare rope positions for positional encoding + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + visual_rope_pos = [ + torch.arange(num_latent_frames, device=device), + torch.arange(height // self.vae_scale_factor_spatial // 2, device=device), + torch.arange(width // self.vae_scale_factor_spatial // 2, device=device), + ] + + text_rope_pos = torch.arange(prompt_cu_seqlens.diff().max().item(), device=device) + + negative_text_rope_pos = ( + torch.arange(negative_prompt_cu_seqlens.diff().max().item(), device=device) + if negative_prompt_cu_seqlens is not None + else None + ) + + # 7. Calculate dynamic scale factor based on resolution + scale_factor = self._get_scale_factor(height, width) + + # 8. Sparse Params for efficient attention + sparse_params = self.get_sparse_params(latents, device) + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + timestep = t.unsqueeze(0).repeat(batch_size * num_videos_per_prompt) + + # Predict noise residual + pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=prompt_embeds_qwen.to(dtype), + pooled_projections=prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + if self.guidance_scale > 1.0 and negative_prompt_embeds_qwen is not None: + uncond_pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=negative_prompt_embeds_qwen.to(dtype), + pooled_projections=negative_prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=negative_text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + pred_velocity = uncond_pred_velocity + guidance_scale * (pred_velocity - uncond_pred_velocity) + + latents[:, 1:, :, :, :num_channels_latents] = self.scheduler.step( + pred_velocity[:, 1:], t, latents[:, 1:, :, :, :num_channels_latents], return_dict=False + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds_qwen = callback_outputs.pop("prompt_embeds_qwen", prompt_embeds_qwen) + prompt_embeds_clip = callback_outputs.pop("prompt_embeds_clip", prompt_embeds_clip) + negative_prompt_embeds_qwen = callback_outputs.pop( + "negative_prompt_embeds_qwen", negative_prompt_embeds_qwen + ) + negative_prompt_embeds_clip = callback_outputs.pop( + "negative_prompt_embeds_clip", negative_prompt_embeds_clip + ) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing - extract main latents + latents = latents[:, :, :, :, :num_channels_latents] + + # 10. fix mesh artifacts + latents = self.normalize_first_frame(latents) + + # 11. Decode latents to video + if output_type != "latent": + latents = latents.to(self.vae.dtype) + # Reshape and normalize latents + video = latents.reshape( + batch_size, + num_videos_per_prompt, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + num_channels_latents, + ) + video = video.permute(0, 1, 5, 2, 3, 4) # [batch, num_videos, channels, frames, height, width] + video = video.reshape( + batch_size * num_videos_per_prompt, + num_channels_latents, + (num_frames - 1) // self.vae_scale_factor_temporal + 1, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + ) + + # Normalize and decode through VAE + video = video / self.vae.config.scaling_factor + video = self.vae.decode(video).sample + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return KandinskyPipelineOutput(frames=video) diff --git a/diffusers/pipelines/kandinsky5/pipeline_kandinsky_t2i.py b/diffusers/pipelines/kandinsky5/pipeline_kandinsky_t2i.py new file mode 100644 index 0000000000000000000000000000000000000000..2a58d4bed33a42fe71c77f134bd28d73155a96ed --- /dev/null +++ b/diffusers/pipelines/kandinsky5/pipeline_kandinsky_t2i.py @@ -0,0 +1,816 @@ +# Copyright 2025 The Kandinsky Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Callable + +import numpy as np +import regex as re +import torch +from torch.nn import functional as F +from transformers import CLIPTextModel, CLIPTokenizer, Qwen2_5_VLForConditionalGeneration, Qwen2VLProcessor + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import VaeImageProcessor +from ...loaders import KandinskyLoraLoaderMixin +from ...models import AutoencoderKL +from ...models.transformers import Kandinsky5Transformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler + +# Add imports for offloading and tiling +from ...utils import ( + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import KandinskyImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + + ```python + >>> import torch + >>> from diffusers import Kandinsky5T2IPipeline + + >>> # Available models: + >>> # kandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers + >>> # kandinskylab/Kandinsky-5.0-T2I-Lite-pretrain-Diffusers + + >>> model_id = "kandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers" + >>> pipe = Kandinsky5T2IPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen." + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt="", + ... height=1024, + ... width=1024, + ... num_inference_steps=50, + ... guidance_scale=3.5, + ... ).frames[0] + ``` +""" + + +def basic_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Clean text using ftfy if available and unescape HTML entities. + """ + if is_ftfy_available(): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Normalize whitespace in text by replacing multiple spaces with single space. + """ + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + """ + Copied from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/wan/pipeline_wan.py + + Apply both basic cleaning and whitespace normalization to prompts. + """ + text = whitespace_clean(basic_clean(text)) + return text + + +class Kandinsky5T2IPipeline(DiffusionPipeline, KandinskyLoraLoaderMixin): + r""" + Pipeline for text-to-image generation using Kandinsky 5.0. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + transformer ([`Kandinsky5Transformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder Model [black-forest-labs/FLUX.1-dev + (vae)](https://huggingface.co/black-forest-labs/FLUX.1-dev) to encode and decode videos to and from latent + representations. + text_encoder ([`Qwen2_5_VLForConditionalGeneration`]): + Frozen text-encoder [Qwen2.5-VL](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct). + tokenizer ([`AutoProcessor`]): + Tokenizer for Qwen2.5-VL. + text_encoder_2 ([`CLIPTextModel`]): + Frozen [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer_2 ([`CLIPTokenizer`]): + Tokenizer for CLIP. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _callback_tensor_inputs = [ + "latents", + "prompt_embeds_qwen", + "prompt_embeds_clip", + "negative_prompt_embeds_qwen", + "negative_prompt_embeds_clip", + ] + + def __init__( + self, + transformer: Kandinsky5Transformer3DModel, + vae: AutoencoderKL, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2VLProcessor, + text_encoder_2: CLIPTextModel, + tokenizer_2: CLIPTokenizer, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + transformer=transformer, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + scheduler=scheduler, + ) + + self.prompt_template = "<|im_start|>system\nYou are a promt engineer. Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>" + self.prompt_template_encode_start_idx = 41 + + self.vae_scale_factor_spatial = 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.resolutions = [(1024, 1024), (640, 1408), (1408, 640), (768, 1280), (1280, 768), (896, 1152), (1152, 896)] + + def _encode_prompt_qwen( + self, + prompt: list[str], + device: torch.device | None = None, + max_sequence_length: int = 512, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using Qwen2.5-VL text encoder. + + This method processes the input prompt through the Qwen2.5-VL model to generate text embeddings suitable for + image generation. + + Args: + prompt list[str]: Input list of prompts + device (torch.device): Device to run encoding on + max_sequence_length (int): Maximum sequence length for tokenization + dtype (torch.dtype): Data type for embeddings + + Returns: + tuple[torch.Tensor, torch.Tensor]: Text embeddings and cumulative sequence lengths + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + full_texts = [self.prompt_template.format(p) for p in prompt] + max_allowed_len = self.prompt_template_encode_start_idx + max_sequence_length + + untruncated_ids = self.tokenizer( + text=full_texts, + images=None, + videos=None, + return_tensors="pt", + padding="longest", + )["input_ids"] + + if untruncated_ids.shape[-1] > max_allowed_len: + for i, text in enumerate(full_texts): + tokens = untruncated_ids[i][self.prompt_template_encode_start_idx : -2] + removed_text = self.tokenizer.decode(tokens[max_sequence_length - 2 :]) + if len(removed_text) > 0: + full_texts[i] = text[: -len(removed_text)] + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + inputs = self.tokenizer( + text=full_texts, + images=None, + videos=None, + max_length=max_allowed_len, + truncation=True, + return_tensors="pt", + padding=True, + ).to(device) + + embeds = self.text_encoder( + input_ids=inputs["input_ids"], + return_dict=True, + output_hidden_states=True, + )["hidden_states"][-1][:, self.prompt_template_encode_start_idx :] + attention_mask = inputs["attention_mask"][:, self.prompt_template_encode_start_idx :] + cu_seqlens = torch.cumsum(attention_mask.sum(1), dim=0) + cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0).to(dtype=torch.int32) + + return embeds.to(dtype), cu_seqlens + + def _encode_prompt_clip( + self, + prompt: str | list[str], + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + """ + Encode prompt using CLIP text encoder. + + This method processes the input prompt through the CLIP model to generate pooled embeddings that capture + semantic information. + + Args: + prompt (str | list[str]): Input prompt or list of prompts + device (torch.device): Device to run encoding on + dtype (torch.dtype): Data type for embeddings + + Returns: + torch.Tensor: Pooled text embeddings from CLIP + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder_2.dtype + + inputs = self.tokenizer_2( + prompt, + max_length=77, + truncation=True, + add_special_tokens=True, + padding="max_length", + return_tensors="pt", + ).to(device) + + pooled_embed = self.text_encoder_2(**inputs)["pooler_output"] + + return pooled_embed.to(dtype) + + def encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes a single prompt (positive or negative) into text encoder hidden states. + + This method combines embeddings from both Qwen2.5-VL and CLIP text encoders to create comprehensive text + representations for image generation. + + Args: + prompt (`str` or `list[str]`): + Prompt to be encoded. + num_images_per_prompt (`int`, *optional*, defaults to 1): + Number of images to generate per prompt. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length for text encoding. Must be less than 1024 + device (`torch.device`, *optional*): + Torch device. + dtype (`torch.dtype`, *optional*): + Torch dtype. + + Returns: + tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + - Qwen text embeddings of shape (batch_size * num_images_per_prompt, sequence_length, embedding_dim) + - CLIP pooled embeddings of shape (batch_size * num_images_per_prompt, clip_embedding_dim) + - Cumulative sequence lengths (`cu_seqlens`) for Qwen embeddings of shape (batch_size * + num_images_per_prompt + 1,) + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + if not isinstance(prompt, list): + prompt = [prompt] + + batch_size = len(prompt) + + prompt = [prompt_clean(p) for p in prompt] + + # Encode with Qwen2.5-VL + prompt_embeds_qwen, prompt_cu_seqlens = self._encode_prompt_qwen( + prompt=prompt, + device=device, + max_sequence_length=max_sequence_length, + dtype=dtype, + ) + # prompt_embeds_qwen shape: [batch_size, seq_len, embed_dim] + + # Encode with CLIP + prompt_embeds_clip = self._encode_prompt_clip( + prompt=prompt, + device=device, + dtype=dtype, + ) + # prompt_embeds_clip shape: [batch_size, clip_embed_dim] + + # Repeat embeddings for num_images_per_prompt + # Qwen embeddings: repeat sequence for each image, then reshape + prompt_embeds_qwen = prompt_embeds_qwen.repeat( + 1, num_images_per_prompt, 1 + ) # [batch_size, seq_len * num_images_per_prompt, embed_dim] + # Reshape to [batch_size * num_images_per_prompt, seq_len, embed_dim] + prompt_embeds_qwen = prompt_embeds_qwen.view( + batch_size * num_images_per_prompt, -1, prompt_embeds_qwen.shape[-1] + ) + + # CLIP embeddings: repeat for each image + prompt_embeds_clip = prompt_embeds_clip.repeat( + 1, num_images_per_prompt, 1 + ) # [batch_size, num_images_per_prompt, clip_embed_dim] + # Reshape to [batch_size * num_images_per_prompt, clip_embed_dim] + prompt_embeds_clip = prompt_embeds_clip.view(batch_size * num_images_per_prompt, -1) + + # Repeat cumulative sequence lengths for num_images_per_prompt + # Original differences (lengths) for each prompt in the batch + original_lengths = prompt_cu_seqlens.diff() # [len1, len2, ...] + # Repeat the lengths for num_images_per_prompt + repeated_lengths = original_lengths.repeat_interleave( + num_images_per_prompt + ) # [len1, len1, ..., len2, len2, ...] + # Reconstruct the cumulative lengths + repeated_cu_seqlens = torch.cat( + [torch.tensor([0], device=device, dtype=torch.int32), repeated_lengths.cumsum(0)] + ) + + return prompt_embeds_qwen, prompt_embeds_clip, repeated_cu_seqlens + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds_qwen=None, + prompt_embeds_clip=None, + negative_prompt_embeds_qwen=None, + negative_prompt_embeds_clip=None, + prompt_cu_seqlens=None, + negative_prompt_cu_seqlens=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + """ + Validate input parameters for the pipeline. + + Args: + prompt: Input prompt + negative_prompt: Negative prompt for guidance + height: Image height + width: Image width + prompt_embeds_qwen: Pre-computed Qwen prompt embeddings + prompt_embeds_clip: Pre-computed CLIP prompt embeddings + negative_prompt_embeds_qwen: Pre-computed Qwen negative prompt embeddings + negative_prompt_embeds_clip: Pre-computed CLIP negative prompt embeddings + prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen positive prompt + negative_prompt_cu_seqlens: Pre-computed cumulative sequence lengths for Qwen negative prompt + callback_on_step_end_tensor_inputs: Callback tensor inputs + + Raises: + ValueError: If inputs are invalid + """ + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError("max_sequence_length must be less than 1024") + + if (width, height) not in self.resolutions: + resolutions_str = ",".join([f"({w},{h})" for w, h in self.resolutions]) + logger.warning( + f"`height` and `width` have to be one of {resolutions_str}, but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Check for consistency within positive prompt embeddings and sequence lengths + if prompt_embeds_qwen is not None or prompt_embeds_clip is not None or prompt_cu_seqlens is not None: + if prompt_embeds_qwen is None or prompt_embeds_clip is None or prompt_cu_seqlens is None: + raise ValueError( + "If any of `prompt_embeds_qwen`, `prompt_embeds_clip`, or `prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check for consistency within negative prompt embeddings and sequence lengths + if ( + negative_prompt_embeds_qwen is not None + or negative_prompt_embeds_clip is not None + or negative_prompt_cu_seqlens is not None + ): + if ( + negative_prompt_embeds_qwen is None + or negative_prompt_embeds_clip is None + or negative_prompt_cu_seqlens is None + ): + raise ValueError( + "If any of `negative_prompt_embeds_qwen`, `negative_prompt_embeds_clip`, or `negative_prompt_cu_seqlens` is provided, " + "all three must be provided." + ) + + # Check if prompt or embeddings are provided (either prompt or all required embedding components for positive) + if prompt is None and prompt_embeds_qwen is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_qwen` (and corresponding `prompt_embeds_clip` and `prompt_cu_seqlens`). Cannot leave all undefined." + ) + + # Validate types for prompt and negative_prompt if provided + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + if negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 1024, + width: int = 1024, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + Prepare initial latent variables for text-to-image generation. + + This method creates random noise latents + + Args: + batch_size (int): Number of images to generate + num_channels_latents (int): Number of channels in latent space + height (int): Height of generated image + width (int): Width of generated image + dtype (torch.dtype): Data type for latents + device (torch.device): Device to create latents on + generator (torch.Generator): Random number generator + latents (torch.Tensor): Pre-existing latents to use + + Returns: + torch.Tensor: Prepared latent tensor + """ + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + 1, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + num_channels_latents, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Generate random noise + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + """Get the current guidance scale value.""" + return self._guidance_scale + + @property + def num_timesteps(self): + """Get the number of denoising timesteps.""" + return self._num_timesteps + + @property + def interrupt(self): + """Check if generation has been interrupted.""" + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 1024, + width: int = 1024, + num_inference_steps: int = 50, + guidance_scale: float = 3.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds_qwen: torch.Tensor | None = None, + prompt_embeds_clip: torch.Tensor | None = None, + negative_prompt_embeds_qwen: torch.Tensor | None = None, + negative_prompt_embeds_clip: torch.Tensor | None = None, + prompt_cu_seqlens: torch.Tensor | None = None, + negative_prompt_cu_seqlens: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int, None], PipelineCallback | MultiPipelineCallbacks] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for text-to-image generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to avoid during image generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`, defaults to `1024`): + The height in pixels of the generated image. + width (`int`, defaults to `1024`): + The width in pixels of the generated image. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in classifier-free guidance. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A torch generator to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents. + prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated Qwen text embeddings. + prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated CLIP text embeddings. + negative_prompt_embeds_qwen (`torch.Tensor`, *optional*): + Pre-generated Qwen negative text embeddings. + negative_prompt_embeds_clip (`torch.Tensor`, *optional*): + Pre-generated CLIP negative text embeddings. + prompt_cu_seqlens (`torch.Tensor`, *optional*): + Pre-generated cumulative sequence lengths for Qwen positive prompt. + negative_prompt_cu_seqlens (`torch.Tensor`, *optional*): + Pre-generated cumulative sequence lengths for Qwen negative prompt. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`KandinskyImagePipelineOutput`]. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function that is called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length for text encoding. + + Examples: + + Returns: + [`~KandinskyImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`KandinskyImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + self.check_inputs( + prompt=prompt, + negative_prompt=negative_prompt, + height=height, + width=width, + prompt_embeds_qwen=prompt_embeds_qwen, + prompt_embeds_clip=prompt_embeds_clip, + negative_prompt_embeds_qwen=negative_prompt_embeds_qwen, + negative_prompt_embeds_clip=negative_prompt_embeds_clip, + prompt_cu_seqlens=prompt_cu_seqlens, + negative_prompt_cu_seqlens=negative_prompt_cu_seqlens, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + if (width, height) not in self.resolutions: + width, height = self.resolutions[ + np.argmin([abs((i[0] / i[1]) - (width / height)) for i in self.resolutions]) + ] + + self._guidance_scale = guidance_scale + self._interrupt = False + + device = self._execution_device + dtype = self.transformer.dtype + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + prompt = [prompt] + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds_qwen.shape[0] + + # 3. Encode input prompt + if prompt_embeds_qwen is None: + prompt_embeds_qwen, prompt_embeds_clip, prompt_cu_seqlens = self.encode_prompt( + prompt=prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if self.guidance_scale > 1.0: + if negative_prompt is None: + negative_prompt = "" + + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * len(prompt) if prompt is not None else [negative_prompt] + elif len(negative_prompt) != len(prompt): + raise ValueError( + f"`negative_prompt` must have same length as `prompt`. Got {len(negative_prompt)} vs {len(prompt)}." + ) + + if negative_prompt_embeds_qwen is None: + negative_prompt_embeds_qwen, negative_prompt_embeds_clip, negative_prompt_cu_seqlens = ( + self.encode_prompt( + prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_visual_dim + latents = self.prepare_latents( + batch_size=batch_size * num_images_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + dtype=dtype, + device=device, + generator=generator, + latents=latents, + ) + + # 6. Prepare rope positions for positional encoding + visual_rope_pos = [ + torch.arange(1, device=device), + torch.arange(height // self.vae_scale_factor_spatial // 2, device=device), + torch.arange(width // self.vae_scale_factor_spatial // 2, device=device), + ] + + text_rope_pos = torch.arange(prompt_cu_seqlens.diff().max().item(), device=device) + + negative_text_rope_pos = ( + torch.arange(negative_prompt_cu_seqlens.diff().max().item(), device=device) + if negative_prompt_cu_seqlens is not None + else None + ) + + # 7. Calculate dynamic scale factor based on resolution + scale_factor = [1.0, 1.0, 1.0] + + # 8. Sparse Params for efficient attention + sparse_params = None + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + timestep = t.unsqueeze(0).repeat(batch_size * num_images_per_prompt) + + # Predict noise residual + pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=prompt_embeds_qwen.to(dtype), + pooled_projections=prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + if self.guidance_scale > 1.0 and negative_prompt_embeds_qwen is not None: + uncond_pred_velocity = self.transformer( + hidden_states=latents.to(dtype), + encoder_hidden_states=negative_prompt_embeds_qwen.to(dtype), + pooled_projections=negative_prompt_embeds_clip.to(dtype), + timestep=timestep.to(dtype), + visual_rope_pos=visual_rope_pos, + text_rope_pos=negative_text_rope_pos, + scale_factor=scale_factor, + sparse_params=sparse_params, + return_dict=True, + ).sample + + pred_velocity = uncond_pred_velocity + guidance_scale * (pred_velocity - uncond_pred_velocity) + + latents = self.scheduler.step(pred_velocity[:, :], t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds_qwen = callback_outputs.pop("prompt_embeds_qwen", prompt_embeds_qwen) + prompt_embeds_clip = callback_outputs.pop("prompt_embeds_clip", prompt_embeds_clip) + negative_prompt_embeds_qwen = callback_outputs.pop( + "negative_prompt_embeds_qwen", negative_prompt_embeds_qwen + ) + negative_prompt_embeds_clip = callback_outputs.pop( + "negative_prompt_embeds_clip", negative_prompt_embeds_clip + ) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing - extract main latents + latents = latents[:, :, :, :, :num_channels_latents] + + # 10. Decode latents to image + if output_type != "latent": + latents = latents.to(self.vae.dtype) + # Reshape and normalize latents + latents = latents.reshape( + batch_size, + num_images_per_prompt, + 1, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + num_channels_latents, + ) + latents = latents.permute(0, 1, 5, 2, 3, 4) # [batch, num_images, channels, 1, height, width] + latents = latents.reshape( + batch_size * num_images_per_prompt, + num_channels_latents, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + ) + + # Normalize and decode through VAE + latents = latents / self.vae.config.scaling_factor + image = self.vae.decode(latents).sample + image = self.image_processor.postprocess(image, output_type=output_type) + else: + image = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return KandinskyImagePipelineOutput(image=image) diff --git a/diffusers/pipelines/kandinsky5/pipeline_output.py b/diffusers/pipelines/kandinsky5/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..9b41d2f14424f29c90cea1d064fe144d5513d294 --- /dev/null +++ b/diffusers/pipelines/kandinsky5/pipeline_output.py @@ -0,0 +1,35 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class KandinskyPipelineOutput(BaseOutput): + r""" + Output class for kandinsky video pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor + + +@dataclass +class KandinskyImagePipelineOutput(BaseOutput): + r""" + Output class for kandinsky image pipelines. + + Args: + image (`torch.Tensor`, `np.ndarray`, or list[PIL.Image.Image]): + List of image outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image. It can also be a NumPy array or Torch tensor of shape `(batch_size, channels, height, + width)`. + """ + + image: torch.Tensor diff --git a/diffusers/pipelines/kolors/__init__.py b/diffusers/pipelines/kolors/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..671d22e9f433802ec82ebe89fae4cef6c001064b --- /dev/null +++ b/diffusers/pipelines/kolors/__init__.py @@ -0,0 +1,54 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_sentencepiece_available, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()) and is_sentencepiece_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_and_sentencepiece_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_sentencepiece_objects)) +else: + _import_structure["pipeline_kolors"] = ["KolorsPipeline"] + _import_structure["pipeline_kolors_img2img"] = ["KolorsImg2ImgPipeline"] + _import_structure["text_encoder"] = ["ChatGLMModel"] + _import_structure["tokenizer"] = ["ChatGLMTokenizer"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()) and is_sentencepiece_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_and_sentencepiece_objects import * + + else: + from .pipeline_kolors import KolorsPipeline + from .pipeline_kolors_img2img import KolorsImg2ImgPipeline + from .text_encoder import ChatGLMModel + from .tokenizer import ChatGLMTokenizer + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/kolors/pipeline_kolors.py b/diffusers/pipelines/kolors/pipeline_kolors.py new file mode 100644 index 0000000000000000000000000000000000000000..1e11faf8b9b6b2e3194a315de56b9047199f0649 --- /dev/null +++ b/diffusers/pipelines/kolors/pipeline_kolors.py @@ -0,0 +1,1063 @@ +# Copyright 2025 Stability AI, Kwai-Kolors Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import KolorsPipelineOutput +from .text_encoder import ChatGLMModel +from .tokenizer import ChatGLMTokenizer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import KolorsPipeline + + >>> pipe = KolorsPipeline.from_pretrained( + ... "Kwai-Kolors/Kolors-diffusers", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = ( + ... "A photo of a ladybug, macro, zoom, high quality, film, holding a wooden sign with the text 'KOLORS'" + ... ) + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class KolorsPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffusionLoraLoaderMixin, IPAdapterMixin): + r""" + Pipeline for text-to-image generation using Kolors. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`ChatGLMModel`]): + Frozen text-encoder. Kolors uses [ChatGLM3-6B](https://huggingface.co/THUDM/chatglm3-6b). + tokenizer (`ChatGLMTokenizer`): + Tokenizer of class + [ChatGLMTokenizer](https://huggingface.co/THUDM/chatglm3-6b/blob/main/tokenization_chatglm.py). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"False"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `Kwai-Kolors/Kolors-diffusers`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = [ + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: ChatGLMModel, + tokenizer: ChatGLMTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + def encode_prompt( + self, + prompt, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt=None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 256, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + """ + # from IPython import embed; embed(); exit() + device = device or self._execution_device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer] + text_encoders = [self.text_encoder] + + if prompt_embeds is None: + prompt_embeds_list = [] + for tokenizer, text_encoder in zip(tokenizers, text_encoders): + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ).to(device) + output = text_encoder( + input_ids=text_inputs["input_ids"], + attention_mask=text_inputs["attention_mask"], + position_ids=text_inputs["position_ids"], + output_hidden_states=True, + ) + + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() + # [max_sequence_length, batch, hidden_size] -> [batch, hidden_size] + pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = prompt_embeds_list[0] + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + negative_prompt_embeds_list = [] + + for tokenizer, text_encoder in zip(tokenizers, text_encoders): + uncond_input = tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ).to(device) + output = text_encoder( + input_ids=uncond_input["input_ids"], + attention_mask=uncond_input["attention_mask"], + position_ids=uncond_input["position_ids"], + output_hidden_states=True, + ) + + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + negative_prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() + # [max_sequence_length, batch, hidden_size] -> [batch, hidden_size] + negative_pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=text_encoder.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = negative_prompt_embeds_list[0] + + bs_embed = pooled_prompt_embeds.shape[0] + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + num_inference_steps, + height, + width, + negative_prompt=None, + prompt_embeds=None, + pooled_prompt_embeds=None, + negative_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + if max_sequence_length is not None and max_sequence_length > 256: + raise ValueError(f"`max_sequence_length` cannot be greater than 256 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + original_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [Kwai-Kolors/Kolors-diffusers](https://huggingface.co/Kwai-Kolors/Kolors-diffusers) and checkpoints + that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [Kwai-Kolors/Kolors-diffusers](https://huggingface.co/Kwai-Kolors/Kolors-diffusers) and checkpoints + that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.kolors.KolorsPipelineOutput`] instead of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.kolors.KolorsPipelineOutput`] or `tuple`: [`~pipelines.kolors.KolorsPipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + num_inference_steps, + height, + width, + negative_prompt, + prompt_embeds, + pooled_prompt_embeds, + negative_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return KolorsPipelineOutput(images=image) diff --git a/diffusers/pipelines/kolors/pipeline_kolors_img2img.py b/diffusers/pipelines/kolors/pipeline_kolors_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..d9b5192672168d2105679de7f37d110936f02acb --- /dev/null +++ b/diffusers/pipelines/kolors/pipeline_kolors_img2img.py @@ -0,0 +1,1243 @@ +# Copyright 2025 Stability AI, Kwai-Kolors Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import IPAdapterMixin, StableDiffusionXLLoraLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import empty_device_cache, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import KolorsPipelineOutput +from .text_encoder import ChatGLMModel +from .tokenizer import ChatGLMTokenizer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import KolorsImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = KolorsImg2ImgPipeline.from_pretrained( + ... "Kwai-Kolors/Kolors-diffusers", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + >>> url = ( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/kolors/bunny_source.png" + ... ) + + + >>> init_image = load_image(url) + >>> prompt = "high quality image of a capybara wearing sunglasses. In the background of the image there are trees, poles, grass and other objects. At the bottom of the object there is the road., 8k, highly detailed." + >>> image = pipe(prompt, image=init_image).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class KolorsImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin, StableDiffusionXLLoraLoaderMixin, IPAdapterMixin): + r""" + Pipeline for text-to-image generation using Kolors. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`ChatGLMModel`]): + Frozen text-encoder. Kolors uses [ChatGLM3-6B](https://huggingface.co/THUDM/chatglm3-6b). + tokenizer (`ChatGLMTokenizer`): + Tokenizer of class + [ChatGLMTokenizer](https://huggingface.co/THUDM/chatglm3-6b/blob/main/tokenization_chatglm.py). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"False"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `Kwai-Kolors/Kolors-diffusers`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder-unet->vae" + _optional_components = [ + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: ChatGLMModel, + tokenizer: ChatGLMTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + # Copied from diffusers.pipelines.kolors.pipeline_kolors.KolorsPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt=None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 256, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + """ + # from IPython import embed; embed(); exit() + device = device or self._execution_device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer] + text_encoders = [self.text_encoder] + + if prompt_embeds is None: + prompt_embeds_list = [] + for tokenizer, text_encoder in zip(tokenizers, text_encoders): + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ).to(device) + output = text_encoder( + input_ids=text_inputs["input_ids"], + attention_mask=text_inputs["attention_mask"], + position_ids=text_inputs["position_ids"], + output_hidden_states=True, + ) + + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() + # [max_sequence_length, batch, hidden_size] -> [batch, hidden_size] + pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = prompt_embeds_list[0] + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + negative_prompt_embeds_list = [] + + for tokenizer, text_encoder in zip(tokenizers, text_encoders): + uncond_input = tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ).to(device) + output = text_encoder( + input_ids=uncond_input["input_ids"], + attention_mask=uncond_input["attention_mask"], + position_ids=uncond_input["position_ids"], + output_hidden_states=True, + ) + + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + negative_prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() + # [max_sequence_length, batch, hidden_size] -> [batch, hidden_size] + negative_pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=text_encoder.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = negative_prompt_embeds_list[0] + + bs_embed = pooled_prompt_embeds.shape[0] + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + num_inference_steps, + height, + width, + negative_prompt=None, + prompt_embeds=None, + pooled_prompt_embeds=None, + negative_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + if max_sequence_length is not None and max_sequence_length > 256: + raise ValueError(f"`max_sequence_length` cannot be greater than 256 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.prepare_latents + def prepare_latents( + self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + latents_mean = latents_std = None + if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None: + latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1) + if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None: + latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + empty_device_cache() + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + if latents_mean is not None and latents_std is not None: + latents_mean = latents_mean.to(device=device, dtype=dtype) + latents_std = latents_std.to(device=device, dtype=dtype) + init_latents = (init_latents - latents_mean) * self.vae.config.scaling_factor / latents_std + else: + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_start(self): + return self._denoising_start + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + strength: float = 0.3, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_start: float | None = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + original_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor` or `PIL.Image.Image` or `np.ndarray` or `list[torch.Tensor]` or `list[PIL.Image.Image]` or `list[np.ndarray]`): + The image(s) to modify with the pipeline. + strength (`float`, *optional*, defaults to 0.3): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. Note that in the case of + `denoising_start` being declared as an integer, the value of `strength` will be ignored. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [Kwai-Kolors/Kolors-diffusers](https://huggingface.co/Kwai-Kolors/Kolors-diffusers) and checkpoints + that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [Kwai-Kolors/Kolors-diffusers](https://huggingface.co/Kwai-Kolors/Kolors-diffusers) and checkpoints + that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refine Image + Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.kolors.KolorsPipelineOutput`] instead of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.kolors.KolorsPipelineOutput`] or `tuple`: [`~pipelines.kolors.KolorsPipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + strength, + num_inference_steps, + height, + width, + negative_prompt, + prompt_embeds, + pooled_prompt_embeds, + negative_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. Prepare timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None, + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + add_noise = True if self.denoising_start is None else False + + # 6. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + add_noise, + ) + + # 7. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 8. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 9. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 9.1 Apply denoising_end + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return KolorsPipelineOutput(images=image) diff --git a/diffusers/pipelines/kolors/pipeline_output.py b/diffusers/pipelines/kolors/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..0418191cc6e50059d7b564b641ef8ca35a61deef --- /dev/null +++ b/diffusers/pipelines/kolors/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class KolorsPipelineOutput(BaseOutput): + """ + Output class for Kolors pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/kolors/text_encoder.py b/diffusers/pipelines/kolors/text_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..434f4fed6fbbb5f359e8ed70932a02d183bb83a5 --- /dev/null +++ b/diffusers/pipelines/kolors/text_encoder.py @@ -0,0 +1,864 @@ +# Copyright 2025 ChatGLM3-6B Model Team, Kwai-Kolors Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math + +import torch +import torch.nn.functional as F +from torch import nn +from torch.nn import LayerNorm +from torch.nn.utils import skip_init +from transformers import PretrainedConfig, PreTrainedModel +from transformers.modeling_outputs import BaseModelOutputWithPast + +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +class ChatGLMConfig(PretrainedConfig): + model_type = "chatglm" + + def __init__( + self, + num_layers=28, + padded_vocab_size=65024, + hidden_size=4096, + ffn_hidden_size=13696, + kv_channels=128, + num_attention_heads=32, + seq_length=2048, + hidden_dropout=0.0, + classifier_dropout=None, + attention_dropout=0.0, + layernorm_epsilon=1e-5, + rmsnorm=True, + apply_residual_connection_post_layernorm=False, + post_layer_norm=True, + add_bias_linear=False, + add_qkv_bias=False, + bias_dropout_fusion=True, + multi_query_attention=False, + multi_query_group_num=1, + apply_query_key_layer_scaling=True, + attention_softmax_in_fp32=True, + fp32_residual_connection=False, + quantization_bit=0, + pre_seq_len=None, + prefix_projection=False, + **kwargs, + ): + self.num_layers = num_layers + self.vocab_size = padded_vocab_size + self.padded_vocab_size = padded_vocab_size + self.hidden_size = hidden_size + self.ffn_hidden_size = ffn_hidden_size + self.kv_channels = kv_channels + self.num_attention_heads = num_attention_heads + self.seq_length = seq_length + self.hidden_dropout = hidden_dropout + self.classifier_dropout = classifier_dropout + self.attention_dropout = attention_dropout + self.layernorm_epsilon = layernorm_epsilon + self.rmsnorm = rmsnorm + self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm + self.post_layer_norm = post_layer_norm + self.add_bias_linear = add_bias_linear + self.add_qkv_bias = add_qkv_bias + self.bias_dropout_fusion = bias_dropout_fusion + self.multi_query_attention = multi_query_attention + self.multi_query_group_num = multi_query_group_num + self.apply_query_key_layer_scaling = apply_query_key_layer_scaling + self.attention_softmax_in_fp32 = attention_softmax_in_fp32 + self.fp32_residual_connection = fp32_residual_connection + self.quantization_bit = quantization_bit + self.pre_seq_len = pre_seq_len + self.prefix_projection = prefix_projection + super().__init__(**kwargs) + + +class RMSNorm(torch.nn.Module): + def __init__(self, normalized_shape, eps=1e-5, device=None, dtype=None, **kwargs): + super().__init__() + self.weight = torch.nn.Parameter(torch.empty(normalized_shape, device=device, dtype=dtype)) + self.eps = eps + + def forward(self, hidden_states: torch.Tensor): + input_dtype = hidden_states.dtype + variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.eps) + + return (self.weight * hidden_states).to(input_dtype) + + +class CoreAttention(torch.nn.Module): + def __init__(self, config: ChatGLMConfig, layer_number): + super(CoreAttention, self).__init__() + + self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling + self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32 + if self.apply_query_key_layer_scaling: + self.attention_softmax_in_fp32 = True + self.layer_number = max(1, layer_number) + + projection_size = config.kv_channels * config.num_attention_heads + + # Per attention head and per partition values. + self.hidden_size_per_partition = projection_size + self.hidden_size_per_attention_head = projection_size // config.num_attention_heads + self.num_attention_heads_per_partition = config.num_attention_heads + + coeff = None + self.norm_factor = math.sqrt(self.hidden_size_per_attention_head) + if self.apply_query_key_layer_scaling: + coeff = self.layer_number + self.norm_factor *= coeff + self.coeff = coeff + + self.attention_dropout = torch.nn.Dropout(config.attention_dropout) + + def forward(self, query_layer, key_layer, value_layer, attention_mask): + pytorch_major_version = int(torch.__version__.split(".")[0]) + if pytorch_major_version >= 2: + query_layer, key_layer, value_layer = [ + k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer] + ] + if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]: + context_layer = torch.nn.functional.scaled_dot_product_attention( + query_layer, key_layer, value_layer, is_causal=True + ) + else: + if attention_mask is not None: + attention_mask = ~attention_mask + context_layer = torch.nn.functional.scaled_dot_product_attention( + query_layer, key_layer, value_layer, attention_mask + ) + context_layer = context_layer.permute(2, 0, 1, 3) + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + context_layer = context_layer.reshape(*new_context_layer_shape) + else: + # Raw attention scores + + # [b, np, sq, sk] + output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0)) + + # [sq, b, np, hn] -> [sq, b * np, hn] + query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1) + # [sk, b, np, hn] -> [sk, b * np, hn] + key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1) + + # preallocting input tensor: [b * np, sq, sk] + matmul_input_buffer = torch.empty( + output_size[0] * output_size[1], + output_size[2], + output_size[3], + dtype=query_layer.dtype, + device=query_layer.device, + ) + + # Raw attention scores. [b * np, sq, sk] + matmul_result = torch.baddbmm( + matmul_input_buffer, + query_layer.transpose(0, 1), # [b * np, sq, hn] + key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk] + beta=0.0, + alpha=(1.0 / self.norm_factor), + ) + + # change view to [b, np, sq, sk] + attention_scores = matmul_result.view(*output_size) + + # =========================== + # Attention probs and dropout + # =========================== + + # attention scores and attention mask [b, np, sq, sk] + if self.attention_softmax_in_fp32: + attention_scores = attention_scores.float() + if self.coeff is not None: + attention_scores = attention_scores * self.coeff + if attention_mask is None and attention_scores.shape[2] == attention_scores.shape[3]: + attention_mask = torch.ones( + output_size[0], 1, output_size[2], output_size[3], device=attention_scores.device, dtype=torch.bool + ) + attention_mask.tril_() + attention_mask = ~attention_mask + if attention_mask is not None: + attention_scores = attention_scores.masked_fill(attention_mask, float("-inf")) + attention_probs = F.softmax(attention_scores, dim=-1) + attention_probs = attention_probs.type_as(value_layer) + + # This is actually dropping out entire tokens to attend to, which might + # seem a bit unusual, but is taken from the original Transformer paper. + attention_probs = self.attention_dropout(attention_probs) + # ========================= + # Context layer. [sq, b, hp] + # ========================= + + # value_layer -> context layer. + # [sk, b, np, hn] --> [b, np, sq, hn] + + # context layer shape: [b, np, sq, hn] + output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3)) + # change view [sk, b * np, hn] + value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1) + # change view [b * np, sq, sk] + attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1) + # matmul: [b * np, sq, hn] + context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1)) + # change view [b, np, sq, hn] + context_layer = context_layer.view(*output_size) + # [b, np, sq, hn] --> [sq, b, np, hn] + context_layer = context_layer.permute(2, 0, 1, 3).contiguous() + # [sq, b, np, hn] --> [sq, b, hp] + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + context_layer = context_layer.view(*new_context_layer_shape) + + return context_layer + + +def split_tensor_along_last_dim( + tensor: torch.Tensor, + num_partitions: int, + contiguous_split_chunks: bool = False, +) -> list[torch.Tensor]: + """Split a tensor along its last dimension. + + Arguments: + tensor: input tensor. + num_partitions: number of partitions to split the tensor + contiguous_split_chunks: If True, make each chunk contiguous + in memory. + + Returns: + A list of Tensors + """ + # Get the size and dimension. + last_dim = tensor.dim() - 1 + last_dim_size = tensor.size()[last_dim] // num_partitions + # Split. + tensor_list = torch.split(tensor, last_dim_size, dim=last_dim) + # Note: torch.split does not create contiguous tensors by default. + if contiguous_split_chunks: + return tuple(chunk.contiguous() for chunk in tensor_list) + + return tensor_list + + +@torch.jit.script +def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor: + # x: [sq, b, np, hn] + sq, _b, np, _hn = x.size(0), x.size(1), x.size(2), x.size(3) + rot_dim = rope_cache.shape[-2] * 2 + x, x_pass = x[..., :rot_dim], x[..., rot_dim:] + # truncate to support variable sizes + rope_cache = rope_cache[:sq] + xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2) + rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2) + x_out2 = torch.stack( + [ + xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1], + xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1], + ], + -1, + ) + x_out2 = x_out2.flatten(3) + return torch.cat((x_out2, x_pass), dim=-1) + + +class SelfAttention(torch.nn.Module): + """Parallel self-attention layer abstract class. + + Self-attention layer takes input with size [s, b, h] and returns output of the same size. + """ + + def __init__(self, config: ChatGLMConfig, layer_number, device=None): + super(SelfAttention, self).__init__() + self.layer_number = max(1, layer_number) + + self.projection_size = config.kv_channels * config.num_attention_heads + + # Per attention head and per partition values. + self.hidden_size_per_attention_head = self.projection_size // config.num_attention_heads + self.num_attention_heads_per_partition = config.num_attention_heads + + self.multi_query_attention = config.multi_query_attention + self.qkv_hidden_size = 3 * self.projection_size + if self.multi_query_attention: + self.num_multi_query_groups_per_partition = config.multi_query_group_num + self.qkv_hidden_size = ( + self.projection_size + 2 * self.hidden_size_per_attention_head * config.multi_query_group_num + ) + self.query_key_value = nn.Linear( + config.hidden_size, + self.qkv_hidden_size, + bias=config.add_bias_linear or config.add_qkv_bias, + device=device, + ) + + self.core_attention = CoreAttention(config, self.layer_number) + + # Output. + self.dense = nn.Linear( + self.projection_size, + config.hidden_size, + bias=config.add_bias_linear, + device=device, + ) + + def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None): + if self.multi_query_attention: + num_attention_heads = self.num_multi_query_groups_per_partition + else: + num_attention_heads = self.num_attention_heads_per_partition + return torch.empty( + inference_max_sequence_len, + batch_size, + num_attention_heads, + self.hidden_size_per_attention_head, + dtype=dtype, + device=device, + ) + + def forward(self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True): + # hidden_states: [sq, b, h] + + # ================================================= + # Pre-allocate memory for key-values for inference. + # ================================================= + # ===================== + # Query, Key, and Value + # ===================== + + # Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)] + mixed_x_layer = self.query_key_value(hidden_states) + + if self.multi_query_attention: + (query_layer, key_layer, value_layer) = mixed_x_layer.split( + [ + self.num_attention_heads_per_partition * self.hidden_size_per_attention_head, + self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head, + self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head, + ], + dim=-1, + ) + query_layer = query_layer.view( + query_layer.size()[:-1] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head) + ) + key_layer = key_layer.view( + key_layer.size()[:-1] + + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head) + ) + value_layer = value_layer.view( + value_layer.size()[:-1] + + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head) + ) + else: + new_tensor_shape = mixed_x_layer.size()[:-1] + ( + self.num_attention_heads_per_partition, + 3 * self.hidden_size_per_attention_head, + ) + mixed_x_layer = mixed_x_layer.view(*new_tensor_shape) + + # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn] + (query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3) + + # apply relative positional encoding (rotary embedding) + if rotary_pos_emb is not None: + query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb) + key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb) + + # adjust key and value for inference + if kv_cache is not None: + cache_k, cache_v = kv_cache + key_layer = torch.cat((cache_k, key_layer), dim=0) + value_layer = torch.cat((cache_v, value_layer), dim=0) + if use_cache: + kv_cache = (key_layer, value_layer) + else: + kv_cache = None + + if self.multi_query_attention: + key_layer = key_layer.unsqueeze(-2) + key_layer = key_layer.expand( + -1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1 + ) + key_layer = key_layer.contiguous().view( + key_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head) + ) + value_layer = value_layer.unsqueeze(-2) + value_layer = value_layer.expand( + -1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1 + ) + value_layer = value_layer.contiguous().view( + value_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head) + ) + + # ================================== + # core attention computation + # ================================== + + context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask) + + # ================= + # Output. [sq, b, h] + # ================= + + output = self.dense(context_layer) + + return output, kv_cache + + +class MLP(torch.nn.Module): + """MLP. + + MLP will take the input with h hidden state, project it to 4*h hidden dimension, perform nonlinear transformation, + and project the state back into h hidden dimension. + """ + + def __init__(self, config: ChatGLMConfig, device=None): + super(MLP, self).__init__() + + self.add_bias = config.add_bias_linear + + # Project to 4h. If using swiglu double the output width, see https://huggingface.co/papers/2002.05202 + self.dense_h_to_4h = nn.Linear( + config.hidden_size, + config.ffn_hidden_size * 2, + bias=self.add_bias, + device=device, + ) + + def swiglu(x): + x = torch.chunk(x, 2, dim=-1) + return F.silu(x[0]) * x[1] + + self.activation_func = swiglu + + # Project back to h. + self.dense_4h_to_h = nn.Linear(config.ffn_hidden_size, config.hidden_size, bias=self.add_bias, device=device) + + def forward(self, hidden_states): + # [s, b, 4hp] + intermediate_parallel = self.dense_h_to_4h(hidden_states) + intermediate_parallel = self.activation_func(intermediate_parallel) + # [s, b, h] + output = self.dense_4h_to_h(intermediate_parallel) + return output + + +class GLMBlock(torch.nn.Module): + """A single transformer layer. + + Transformer layer takes input with size [s, b, h] and returns an output of the same size. + """ + + def __init__(self, config: ChatGLMConfig, layer_number, device=None): + super(GLMBlock, self).__init__() + self.layer_number = layer_number + + self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm + + self.fp32_residual_connection = config.fp32_residual_connection + + LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm + # Layernorm on the input data. + self.input_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device) + + # Self attention. + self.self_attention = SelfAttention(config, layer_number, device=device) + self.hidden_dropout = config.hidden_dropout + + # Layernorm on the attention output + self.post_attention_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device) + + # MLP + self.mlp = MLP(config, device=device) + + def forward( + self, + hidden_states, + attention_mask, + rotary_pos_emb, + kv_cache=None, + use_cache=True, + ): + # hidden_states: [s, b, h] + + # Layer norm at the beginning of the transformer layer. + layernorm_output = self.input_layernorm(hidden_states) + # Self attention. + attention_output, kv_cache = self.self_attention( + layernorm_output, attention_mask, rotary_pos_emb, kv_cache=kv_cache, use_cache=use_cache + ) + + # Residual connection. + if self.apply_residual_connection_post_layernorm: + residual = layernorm_output + else: + residual = hidden_states + + layernorm_input = torch.nn.functional.dropout(attention_output, p=self.hidden_dropout, training=self.training) + layernorm_input = residual + layernorm_input + + # Layer norm post the self attention. + layernorm_output = self.post_attention_layernorm(layernorm_input) + + # MLP. + mlp_output = self.mlp(layernorm_output) + + # Second residual connection. + if self.apply_residual_connection_post_layernorm: + residual = layernorm_output + else: + residual = layernorm_input + + output = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training) + output = residual + output + + return output, kv_cache + + +class GLMTransformer(torch.nn.Module): + """Transformer class.""" + + def __init__(self, config: ChatGLMConfig, device=None): + super(GLMTransformer, self).__init__() + + self.fp32_residual_connection = config.fp32_residual_connection + self.post_layer_norm = config.post_layer_norm + + # Number of layers. + self.num_layers = config.num_layers + + # Transformer layers. + def build_layer(layer_number): + return GLMBlock(config, layer_number, device=device) + + self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_layers)]) + + if self.post_layer_norm: + LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm + # Final layer norm before output. + self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device) + + self.gradient_checkpointing = False + + def _get_layer(self, layer_number): + return self.layers[layer_number] + + def forward( + self, + hidden_states, + attention_mask, + rotary_pos_emb, + kv_caches=None, + use_cache: bool | None = True, + output_hidden_states: bool | None = False, + ): + if not kv_caches: + kv_caches = [None for _ in range(self.num_layers)] + presents = () if use_cache else None + if torch.is_grad_enabled() and self.gradient_checkpointing: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + all_self_attentions = None + all_hidden_states = () if output_hidden_states else None + for index in range(self.num_layers): + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + layer = self._get_layer(index) + if torch.is_grad_enabled() and self.gradient_checkpointing: + layer_ret = self._gradient_checkpointing_func( + layer, hidden_states, attention_mask, rotary_pos_emb, kv_caches[index], use_cache + ) + else: + layer_ret = layer( + hidden_states, attention_mask, rotary_pos_emb, kv_cache=kv_caches[index], use_cache=use_cache + ) + hidden_states, kv_cache = layer_ret + if use_cache: + presents = presents + (kv_cache,) + + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + # Final layer norm. + if self.post_layer_norm: + hidden_states = self.final_layernorm(hidden_states) + + return hidden_states, presents, all_hidden_states, all_self_attentions + + +class ChatGLMPreTrainedModel(PreTrainedModel): + """ + An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained + models. + """ + + is_parallelizable = False + supports_gradient_checkpointing = True + config_class = ChatGLMConfig + base_model_prefix = "transformer" + _no_split_modules = ["GLMBlock"] + + def _init_weights(self, module: nn.Module): + """Initialize the weights.""" + return + + def get_masks(self, input_ids, past_key_values, padding_mask=None): + batch_size, seq_length = input_ids.shape + full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device) + full_attention_mask.tril_() + past_length = 0 + if past_key_values: + past_length = past_key_values[0][0].shape[0] + if past_length: + full_attention_mask = torch.cat( + (torch.ones(batch_size, seq_length, past_length, device=input_ids.device), full_attention_mask), dim=-1 + ) + if padding_mask is not None: + full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1) + if not past_length and padding_mask is not None: + full_attention_mask -= padding_mask.unsqueeze(-1) - 1 + full_attention_mask = (full_attention_mask < 0.5).bool() + full_attention_mask.unsqueeze_(1) + return full_attention_mask + + def get_position_ids(self, input_ids, device): + batch_size, seq_length = input_ids.shape + position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) + return position_ids + + +def default_init(cls, *args, **kwargs): + return cls(*args, **kwargs) + + +class Embedding(torch.nn.Module): + """Language model embeddings.""" + + def __init__(self, config: ChatGLMConfig, device=None): + super(Embedding, self).__init__() + + self.hidden_size = config.hidden_size + # Word embeddings (parallel). + self.word_embeddings = nn.Embedding(config.padded_vocab_size, self.hidden_size, device=device) + self.fp32_residual_connection = config.fp32_residual_connection + + def forward(self, input_ids): + # Embeddings. + words_embeddings = self.word_embeddings(input_ids) + embeddings = words_embeddings + # Data format change to avoid explicit transposes : [b s h] --> [s b h]. + embeddings = embeddings.transpose(0, 1).contiguous() + # If the input flag for fp32 residual connection is set, convert for float. + if self.fp32_residual_connection: + embeddings = embeddings.float() + return embeddings + + +class RotaryEmbedding(nn.Module): + def __init__(self, dim, original_impl=False, device=None, dtype=None): + super().__init__() + inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device).to(dtype=dtype) / dim)) + self.register_buffer("inv_freq", inv_freq) + self.dim = dim + self.original_impl = original_impl + + def forward_impl(self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000): + """Enhanced Transformer with Rotary Position Embedding. + + Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/ + transformers/rope/__init__.py. MIT License: + https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license. + """ + # $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$ + theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=torch.float, device=device) / n_elem)) + + # Create position indexes `[0, 1, ..., seq_len - 1]` + seq_idx = torch.arange(seq_len, dtype=torch.float, device=device) + + # Calculate the product of position index and $\theta_i$ + idx_theta = torch.outer(seq_idx, theta).float() + + cache = torch.stack([torch.cos(idx_theta), torch.sin(idx_theta)], dim=-1) + + # this is to mimic the behaviour of complex32, else we will get different results + if dtype in (torch.float16, torch.bfloat16, torch.int8): + cache = cache.bfloat16() if dtype == torch.bfloat16 else cache.half() + return cache + + def forward(self, max_seq_len, offset=0): + return self.forward_impl(max_seq_len, self.dim, dtype=self.inv_freq.dtype, device=self.inv_freq.device) + + +class PrefixEncoder(torch.nn.Module): + """ + The torch.nn model to encode the prefix Input shape: (batch-size, prefix-length) Output shape: (batch-size, + prefix-length, 2*layers*hidden) + """ + + def __init__(self, config: ChatGLMConfig): + super().__init__() + self.prefix_projection = config.prefix_projection + if self.prefix_projection: + # Use a two-layer MLP to encode the prefix + kv_size = config.num_layers * config.kv_channels * config.multi_query_group_num * 2 + self.embedding = torch.nn.Embedding(config.pre_seq_len, kv_size) + self.trans = torch.nn.Sequential( + torch.nn.Linear(kv_size, config.hidden_size), + torch.nn.Tanh(), + torch.nn.Linear(config.hidden_size, kv_size), + ) + else: + self.embedding = torch.nn.Embedding( + config.pre_seq_len, config.num_layers * config.kv_channels * config.multi_query_group_num * 2 + ) + + def forward(self, prefix: torch.Tensor): + if self.prefix_projection: + prefix_tokens = self.embedding(prefix) + past_key_values = self.trans(prefix_tokens) + else: + past_key_values = self.embedding(prefix) + return past_key_values + + +class ChatGLMModel(ChatGLMPreTrainedModel): + def __init__(self, config: ChatGLMConfig, device=None, empty_init=True): + super().__init__(config) + if empty_init: + init_method = skip_init + else: + init_method = default_init + init_kwargs = {} + if device is not None: + init_kwargs["device"] = device + self.embedding = init_method(Embedding, config, **init_kwargs) + self.num_layers = config.num_layers + self.multi_query_group_num = config.multi_query_group_num + self.kv_channels = config.kv_channels + + # Rotary positional embeddings + self.seq_length = config.seq_length + rotary_dim = ( + config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels + ) + + self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, original_impl=config.original_rope, device=device) + self.encoder = init_method(GLMTransformer, config, **init_kwargs) + self.output_layer = init_method( + nn.Linear, + config.hidden_size, + config.padded_vocab_size, + bias=False, + **init_kwargs, + ) + self.pre_seq_len = config.pre_seq_len + self.prefix_projection = config.prefix_projection + if self.pre_seq_len is not None: + for param in self.parameters(): + param.requires_grad = False + self.prefix_tokens = torch.arange(self.pre_seq_len).long() + self.prefix_encoder = PrefixEncoder(config) + self.dropout = torch.nn.Dropout(0.1) + + if hasattr(self, "post_init"): + self.post_init() + + def get_input_embeddings(self): + return self.embedding.word_embeddings + + def get_prompt(self, batch_size, device, dtype=torch.half): + prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device) + past_key_values = self.prefix_encoder(prefix_tokens).type(dtype) + past_key_values = past_key_values.view( + batch_size, self.pre_seq_len, self.num_layers * 2, self.multi_query_group_num, self.kv_channels + ) + # seq_len, b, nh, hidden_size + past_key_values = self.dropout(past_key_values) + past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2) + return past_key_values + + def forward( + self, + input_ids, + position_ids: torch.Tensor | None = None, + attention_mask: torch.BoolTensor | None = None, + full_attention_mask: torch.BoolTensor | None = None, + past_key_values: tuple[tuple[torch.Tensor, torch.Tensor], ...] | None = None, + inputs_embeds: torch.Tensor | None = None, + use_cache: bool | None = None, + output_hidden_states: bool | None = None, + return_dict: bool | None = None, + ): + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else getattr(self.config, "use_cache", None) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + batch_size, seq_length = input_ids.shape + + if inputs_embeds is None: + inputs_embeds = self.embedding(input_ids) + + if self.pre_seq_len is not None: + if past_key_values is None: + past_key_values = self.get_prompt( + batch_size=batch_size, device=input_ids.device, dtype=inputs_embeds.dtype + ) + if attention_mask is not None: + attention_mask = torch.cat( + [attention_mask.new_ones((batch_size, self.pre_seq_len)), attention_mask], dim=-1 + ) + + if full_attention_mask is None: + if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1): + full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask) + + # Rotary positional embeddings + rotary_pos_emb = self.rotary_pos_emb(self.seq_length) + if position_ids is not None: + rotary_pos_emb = rotary_pos_emb[position_ids] + else: + rotary_pos_emb = rotary_pos_emb[None, :seq_length] + rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous() + + # Run encoder. + hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder( + inputs_embeds, + full_attention_mask, + rotary_pos_emb=rotary_pos_emb, + kv_caches=past_key_values, + use_cache=use_cache, + output_hidden_states=output_hidden_states, + ) + + if not return_dict: + return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None) + + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=presents, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + ) diff --git a/diffusers/pipelines/kolors/tokenizer.py b/diffusers/pipelines/kolors/tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..5fc475ddc0791cb95950515472fd3b0bcaee6ec3 --- /dev/null +++ b/diffusers/pipelines/kolors/tokenizer.py @@ -0,0 +1,345 @@ +# Copyright 2025 ChatGLM3-6B Model Team, Kwai-Kolors Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import json +import os +import re + +from transformers import PreTrainedTokenizer +from transformers.tokenization_utils_base import BatchEncoding, EncodedInput +from transformers.utils import PaddingStrategy + +from ...utils import is_sentencepiece_available + + +class SPTokenizer: + def __init__(self, model_path: str): + if not is_sentencepiece_available(): + raise ImportError( + "`SPTokenizer` requires the `sentencepiece` library but it was not found in your environment. You can " + "install it with `pip install sentencepiece`." + ) + from sentencepiece import SentencePieceProcessor + + # reload tokenizer + assert os.path.isfile(model_path), model_path + self.sp_model = SentencePieceProcessor(model_file=model_path) + + # BOS / EOS token IDs + self.n_words: int = self.sp_model.vocab_size() + self.bos_id: int = self.sp_model.bos_id() + self.eos_id: int = self.sp_model.eos_id() + self.pad_id: int = self.sp_model.unk_id() + assert self.sp_model.vocab_size() == self.sp_model.get_piece_size() + + role_special_tokens = ["<|system|>", "<|user|>", "<|assistant|>", "<|observation|>"] + special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"] + role_special_tokens + self.special_tokens = {} + self.index_special_tokens = {} + for token in special_tokens: + self.special_tokens[token] = self.n_words + self.index_special_tokens[self.n_words] = token + self.n_words += 1 + self.role_special_token_expression = "|".join([re.escape(token) for token in role_special_tokens]) + + def tokenize(self, s: str, encode_special_tokens=False): + if encode_special_tokens: + last_index = 0 + t = [] + for match in re.finditer(self.role_special_token_expression, s): + if last_index < match.start(): + t.extend(self.sp_model.EncodeAsPieces(s[last_index : match.start()])) + t.append(s[match.start() : match.end()]) + last_index = match.end() + if last_index < len(s): + t.extend(self.sp_model.EncodeAsPieces(s[last_index:])) + return t + else: + return self.sp_model.EncodeAsPieces(s) + + def encode(self, s: str, bos: bool = False, eos: bool = False) -> list[int]: + assert isinstance(s, str) + t = self.sp_model.encode(s) + if bos: + t = [self.bos_id] + t + if eos: + t = t + [self.eos_id] + return t + + def decode(self, t: list[int]) -> str: + text, buffer = "", [] + for token in t: + if token in self.index_special_tokens: + if buffer: + text += self.sp_model.decode(buffer) + buffer = [] + text += self.index_special_tokens[token] + else: + buffer.append(token) + if buffer: + text += self.sp_model.decode(buffer) + return text + + def decode_tokens(self, tokens: list[str]) -> str: + text = self.sp_model.DecodePieces(tokens) + return text + + def convert_token_to_id(self, token): + """Converts a token (str) in an id using the vocab.""" + if token in self.special_tokens: + return self.special_tokens[token] + return self.sp_model.PieceToId(token) + + def convert_id_to_token(self, index): + """Converts an index (integer) in a token (str) using the vocab.""" + if index in self.index_special_tokens: + return self.index_special_tokens[index] + if index in [self.eos_id, self.bos_id, self.pad_id] or index < 0: + return "" + return self.sp_model.IdToPiece(index) + + +class ChatGLMTokenizer(PreTrainedTokenizer): + vocab_files_names = {"vocab_file": "tokenizer.model"} + + model_input_names = ["input_ids", "attention_mask", "position_ids"] + + def __init__( + self, + vocab_file, + padding_side="left", + clean_up_tokenization_spaces=False, + encode_special_tokens=False, + **kwargs, + ): + self.name = "GLMTokenizer" + + self.vocab_file = vocab_file + self.tokenizer = SPTokenizer(vocab_file) + self.special_tokens = { + "": self.tokenizer.bos_id, + "": self.tokenizer.eos_id, + "": self.tokenizer.pad_id, + } + self.encode_special_tokens = encode_special_tokens + super().__init__( + padding_side=padding_side, + clean_up_tokenization_spaces=clean_up_tokenization_spaces, + encode_special_tokens=encode_special_tokens, + **kwargs, + ) + + def get_command(self, token): + if token in self.special_tokens: + return self.special_tokens[token] + assert token in self.tokenizer.special_tokens, f"{token} is not a special token for {self.name}" + return self.tokenizer.special_tokens[token] + + @property + def unk_token(self) -> str: + return "" + + @unk_token.setter + def unk_token(self, value: str): + self._unk_token = value + + @property + def pad_token(self) -> str: + return "" + + @pad_token.setter + def pad_token(self, value: str): + self._pad_token = value + + @property + def pad_token_id(self): + return self.get_command("") + + @property + def eos_token(self) -> str: + return "" + + @eos_token.setter + def eos_token(self, value: str): + self._eos_token = value + + @property + def eos_token_id(self): + return self.get_command("") + + @property + def vocab_size(self): + return self.tokenizer.n_words + + def get_vocab(self): + """Returns vocab as a dict""" + vocab = {self._convert_id_to_token(i): i for i in range(self.vocab_size)} + vocab.update(self.added_tokens_encoder) + return vocab + + def _tokenize(self, text, **kwargs): + return self.tokenizer.tokenize(text, encode_special_tokens=self.encode_special_tokens) + + def _convert_token_to_id(self, token): + """Converts a token (str) in an id using the vocab.""" + return self.tokenizer.convert_token_to_id(token) + + def _convert_id_to_token(self, index): + """Converts an index (integer) in a token (str) using the vocab.""" + return self.tokenizer.convert_id_to_token(index) + + def convert_tokens_to_string(self, tokens: list[str]) -> str: + return self.tokenizer.decode_tokens(tokens) + + def save_vocabulary(self, save_directory, filename_prefix=None): + """ + Save the vocabulary and special tokens file to a directory. + + Args: + save_directory (`str`): + The directory in which to save the vocabulary. + filename_prefix (`str`, *optional*): + An optional prefix to add to the named of the saved files. + + Returns: + `tuple(str)`: Paths to the files saved. + """ + if os.path.isdir(save_directory): + vocab_file = os.path.join(save_directory, self.vocab_files_names["vocab_file"]) + else: + vocab_file = save_directory + + with open(self.vocab_file, "rb") as fin: + proto_str = fin.read() + + with open(vocab_file, "wb") as writer: + writer.write(proto_str) + + return (vocab_file,) + + def get_prefix_tokens(self): + prefix_tokens = [self.get_command("[gMASK]"), self.get_command("sop")] + return prefix_tokens + + def build_single_message(self, role, metadata, message): + assert role in ["system", "user", "assistant", "observation"], role + role_tokens = [self.get_command(f"<|{role}|>")] + self.tokenizer.encode(f"{metadata}\n") + message_tokens = self.tokenizer.encode(message) + tokens = role_tokens + message_tokens + return tokens + + def build_chat_input(self, query, history=None, role="user"): + if history is None: + history = [] + input_ids = [] + for item in history: + content = item["content"] + if item["role"] == "system" and "tools" in item: + content = content + "\n" + json.dumps(item["tools"], indent=4, ensure_ascii=False) + input_ids.extend(self.build_single_message(item["role"], item.get("metadata", ""), content)) + input_ids.extend(self.build_single_message(role, "", query)) + input_ids.extend([self.get_command("<|assistant|>")]) + return self.batch_encode_plus([input_ids], return_tensors="pt", is_split_into_words=True) + + def build_inputs_with_special_tokens( + self, token_ids_0: list[int], token_ids_1: list[int] | None = None + ) -> list[int]: + """ + Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and + adding special tokens. A BERT sequence has the following format: + + - single sequence: `[CLS] X [SEP]` + - pair of sequences: `[CLS] A [SEP] B [SEP]` + + Args: + token_ids_0 (`list[int]`): + list of IDs to which the special tokens will be added. + token_ids_1 (`list[int]`, *optional*): + Optional second list of IDs for sequence pairs. + + Returns: + `list[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens. + """ + prefix_tokens = self.get_prefix_tokens() + token_ids_0 = prefix_tokens + token_ids_0 + if token_ids_1 is not None: + token_ids_0 = token_ids_0 + token_ids_1 + [self.get_command("")] + return token_ids_0 + + def _pad( + self, + encoded_inputs: dict[str, EncodedInput] | BatchEncoding, + max_length: int | None = None, + padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD, + pad_to_multiple_of: int | None = None, + return_attention_mask: bool | None = None, + padding_side: bool | None = None, + ) -> dict: + """ + Pad encoded inputs (on left/right and up to predefined length or max length in the batch) + + Args: + encoded_inputs: + Dictionary of tokenized inputs (`list[int]`) or batch of tokenized inputs (`list[list[int]]`). + max_length: maximum length of the returned list and optionally padding length (see below). + Will truncate by taking into account the special tokens. + padding_strategy: PaddingStrategy to use for padding. + + - PaddingStrategy.LONGEST Pad to the longest sequence in the batch + - PaddingStrategy.MAX_LENGTH: Pad to the max length (default) + - PaddingStrategy.DO_NOT_PAD: Do not pad + The tokenizer padding sides are defined in self.padding_side: + + - 'left': pads on the left of the sequences + - 'right': pads on the right of the sequences + pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value. + This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability + `>= 7.5` (Volta). + padding_side (`str`, *optional*): + The side on which the model should have padding applied. Should be selected between ['right', 'left']. + Default value is picked from the class attribute of the same name. + return_attention_mask: + (optional) Set to False to avoid returning attention mask (default: set to model specifics) + """ + # Load from model defaults + assert self.padding_side == "left" + + required_input = encoded_inputs[self.model_input_names[0]] + seq_length = len(required_input) + + if padding_strategy == PaddingStrategy.LONGEST: + max_length = len(required_input) + + if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0): + max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of + + needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length + + # Initialize attention mask if not present. + if "attention_mask" not in encoded_inputs: + encoded_inputs["attention_mask"] = [1] * seq_length + + if "position_ids" not in encoded_inputs: + encoded_inputs["position_ids"] = list(range(seq_length)) + + if needs_to_be_padded: + difference = max_length - len(required_input) + + if "attention_mask" in encoded_inputs: + encoded_inputs["attention_mask"] = [0] * difference + encoded_inputs["attention_mask"] + if "position_ids" in encoded_inputs: + encoded_inputs["position_ids"] = [0] * difference + encoded_inputs["position_ids"] + encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input + + return encoded_inputs diff --git a/diffusers/pipelines/krea2/__init__.py b/diffusers/pipelines/krea2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..cff52b1b5488b8316bd06319a61b97cdbd061d3b --- /dev/null +++ b/diffusers/pipelines/krea2/__init__.py @@ -0,0 +1,47 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa: F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_krea2"] = ["Krea2Pipeline"] + _import_structure["pipeline_output"] = ["Krea2PipelineOutput"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_krea2 import Krea2Pipeline + from .pipeline_output import Krea2PipelineOutput +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/krea2/pipeline_krea2.py b/diffusers/pipelines/krea2/pipeline_krea2.py new file mode 100644 index 0000000000000000000000000000000000000000..51d33cb4861903ee1ac682f8da3b7256013656aa --- /dev/null +++ b/diffusers/pipelines/krea2/pipeline_krea2.py @@ -0,0 +1,719 @@ +# Copyright 2026 Krea AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import AutoTokenizer, Qwen3VLModel + +from ...image_processor import VaeImageProcessor +from ...loaders import Krea2LoraLoaderMixin +from ...models import AutoencoderKLQwenImage, Krea2Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import Krea2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import Krea2Pipeline + + >>> # Load from a local directory produced by the Krea 2 conversion (no hub repo yet). + >>> pipe = Krea2Pipeline.from_pretrained("path/to/krea2-diffusers", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "a fox in the snow" + >>> # Base (midtrain) checkpoint defaults. For the few-step distilled (TDM) checkpoint use + >>> # `num_inference_steps=8, guidance_scale=0.0` instead. + >>> image = pipe(prompt, num_inference_steps=28, guidance_scale=4.5).images[0] + >>> image.save("krea2.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class Krea2Pipeline(DiffusionPipeline, Krea2LoraLoaderMixin): + r""" + The Krea 2 pipeline for text-to-image generation. + + Args: + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + Euler flow-matching scheduler. The Krea 2 sigma schedule is the resolution-aware exponential time shift, so + the scheduler config is expected to set `use_dynamic_shifting=True` together with the Krea 2 shift + parameters (`base_shift=0.5`, `max_shift=1.15`, `base_image_seq_len=256`, `max_image_seq_len=6400`). + vae ([`AutoencoderKLQwenImage`]): + The Qwen-Image variational auto-encoder (f8, 16 latent channels) used to decode latents to images. + text_encoder ([`~transformers.PreTrainedModel`]): + A Qwen3-VL model (e.g. `Qwen3VLModel` of `Qwen/Qwen3-VL-4B-Instruct`). The pipeline consumes a stack of + hidden states tapped from several decoder layers rather than the last hidden state. + tokenizer ([`~transformers.AutoTokenizer`]): + The tokenizer paired with the text encoder. + transformer ([`Krea2Transformer2DModel`]): + The Krea 2 single-stream MMDiT that predicts the flow-matching velocity. + text_encoder_select_layers (`tuple[int, ...]`, *optional*): + Indices into the text encoder's `hidden_states` tuple (0 is the embedding output) whose states are stacked + per token as the transformer's text conditioning. Must have `transformer.config.num_text_layers` entries. + is_distilled (`bool`, *optional*, defaults to `False`): + Whether the transformer is the few-step distilled (TDM/turbo) checkpoint. When `True` a fixed timestep + shift `mu=1.15` is used; otherwise `mu` is computed from the image resolution. + patch_size (`int`, *optional*, defaults to 2): + Side length of the square patches the latents are packed into before entering the transformer. The + effective pixel-to-token downsampling factor is `vae_scale_factor * patch_size`. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen3VLModel, + tokenizer: AutoTokenizer, + transformer: Krea2Transformer2DModel, + text_encoder_select_layers: tuple[int, ...] | list[int] | None = None, + is_distilled: bool = False, + patch_size: int = 2, + ): + super().__init__() + + self.register_modules( + scheduler=scheduler, + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + ) + # Indices into the text encoder's `hidden_states` tuple (0 is the embedding output) whose states are stacked + # per token and fed to the transformer's text fusion stage. `None` selects the Krea 2 (Qwen3-VL-4B) taps. + if text_encoder_select_layers is None: + text_encoder_select_layers = (2, 5, 8, 11, 14, 17, 20, 23, 26, 29, 32, 35) + self.register_to_config(text_encoder_select_layers=tuple(text_encoder_select_layers)) + self.text_encoder_select_layers = tuple(text_encoder_select_layers) + # The few-step distilled (TDM/turbo) checkpoint uses a fixed timestep-shift `mu=1.15`; the base (midtrain) + # checkpoint computes `mu` from the image resolution. Encoded here so each checkpoint carries the right schedule. + self.register_to_config(is_distilled=is_distilled) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # Latents are packed into `patch_size`-square patches before entering the transformer, so the effective + # pixel-to-token downsampling factor is vae_scale_factor * patch_size. + self.register_to_config(patch_size=patch_size) + self.patch_size = patch_size + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * self.patch_size) + + # Text conditioning uses the Qwen-Image chat template, tokenized as a fixed-length block: the prompt is padded + # to a fixed length first and the assistant suffix is appended after the padding (matching how the model was + # sampled at training time). The first `prompt_template_encode_start_idx` (system prefix) tokens are dropped + # from the encoder outputs. + self.prompt_template_encode_prefix = ( + "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, " + "spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n" + ) + self.prompt_template_encode_suffix = "<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.prompt_template_encode_num_suffix_tokens = 5 + + def get_text_hidden_states( + self, + prompt: str | list[str], + max_sequence_length: int = 512, + device: torch.device | None = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + """Tokenize `prompt` into the fixed-length Krea 2 layout and tap the selected encoder hidden states. + + Returns a `(hidden_states, attention_mask)` tuple of shapes `(batch_size, text_seq_len, num_text_layers, + text_hidden_dim)` and `(batch_size, text_seq_len)` (bool). + """ + device = device or self._execution_device + prompt = [prompt] if isinstance(prompt, str) else prompt + prefix_idx = self.prompt_template_encode_start_idx + text = [self.prompt_template_encode_prefix + e for e in prompt] + text_tokens = self.tokenizer( + text, + truncation=True, + padding="max_length", + max_length=max_sequence_length + prefix_idx - self.prompt_template_encode_num_suffix_tokens, + return_tensors="pt", + ).to(device) + suffix_tokens = self.tokenizer([self.prompt_template_encode_suffix] * len(text), return_tensors="pt").to( + device + ) + + input_ids = torch.cat([text_tokens.input_ids, suffix_tokens.input_ids], dim=1) + attention_mask = torch.cat([text_tokens.attention_mask, suffix_tokens.attention_mask], dim=1).bool() + + # Krea 2 pads in the middle of the template (`[prefix | prompt | PAD | suffix]`), so the suffix tokens sit + # downstream of the padding. The text features must use positions that count only real tokens (padding does + # not consume a position) to match how the model was trained; otherwise the suffix gets a shifted mRoPE phase. + # `Qwen3VLModel`'s default raw-index positions would place the suffix at ~max_length instead. Build the + # cumulative-valid-token positions explicitly and broadcast across the 3 mRoPE axes (T/H/W are equal for text). + position_ids = (attention_mask.long().cumsum(dim=-1) - 1).clamp(min=0) + position_ids = position_ids.unsqueeze(0).expand(3, -1, -1) + + outputs = self.text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + output_hidden_states=True, + ) + hidden_states = torch.stack([outputs.hidden_states[i] for i in self.text_encoder_select_layers], dim=2) + + hidden_states = hidden_states[:, prefix_idx:] + attention_mask = attention_mask[:, prefix_idx:] + return hidden_states, attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 512, + ) -> tuple[torch.Tensor, torch.Tensor]: + r""" + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings of shape `(batch_size, text_seq_len, num_text_layers, text_hidden_dim)`. + Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will + be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Pre-generated boolean mask marking valid text tokens, of shape `(batch_size, text_seq_len)`. Required + when `prompt_embeds` is passed. + max_sequence_length (`int`, defaults to 512): + Fixed text sequence length consumed by the transformer; prompts are padded or truncated to it. + """ + device = device or self._execution_device + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self.get_text_hidden_states(prompt, max_sequence_length, device) + + batch_size, seq_len, num_text_layers, dim = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, num_text_layers, dim) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + multiple = self.vae_scale_factor * self.patch_size + if height % multiple != 0 or width % multiple != 0: + raise ValueError(f"`height` and `width` must be divisible by {multiple} but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length <= 0: + raise ValueError(f"`max_sequence_length` must be a positive integer but is {max_sequence_length}") + + def _pack_latents(self, latents, batch_size, num_channels_latents, height, width): + p = self.patch_size + latents = latents.view(batch_size, num_channels_latents, height // p, p, width // p, p) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // p) * (width // p), num_channels_latents * p * p) + + return latents + + def _unpack_latents(self, latents, height, width): + batch_size, _, channels = latents.shape + p = self.patch_size + + # The VAE applies `vae_scale_factor`x compression, and latents are packed into `p`-square patches, so latent + # height and width must be divisible by `p`. + height = p * (int(height) // (self.vae_scale_factor * p)) + width = p * (int(width) // (self.vae_scale_factor * p)) + + latents = latents.view(batch_size, height // p, width // p, channels // (p * p), p, p) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (p * p), 1, height, width) + + return latents + + @staticmethod + def prepare_position_ids(text_seq_len: int, grid_height: int, grid_width: int, device: torch.device): + """Build the `(text_seq_len + grid_height * grid_width, 3)` rotary coordinates for the combined sequence: + text tokens sit at the origin, image tokens carry their `(0, h, w)` latent-grid coordinates.""" + text_ids = torch.zeros(text_seq_len, 3, device=device) + image_ids = torch.zeros(grid_height, grid_width, 3, device=device) + image_ids[..., 1] = torch.arange(grid_height, device=device)[:, None] + image_ids[..., 2] = torch.arange(grid_width, device=device)[None, :] + image_ids = image_ids.reshape(grid_height * grid_width, 3) + return torch.cat([text_ids, image_ids], dim=0) + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + latent_height = height // self.vae_scale_factor + latent_width = width // self.vae_scale_factor + shape = (batch_size, num_channels_latents, latent_height, latent_width) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, latent_height, latent_width) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 0 + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + negative_prompt: str | list[str] | None = None, + height: int = 1024, + width: int = 1024, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int, dict], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + attention_kwargs: dict[str, Any] | None = None, + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when `guidance_scale <= 0`; defaults + to an empty prompt when guidance is enabled. + height (`int`, defaults to 1024): + The height in pixels of the generated image. Rounded up to a multiple of 16 if needed. + width (`int`, defaults to 1024): + The width in pixels of the generated image. Rounded up to a multiple of 16 if needed. + num_inference_steps (`int`, defaults to 28): + The number of denoising steps. Use 28 for the base (midtrain) checkpoint and 8 for the few-step + distilled (TDM) checkpoint. + sigmas (`list[float]`, *optional*): + Custom sigmas for the scheduler. If not defined, the default `linspace(1.0, 1/num_inference_steps, + num_inference_steps)` grid is used (the resolution-aware shift is applied inside the scheduler). + guidance_scale (`float`, defaults to 4.5): + Classifier-free guidance scale, following the Krea 2 convention: the velocity is computed as `cond + + guidance_scale * (cond - uncond)` and guidance is enabled whenever `guidance_scale > 0` (this equals + the usual CFG formulation with scale `1 + guidance_scale`). Set to `0.0` to disable (e.g. for the TDM + checkpoint). + num_images_per_prompt (`int`, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or more [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) to + make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents in packed form `(batch_size, image_seq_len, in_channels)`, sampled from a + Gaussian distribution, to be used as inputs for image generation. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings of shape `(batch_size, text_seq_len, num_text_layers, text_hidden_dim)`. + If not provided, embeddings are generated from `prompt`. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Boolean mask for `prompt_embeds`; required when `prompt_embeds` is passed. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings; same layout as `prompt_embeds`. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Boolean mask for `negative_prompt_embeds`; required when `negative_prompt_embeds` is passed. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `"pil"`, `"np"`, `"pt"` or `"latent"`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.krea2.Krea2PipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that is called at the end of each denoising step with `callback_on_step_end(self, step, + timestep, callback_kwargs)`. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*, defaults to `["latents"]`): + The list of tensor inputs for the `callback_on_step_end` function. Must be a subset of + `._callback_tensor_inputs`. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + max_sequence_length (`int`, defaults to 512): + Fixed text sequence length consumed by the transformer; prompts are padded or truncated to it. + + Examples: + + Returns: + [`~pipelines.krea2.Krea2PipelineOutput`] or `tuple`: [`~pipelines.krea2.Krea2PipelineOutput`] if + `return_dict` is True, otherwise a `tuple`, whose first element is a list with the generated images. + """ + multiple = self.vae_scale_factor * self.patch_size + if height % multiple != 0 or width % multiple != 0: + rounded_height = ((height + multiple - 1) // multiple) * multiple + rounded_width = ((width + multiple - 1) // multiple) * multiple + logger.warning( + f"`height` and `width` must be multiples of {multiple}; rounding up from {height}x{width} to" + f" {rounded_height}x{rounded_width}." + ) + height, width = rounded_height, rounded_width + + # 1. Check inputs + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode the prompts + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + max_sequence_length=max_sequence_length, + ) + if self.do_classifier_free_guidance: + if negative_prompt is None and negative_prompt_embeds is None: + negative_prompt = "" + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latents and position ids + num_channels_latents = self.transformer.config.in_channels // (self.patch_size**2) + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + grid_height = height // (self.vae_scale_factor * self.patch_size) + grid_width = width // (self.vae_scale_factor * self.patch_size) + position_ids = self.prepare_position_ids(prompt_embeds.shape[1], grid_height, grid_width, device) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + if self.config.is_distilled: + mu = 1.15 + else: + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 6400), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = (t / self.scheduler.config.num_train_timesteps).expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latents, + encoder_hidden_states=prompt_embeds, + timestep=timestep, + position_ids=position_ids, + encoder_attention_mask=prompt_embeds_mask, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + neg_noise_pred = self.transformer( + hidden_states=latents, + encoder_hidden_states=negative_prompt_embeds, + timestep=timestep, + position_ids=position_ids, + encoder_attention_mask=negative_prompt_embeds_mask, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred + guidance_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + # 7. Decode latents + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return Krea2PipelineOutput(images=image) diff --git a/diffusers/pipelines/krea2/pipeline_output.py b/diffusers/pipelines/krea2/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..dd39a4d37c6275ad212bc8f3f8cb1216167bcbf4 --- /dev/null +++ b/diffusers/pipelines/krea2/pipeline_output.py @@ -0,0 +1,34 @@ +# Copyright 2026 Krea AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class Krea2PipelineOutput(BaseOutput): + """ + Output class for the Krea 2 pipeline. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`): + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/latent_consistency_models/__init__.py b/diffusers/pipelines/latent_consistency_models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8f79d3c4773f393ed689a949041d36ad77e20968 --- /dev/null +++ b/diffusers/pipelines/latent_consistency_models/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_latent_consistency_img2img"] = ["LatentConsistencyModelImg2ImgPipeline"] + _import_structure["pipeline_latent_consistency_text2img"] = ["LatentConsistencyModelPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_latent_consistency_img2img import LatentConsistencyModelImg2ImgPipeline + from .pipeline_latent_consistency_text2img import LatentConsistencyModelPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py b/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..424a2c46e06b9152585835fd54616fd14f3f9e98 --- /dev/null +++ b/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py @@ -0,0 +1,996 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import LCMScheduler +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import AutoPipelineForImage2Image + >>> import torch + >>> import PIL + + >>> pipe = AutoPipelineForImage2Image.from_pretrained("SimianLuo/LCM_Dreamshaper_v7") + >>> # To save GPU memory, torch.float16 can be used, but it may compromise image quality. + >>> pipe.to(torch_device="cuda", torch_dtype=torch.float32) + + >>> prompt = "High altitude snowy mountains" + >>> image = PIL.Image.open("./snowy_mountains.png") + + >>> # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps. + >>> num_inference_steps = 4 + >>> images = pipe( + ... prompt=prompt, image=image, num_inference_steps=num_inference_steps, guidance_scale=8.0 + ... ).images + + >>> images[0].save("image.png") + ``` + +""" + + +class LatentConsistencyModelImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for image-to-image generation using a latent consistency model. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Currently only + supports [`LCMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + requires_safety_checker (`bool`, *optional*, defaults to `True`): + Whether the pipeline requires a safety checker component. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "denoised", "prompt_embeds", "w_embedding"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: LCMScheduler, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection | None = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt: str | list[str], + strength: float, + callback_steps: int, + prompt_embeds: torch.Tensor | None = None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def clip_skip(self): + return self._clip_skip + + @property + def do_classifier_free_guidance(self): + return False + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + num_inference_steps: int = 4, + strength: float = 0.8, + original_inference_steps: int = None, + timesteps: list[int] = None, + guidance_scale: float = 8.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image` or tensor representing an image batch to be used as the starting point. Can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + original_inference_steps (`int`, *optional*): + The original number of inference steps use to generate a linearly-spaced timestep schedule, from which + we will draw `num_inference_steps` evenly spaced timesteps from as our final timestep schedule, + following the Skipping-Step method in the paper (see Section 4.3). If not set this will default to the + scheduler's `original_inference_steps` attribute. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` + timesteps on the original LCM training/distillation timestep schedule are used. Must be in descending + order. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + Note that the original latent consistency models paper uses a different CFG formulation where the + guidance scales are decreased by 1 (so in the paper formulation CFG is enabled when `guidance_scale > + 0`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + strength, + callback_steps, + prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + # NOTE: when a LCM is distilled from an LDM via latent consistency distillation (Algorithm 1) with guided + # distillation, the forward pass of the LCM learns to approximate sampling from the LDM using CFG with the + # unconditional prompt "" (the empty string). Due to this, LCMs currently do not support negative prompts. + prompt_embeds, _ = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=None, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Encode image + image = self.image_processor.preprocess(image) + + # 5. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + timesteps, + original_inference_steps=original_inference_steps, + strength=strength, + ) + + # 6. Prepare latent variables + original_inference_steps = ( + original_inference_steps + if original_inference_steps is not None + else self.scheduler.config.original_inference_steps + ) + latent_timestep = timesteps[:1] + if latents is None: + latents = self.prepare_latents( + image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator + ) + bs = batch_size * num_images_per_prompt + + # 6. Get Guidance Scale Embedding + # NOTE: We use the Imagen CFG formulation that StableDiffusionPipeline uses rather than the original LCM paper + # CFG formulation, so we need to subtract 1 from the input guidance_scale. + # LCM CFG formulation: cfg_noise = noise_cond + cfg_scale * (noise_cond - noise_uncond), (cfg_scale > 0.0 using CFG) + w = torch.tensor(self.guidance_scale - 1).repeat(bs) + w_embedding = self.get_guidance_scale_embedding(w, embedding_dim=self.unet.config.time_cond_proj_dim).to( + device=device, dtype=latents.dtype + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, None) + + # 7.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 8. LCM Multistep Sampling Loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + latents = latents.to(prompt_embeds.dtype) + + # model prediction (v-prediction, eps, x) + model_pred = self.unet( + latents, + t, + timestep_cond=w_embedding, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents, denoised = self.scheduler.step(model_pred, t, latents, **extra_step_kwargs, return_dict=False) + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + w_embedding = callback_outputs.pop("w_embedding", w_embedding) + denoised = callback_outputs.pop("denoised", denoised) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + denoised = denoised.to(prompt_embeds.dtype) + if not output_type == "latent": + image = self.vae.decode(denoised / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = denoised + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py b/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py new file mode 100644 index 0000000000000000000000000000000000000000..60f59ec7f9d3dd6416cacccb2ef11b665dbb8faf --- /dev/null +++ b/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py @@ -0,0 +1,925 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import LCMScheduler +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import DiffusionPipeline + >>> import torch + + >>> pipe = DiffusionPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7") + >>> # To save GPU memory, torch.float16 can be used, but it may compromise image quality. + >>> pipe.to(torch_device="cuda", torch_dtype=torch.float32) + + >>> prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k" + + >>> # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps. + >>> num_inference_steps = 4 + >>> images = pipe(prompt=prompt, num_inference_steps=num_inference_steps, guidance_scale=8.0).images + >>> images[0].save("image.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class LatentConsistencyModelPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-to-image generation using a latent consistency model. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Currently only + supports [`LCMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + requires_safety_checker (`bool`, *optional*, defaults to `True`): + Whether the pipeline requires a safety checker component. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "denoised", "prompt_embeds", "w_embedding"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: LCMScheduler, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection | None = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Currently StableDiffusionPipeline.check_inputs with negative prompt stuff removed + def check_inputs( + self, + prompt: str | list[str], + height: int, + width: int, + callback_steps: int, + prompt_embeds: torch.Tensor | None = None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def clip_skip(self): + return self._clip_skip + + @property + def do_classifier_free_guidance(self): + return False + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 4, + original_inference_steps: int = None, + timesteps: list[int] = None, + guidance_scale: float = 8.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + original_inference_steps (`int`, *optional*): + The original number of inference steps use to generate a linearly-spaced timestep schedule, from which + we will draw `num_inference_steps` evenly spaced timesteps from as our final timestep schedule, + following the Skipping-Step method in the paper (see Section 4.3). If not set this will default to the + scheduler's `original_inference_steps` attribute. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` + timesteps on the original LCM training/distillation timestep schedule are used. Must be in descending + order. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + Note that the original latent consistency models paper uses a different CFG formulation where the + guidance scales are decreased by 1 (so in the paper formulation CFG is enabled when `guidance_scale > + 0`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + # NOTE: when a LCM is distilled from an LDM via latent consistency distillation (Algorithm 1) with guided + # distillation, the forward pass of the LCM learns to approximate sampling from the LDM using CFG with the + # unconditional prompt "" (the empty string). Due to this, LCMs currently do not support negative prompts. + prompt_embeds, _ = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=None, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + timesteps, + original_inference_steps=original_inference_steps, + ) + + # 5. Prepare latent variable + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + bs = batch_size * num_images_per_prompt + + # 6. Get Guidance Scale Embedding + # NOTE: We use the Imagen CFG formulation that StableDiffusionPipeline uses rather than the original LCM paper + # CFG formulation, so we need to subtract 1 from the input guidance_scale. + # LCM CFG formulation: cfg_noise = noise_cond + cfg_scale * (noise_cond - noise_uncond), (cfg_scale > 0.0 using CFG) + w = torch.tensor(self.guidance_scale - 1).repeat(bs) + w_embedding = self.get_guidance_scale_embedding(w, embedding_dim=self.unet.config.time_cond_proj_dim).to( + device=device, dtype=latents.dtype + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, None) + + # 7.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 8. LCM MultiStep Sampling Loop: + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + latents = latents.to(prompt_embeds.dtype) + + # model prediction (v-prediction, eps, x) + model_pred = self.unet( + latents, + t, + timestep_cond=w_embedding, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents, denoised = self.scheduler.step(model_pred, t, latents, **extra_step_kwargs, return_dict=False) + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + w_embedding = callback_outputs.pop("w_embedding", w_embedding) + denoised = callback_outputs.pop("denoised", denoised) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + denoised = denoised.to(prompt_embeds.dtype) + if not output_type == "latent": + image = self.vae.decode(denoised / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = denoised + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/latent_diffusion/__init__.py b/diffusers/pipelines/latent_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..561f96fc71dc7b4404e09571e0b7eaa4ee02fde8 --- /dev/null +++ b/diffusers/pipelines/latent_diffusion/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_latent_diffusion"] = ["LDMBertModel", "LDMTextToImagePipeline"] + _import_structure["pipeline_latent_diffusion_superresolution"] = ["LDMSuperResolutionPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_latent_diffusion import LDMBertModel, LDMTextToImagePipeline + from .pipeline_latent_diffusion_superresolution import LDMSuperResolutionPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py b/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..cd7d527b87127c38a999a462eb2be7045186ea5d --- /dev/null +++ b/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py @@ -0,0 +1,749 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect + +import torch +import torch.nn as nn +from transformers import PretrainedConfig, PreTrainedModel, PreTrainedTokenizer +from transformers.activations import ACT2FN +from transformers.modeling_outputs import BaseModelOutput +from transformers.utils import logging + +from ...models import AutoencoderKL, UNet2DConditionModel, UNet2DModel, VQModel +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import is_torch_xla_available +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +class LDMTextToImagePipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using latent diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) model to encode and decode images to and from latent representations. + bert ([`LDMBertModel`]): + Text-encoder model based on [`~transformers.BERT`]. + tokenizer ([`~transformers.BertTokenizer`]): + A `BertTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "bert->unet->vqvae" + + def __init__( + self, + vqvae: VQModel | AutoencoderKL, + bert: PreTrainedModel, + tokenizer: PreTrainedTokenizer, + unet: UNet2DModel | UNet2DConditionModel, + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler, + ): + super().__init__() + self.register_modules(vqvae=vqvae, bert=bert, tokenizer=tokenizer, unet=unet, scheduler=scheduler) + self.vae_scale_factor = 2 ** (len(self.vqvae.config.block_out_channels) - 1) + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + height: int | None = None, + width: int | None = None, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 1.0, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + **kwargs, + ) -> tuple | ImagePipelineOutput: + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 1.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple. + + Example: + + ```py + >>> from diffusers import DiffusionPipeline + + >>> # load model and scheduler + >>> ldm = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256") + + >>> # run pipeline in inference (sample random noise and denoise) + >>> prompt = "A painting of a squirrel eating a burger" + >>> images = ldm([prompt], num_inference_steps=50, eta=0.3, guidance_scale=6).images + + >>> # save images + >>> for idx, image in enumerate(images): + ... image.save(f"squirrel-{idx}.png") + ``` + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + # get unconditional embeddings for classifier free guidance + if guidance_scale != 1.0: + uncond_input = self.tokenizer( + [""] * batch_size, padding="max_length", max_length=77, truncation=True, return_tensors="pt" + ) + negative_prompt_embeds = self.bert(uncond_input.input_ids.to(self._execution_device))[0] + + # get prompt text embeddings + text_input = self.tokenizer(prompt, padding="max_length", max_length=77, truncation=True, return_tensors="pt") + prompt_embeds = self.bert(text_input.input_ids.to(self._execution_device))[0] + + # get the initial random noise unless the user supplied it + latents_shape = (batch_size, self.unet.config.in_channels, height // 8, width // 8) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor( + latents_shape, generator=generator, device=self._execution_device, dtype=prompt_embeds.dtype + ) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + latents = latents.to(self._execution_device) + + self.scheduler.set_timesteps(num_inference_steps) + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + + extra_kwargs = {} + if accepts_eta: + extra_kwargs["eta"] = eta + + for t in self.progress_bar(self.scheduler.timesteps): + if guidance_scale == 1.0: + # guidance_scale of 1 means no guidance + latents_input = latents + context = prompt_embeds + else: + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + latents_input = torch.cat([latents] * 2) + context = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # predict the noise residual + noise_pred = self.unet(latents_input, t, encoder_hidden_states=context).sample + # perform guidance + if guidance_scale != 1.0: + noise_pred_uncond, noise_prediction_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + # scale and decode the image latents with vae + latents = 1 / self.vqvae.config.scaling_factor * latents + image = self.vqvae.decode(latents).sample + + image = (image / 2 + 0.5).clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).numpy() + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) + + +################################################################################ +# Code for the text transformer model +################################################################################ +""" PyTorch LDMBERT model.""" + + +logger = logging.get_logger(__name__) + +LDMBERT_PRETRAINED_MODEL_ARCHIVE_list = [ + "ldm-bert", + # See all LDMBert models at https://huggingface.co/models?filter=ldmbert +] + + +LDMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = { + "ldm-bert": "https://huggingface.co/valhalla/ldm-bert/blob/main/config.json", +} + + +""" LDMBERT model configuration""" + + +class LDMBertConfig(PretrainedConfig): + model_type = "ldmbert" + keys_to_ignore_at_inference = ["past_key_values"] + attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"} + + def __init__( + self, + vocab_size=30522, + max_position_embeddings=77, + encoder_layers=32, + encoder_ffn_dim=5120, + encoder_attention_heads=8, + head_dim=64, + encoder_layerdrop=0.0, + activation_function="gelu", + d_model=1280, + dropout=0.1, + attention_dropout=0.0, + activation_dropout=0.0, + init_std=0.02, + classifier_dropout=0.0, + scale_embedding=False, + use_cache=True, + pad_token_id=0, + **kwargs, + ): + self.vocab_size = vocab_size + self.max_position_embeddings = max_position_embeddings + self.d_model = d_model + self.encoder_ffn_dim = encoder_ffn_dim + self.encoder_layers = encoder_layers + self.encoder_attention_heads = encoder_attention_heads + self.head_dim = head_dim + self.dropout = dropout + self.attention_dropout = attention_dropout + self.activation_dropout = activation_dropout + self.activation_function = activation_function + self.init_std = init_std + self.encoder_layerdrop = encoder_layerdrop + self.classifier_dropout = classifier_dropout + self.use_cache = use_cache + self.num_hidden_layers = encoder_layers + self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True + + super().__init__(pad_token_id=pad_token_id, **kwargs) + + +def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: int | None = None): + """ + Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. + """ + bsz, src_len = mask.size() + tgt_len = tgt_len if tgt_len is not None else src_len + + expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype) + + inverted_mask = 1.0 - expanded_mask + + return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min) + + +# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->LDMBert +class LDMBertAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__( + self, + embed_dim: int, + num_heads: int, + head_dim: int, + dropout: float = 0.0, + is_decoder: bool = False, + bias: bool = False, + ): + super().__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = head_dim + self.inner_dim = head_dim * num_heads + + self.scaling = self.head_dim**-0.5 + self.is_decoder = is_decoder + + self.k_proj = nn.Linear(embed_dim, self.inner_dim, bias=bias) + self.v_proj = nn.Linear(embed_dim, self.inner_dim, bias=bias) + self.q_proj = nn.Linear(embed_dim, self.inner_dim, bias=bias) + self.out_proj = nn.Linear(self.inner_dim, embed_dim) + + def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): + return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() + + def forward( + self, + hidden_states: torch.Tensor, + key_value_states: torch.Tensor | None = None, + past_key_value: tuple[torch.Tensor] | None = None, + attention_mask: torch.Tensor | None = None, + layer_head_mask: torch.Tensor | None = None, + output_attentions: bool = False, + ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]: + """Input shape: Batch x Time x Channel""" + + # if key_value_states are provided this layer is used as a cross-attention layer + # for the decoder + is_cross_attention = key_value_states is not None + + bsz, tgt_len, _ = hidden_states.size() + + # get query proj + query_states = self.q_proj(hidden_states) * self.scaling + # get key, value proj + if is_cross_attention and past_key_value is not None: + # reuse k,v, cross_attentions + key_states = past_key_value[0] + value_states = past_key_value[1] + elif is_cross_attention: + # cross_attentions + key_states = self._shape(self.k_proj(key_value_states), -1, bsz) + value_states = self._shape(self.v_proj(key_value_states), -1, bsz) + elif past_key_value is not None: + # reuse k, v, self_attention + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + key_states = torch.cat([past_key_value[0], key_states], dim=2) + value_states = torch.cat([past_key_value[1], value_states], dim=2) + else: + # self_attention + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + + if self.is_decoder: + # if cross_attention save tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. + # Further calls to cross_attention layer can then reuse all cross-attention + # key/value_states (first "if" case) + # if uni-directional self-attention (decoder) save tuple(torch.Tensor, torch.Tensor) of + # all previous decoder key/value_states. Further calls to uni-directional self-attention + # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case) + # if encoder bi-directional self-attention `past_key_value` is always `None` + past_key_value = (key_states, value_states) + + proj_shape = (bsz * self.num_heads, -1, self.head_dim) + query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape) + key_states = key_states.view(*proj_shape) + value_states = value_states.view(*proj_shape) + + src_len = key_states.size(1) + attn_weights = torch.bmm(query_states, key_states.transpose(1, 2)) + + if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len): + raise ValueError( + f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is" + f" {attn_weights.size()}" + ) + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, tgt_len, src_len): + raise ValueError( + f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" + ) + attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + attn_weights = nn.functional.softmax(attn_weights, dim=-1) + + if layer_head_mask is not None: + if layer_head_mask.size() != (self.num_heads,): + raise ValueError( + f"Head mask for a single layer should be of size {(self.num_heads,)}, but is" + f" {layer_head_mask.size()}" + ) + attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + if output_attentions: + # this operation is a bit awkward, but it's required to + # make sure that attn_weights keeps its gradient. + # In order to do so, attn_weights have to be reshaped + # twice and have to be reused in the following + attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len) + else: + attn_weights_reshaped = None + + attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training) + + attn_output = torch.bmm(attn_probs, value_states) + + if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim): + raise ValueError( + f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is" + f" {attn_output.size()}" + ) + + attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim) + attn_output = attn_output.transpose(1, 2) + + # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be + # partitioned across GPUs when using tensor-parallelism. + attn_output = attn_output.reshape(bsz, tgt_len, self.inner_dim) + + attn_output = self.out_proj(attn_output) + + return attn_output, attn_weights_reshaped, past_key_value + + +class LDMBertEncoderLayer(nn.Module): + def __init__(self, config: LDMBertConfig): + super().__init__() + self.embed_dim = config.d_model + self.self_attn = LDMBertAttention( + embed_dim=self.embed_dim, + num_heads=config.encoder_attention_heads, + head_dim=config.head_dim, + dropout=config.attention_dropout, + ) + self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) + self.dropout = config.dropout + self.activation_fn = ACT2FN[config.activation_function] + self.activation_dropout = config.activation_dropout + self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim) + self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim) + self.final_layer_norm = nn.LayerNorm(self.embed_dim) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: torch.Tensor, + layer_head_mask: torch.Tensor, + output_attentions: bool | None = False, + ) -> tuple[torch.Tensor, torch.Tensor | None]: + """ + Args: + hidden_states (`torch.Tensor`): input to the layer of shape `(seq_len, batch, embed_dim)` + attention_mask (`torch.Tensor`): attention mask of size + `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. + layer_head_mask (`torch.Tensor`): mask for attention heads in a given layer of size + `(encoder_attention_heads,)`. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + """ + residual = hidden_states + hidden_states = self.self_attn_layer_norm(hidden_states) + hidden_states, attn_weights, _ = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + layer_head_mask=layer_head_mask, + output_attentions=output_attentions, + ) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + hidden_states = residual + hidden_states + + residual = hidden_states + hidden_states = self.final_layer_norm(hidden_states) + hidden_states = self.activation_fn(self.fc1(hidden_states)) + hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training) + hidden_states = self.fc2(hidden_states) + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + hidden_states = residual + hidden_states + + if hidden_states.dtype == torch.float16 and ( + torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any() + ): + clamp_value = torch.finfo(hidden_states.dtype).max - 1000 + hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) + + outputs = (hidden_states,) + + if output_attentions: + outputs += (attn_weights,) + + return outputs + + +# Copied from transformers.models.bart.modeling_bart.BartPretrainedModel with Bart->LDMBert +class LDMBertPreTrainedModel(PreTrainedModel): + config_class = LDMBertConfig + base_model_prefix = "model" + _supports_gradient_checkpointing = True + _keys_to_ignore_on_load_unexpected = [r"encoder\.version", r"decoder\.version"] + + def _init_weights(self, module): + std = self.config.init_std + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + + @property + def dummy_inputs(self): + pad_token = self.config.pad_token_id + input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device) + dummy_inputs = { + "attention_mask": input_ids.ne(pad_token), + "input_ids": input_ids, + } + return dummy_inputs + + +class LDMBertEncoder(LDMBertPreTrainedModel): + """ + Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a + [`LDMBertEncoderLayer`]. + + Args: + config: LDMBertConfig + embed_tokens (nn.Embedding): output embedding + """ + + def __init__(self, config: LDMBertConfig): + super().__init__(config) + + self.dropout = config.dropout + + embed_dim = config.d_model + self.padding_idx = config.pad_token_id + self.max_source_positions = config.max_position_embeddings + + self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim) + self.embed_positions = nn.Embedding(config.max_position_embeddings, embed_dim) + self.layers = nn.ModuleList([LDMBertEncoderLayer(config) for _ in range(config.encoder_layers)]) + self.layer_norm = nn.LayerNorm(embed_dim) + + self.gradient_checkpointing = False + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: torch.Tensor | None = None, + position_ids: torch.LongTensor | None = None, + head_mask: torch.Tensor | None = None, + inputs_embeds: torch.Tensor | None = None, + output_attentions: bool | None = None, + output_hidden_states: bool | None = None, + return_dict: bool | None = None, + ) -> tuple | BaseModelOutput: + r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you + provide it. + + Indices can be obtained using [`BartTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): + Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. + This is useful if you want more control over how to convert `input_ids` indices into associated vectors + than the model's internal embedding lookup matrix. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors + for more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.BaseModelOutput`] instead of a plain tuple. + """ + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + elif input_ids is not None: + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + else: + raise ValueError("You have to specify either input_ids or inputs_embeds") + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + seq_len = input_shape[1] + if position_ids is None: + position_ids = torch.arange(seq_len, dtype=torch.long, device=inputs_embeds.device).expand((1, -1)) + embed_pos = self.embed_positions(position_ids) + + hidden_states = inputs_embeds + embed_pos + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + + # expand attention_mask + if attention_mask is not None: + # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] + attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype) + + encoder_states = () if output_hidden_states else None + all_attentions = () if output_attentions else None + + # check if head_mask has a correct number of layers specified if desired + if head_mask is not None: + if head_mask.size()[0] != (len(self.layers)): + raise ValueError( + f"The head_mask should be specified for {len(self.layers)} layers, but it is for" + f" {head_mask.size()[0]}." + ) + + for idx, encoder_layer in enumerate(self.layers): + if output_hidden_states: + encoder_states = encoder_states + (hidden_states,) + if torch.is_grad_enabled() and self.gradient_checkpointing: + layer_outputs = self._gradient_checkpointing_func( + encoder_layer, + hidden_states, + attention_mask, + (head_mask[idx] if head_mask is not None else None), + ) + else: + layer_outputs = encoder_layer( + hidden_states, + attention_mask, + layer_head_mask=(head_mask[idx] if head_mask is not None else None), + output_attentions=output_attentions, + ) + + hidden_states = layer_outputs[0] + + if output_attentions: + all_attentions = all_attentions + (layer_outputs[1],) + + hidden_states = self.layer_norm(hidden_states) + + if output_hidden_states: + encoder_states = encoder_states + (hidden_states,) + + if not return_dict: + return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) + return BaseModelOutput( + last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions + ) + + +class LDMBertModel(LDMBertPreTrainedModel): + _no_split_modules = [] + + def __init__(self, config: LDMBertConfig): + super().__init__(config) + self.model = LDMBertEncoder(config) + self.to_logits = nn.Linear(config.hidden_size, config.vocab_size) + self.post_init() + + def forward( + self, + input_ids=None, + attention_mask=None, + position_ids=None, + head_mask=None, + inputs_embeds=None, + output_attentions=None, + output_hidden_states=None, + return_dict=None, + ): + outputs = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + return outputs diff --git a/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py b/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py new file mode 100644 index 0000000000000000000000000000000000000000..18cb8274f9b5fefd481346c25297d069e60984aa --- /dev/null +++ b/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py @@ -0,0 +1,196 @@ +import inspect + +import numpy as np +import PIL.Image +import torch + +from ...models import UNet2DModel, VQModel +from ...schedulers import ( + DDIMScheduler, + DPMSolverMultistepScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + LMSDiscreteScheduler, + PNDMScheduler, +) +from ...utils import PIL_INTERPOLATION, is_torch_xla_available +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +def preprocess(image): + w, h = image.size + w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +class LDMSuperResolutionPipeline(DiffusionPipeline): + r""" + A pipeline for image super-resolution using latent diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Parameters: + vqvae ([`VQModel`]): + Vector-quantized (VQ) model to encode and decode images to and from latent representations. + unet ([`UNet2DModel`]): + A `UNet2DModel` to denoise the encoded image. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`], + [`EulerAncestralDiscreteScheduler`], [`DPMSolverMultistepScheduler`], or [`PNDMScheduler`]. + """ + + def __init__( + self, + vqvae: VQModel, + unet: UNet2DModel, + scheduler: DDIMScheduler + | PNDMScheduler + | LMSDiscreteScheduler + | EulerDiscreteScheduler + | EulerAncestralDiscreteScheduler + | DPMSolverMultistepScheduler, + ): + super().__init__() + self.register_modules(vqvae=vqvae, unet=unet, scheduler=scheduler) + + @torch.no_grad() + def __call__( + self, + image: torch.Tensor | PIL.Image.Image = None, + batch_size: int | None = 1, + num_inference_steps: int | None = 100, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + ) -> tuple | ImagePipelineOutput: + r""" + The call function to the pipeline for generation. + + Args: + image (`torch.Tensor` or `PIL.Image.Image`): + `Image` or tensor representing an image batch to be used as the starting point for the process. + batch_size (`int`, *optional*, defaults to 1): + Number of images to generate. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple. + + Example: + + ```py + >>> import requests + >>> from PIL import Image + >>> from io import BytesIO + >>> from diffusers import LDMSuperResolutionPipeline + >>> import torch + + >>> # load model and scheduler + >>> pipeline = LDMSuperResolutionPipeline.from_pretrained("CompVis/ldm-super-resolution-4x-openimages") + >>> pipeline = pipeline.to("cuda") + + >>> # let's download an image + >>> url = ( + ... "https://user-images.githubusercontent.com/38061659/199705896-b48e17b8-b231-47cd-a270-4ffa5a93fa3e.png" + ... ) + >>> response = requests.get(url) + >>> low_res_img = Image.open(BytesIO(response.content)).convert("RGB") + >>> low_res_img = low_res_img.resize((128, 128)) + + >>> # run pipeline in inference (sample random noise and denoise) + >>> upscaled_image = pipeline(low_res_img, num_inference_steps=100, eta=1).images[0] + >>> # save image + >>> upscaled_image.save("ldm_generated_image.png") + ``` + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images + """ + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, torch.Tensor): + batch_size = image.shape[0] + else: + raise ValueError(f"`image` has to be of type `PIL.Image.Image` or `torch.Tensor` but is {type(image)}") + + if isinstance(image, PIL.Image.Image): + image = preprocess(image) + + height, width = image.shape[-2:] + + # in_channels should be 6: 3 for latents, 3 for low resolution image + latents_shape = (batch_size, self.unet.config.in_channels // 2, height, width) + latents_dtype = next(self.unet.parameters()).dtype + + latents = randn_tensor(latents_shape, generator=generator, device=self.device, dtype=latents_dtype) + + image = image.to(device=self.device, dtype=latents_dtype) + + # set timesteps and move to the correct device + self.scheduler.set_timesteps(num_inference_steps, device=self.device) + timesteps_tensor = self.scheduler.timesteps + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature. + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_kwargs = {} + if accepts_eta: + extra_kwargs["eta"] = eta + + for t in self.progress_bar(timesteps_tensor): + # concat latents and low resolution image in the channel dimension. + latents_input = torch.cat([latents, image], dim=1) + latents_input = self.scheduler.scale_model_input(latents_input, t) + # predict the noise residual + noise_pred = self.unet(latents_input, t).sample + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + # decode the image latents with the VQVAE + image = self.vqvae.decode(latents).sample + image = torch.clamp(image, -1.0, 1.0) + image = image / 2 + 0.5 + image = image.cpu().permute(0, 2, 3, 1).numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/latte/__init__.py b/diffusers/pipelines/latte/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4296b42e125303c0e036b9f2deecc36bdf959de3 --- /dev/null +++ b/diffusers/pipelines/latte/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_latte"] = ["LattePipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_latte import LattePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/latte/pipeline_latte.py b/diffusers/pipelines/latte/pipeline_latte.py new file mode 100644 index 0000000000000000000000000000000000000000..7bc7b4aa915eb33dda927dfc5c3dd4703c5b98a7 --- /dev/null +++ b/diffusers/pipelines/latte/pipeline_latte.py @@ -0,0 +1,914 @@ +# Copyright 2025 the Latte Team and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +from dataclasses import dataclass +from typing import Callable + +import torch +from transformers import T5EncoderModel, T5Tokenizer + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...models import AutoencoderKL, LatteTransformer3DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + BACKENDS_MAPPING, + BaseOutput, + deprecate, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import is_compiled_module, randn_tensor +from ...video_processor import VideoProcessor + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LattePipeline + >>> from diffusers.utils import export_to_gif + + >>> # You can replace the checkpoint id with "maxin-cn/Latte-1" too. + >>> pipe = LattePipeline.from_pretrained("maxin-cn/Latte-1", torch_dtype=torch.float16) + >>> # Enable memory optimizations. + >>> pipe.enable_model_cpu_offload() + + >>> prompt = "A small cactus with a happy face in the Sahara desert." + >>> videos = pipe(prompt).frames[0] + >>> export_to_gif(videos, "latte.gif") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +@dataclass +class LattePipelineOutput(BaseOutput): + frames: torch.Tensor + + +class LattePipeline(DiffusionPipeline): + r""" + Pipeline for text-to-video generation using Latte. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder ([`T5EncoderModel`]): + Frozen text-encoder. Latte uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant. + tokenizer (`T5Tokenizer`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + transformer ([`LatteTransformer3DModel`]): + A text conditioned `LatteTransformer3DModel` to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `transformer` to denoise the encoded video latents. + """ + + bad_punct_regex = re.compile(r"[#®•©™&@·º½¾¿¡§~\)\(\]\[\}\{\|\\/\\*]{1,}") + + _optional_components = ["tokenizer", "text_encoder"] + model_cpu_offload_seq = "text_encoder->transformer->vae" + + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + ] + + def __init__( + self, + tokenizer: T5Tokenizer, + text_encoder: T5EncoderModel, + vae: AutoencoderKL, + transformer: LatteTransformer3DModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor) + + # Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/utils.py + def mask_text_embeddings(self, emb, mask): + if emb.shape[0] == 1: + keep_index = mask.sum().item() + return emb[:, :, :keep_index, :], keep_index # 1, 120, 4096 -> 1 7 4096 + else: + masked_feature = emb * mask[:, None, :, None] # 1 120 4096 + return masked_feature, emb.shape[2] + + # Adapted from diffusers.pipelines.deepfloyd_if.pipeline_if.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + clean_caption: bool = False, + mask_feature: bool = True, + dtype=None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the video generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + Latte, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of video that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For Latte, it's should be the embeddings of the "" string. + clean_caption (bool, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + mask_feature: (bool, defaults to `True`): + If `True`, the function will mask the text embeddings. + """ + embeds_initially_provided = prompt_embeds is not None and negative_prompt_embeds is not None + + if device is None: + device = self._execution_device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + max_length = 120 + if prompt_embeds is None: + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {max_length} tokens: {removed_text}" + ) + + attention_mask = text_inputs.attention_mask.to(device) + prompt_embeds_attention_mask = attention_mask + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds_attention_mask = torch.ones_like(prompt_embeds) + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + elif self.transformer is not None: + dtype = self.transformer.dtype + else: + dtype = None + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_embeds_attention_mask = prompt_embeds_attention_mask.view(bs_embed, -1) + prompt_embeds_attention_mask = prompt_embeds_attention_mask.repeat(num_images_per_prompt, 1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt + uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption) + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + attention_mask = uncond_input.attention_mask.to(device) + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + else: + negative_prompt_embeds = None + + # Perform additional masking. + if mask_feature and not embeds_initially_provided: + prompt_embeds = prompt_embeds.unsqueeze(1) + masked_prompt_embeds, keep_indices = self.mask_text_embeddings(prompt_embeds, prompt_embeds_attention_mask) + masked_prompt_embeds = masked_prompt_embeds.squeeze(1) + masked_negative_prompt_embeds = ( + negative_prompt_embeds[:, :keep_indices, :] if negative_prompt_embeds is not None else None + ) + + return masked_prompt_embeds, masked_negative_prompt_embeds + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt, + callback_on_step_end_tensor_inputs, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + # Copied from diffusers.pipelines.deprecated.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents + def prepare_latents( + self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None + ): + shape = ( + batch_size, + num_channels_latents, + num_frames, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 50, + timesteps: list[int] | None = None, + guidance_scale: float = 7.5, + num_images_per_prompt: int = 1, + video_length: int = 16, + height: int = 512, + width: int = 512, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + output_type: str = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + clean_caption: bool = True, + mask_feature: bool = True, + enable_temporal_attentions: bool = True, + decode_chunk_size: int = 14, + ) -> LattePipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the video generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality video at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` + timesteps are used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate videos that are closely linked to + the text `prompt`, usually at the expense of lower video quality. + video_length (`int`, *optional*, defaults to 16): + The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated video. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated video. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For Latte this negative prompt should be "". If not provided, + negative_prompt_embeds will be generated from `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate video. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable[[int, int], None]`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A callback function or a list of callback functions to be called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + A list of tensor inputs that should be passed to the callback function. If not defined, all tensor + inputs will be passed. + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + mask_feature (`bool` defaults to `True`): If set to `True`, the text embeddings will be masked. + enable_temporal_attentions (`bool`, *optional*, defaults to `True`): Whether to enable temporal attentions + decode_chunk_size (`int`, *optional*): + The number of frames to decode at a time. Higher chunk size leads to better temporal consistency at the + expense of more memory usage. By default, the decoder decodes all frames at once for maximal quality. + For lower memory usage, reduce `decode_chunk_size`. + + Examples: + + Returns: + [`~pipelines.latte.pipeline_latte.LattePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.latte.pipeline_latte.LattePipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images + """ + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default + decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else video_length + + # 1. Check inputs. Raise error if not correct + height = height or self.transformer.config.sample_size * self.vae_scale_factor + width = width or self.transformer.config.sample_size * self.vae_scale_factor + self.check_inputs( + prompt, + height, + width, + negative_prompt, + callback_on_step_end_tensor_inputs, + prompt_embeds, + negative_prompt_embeds, + ) + self._guidance_scale = guidance_scale + self._current_timestep = None + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clean_caption=clean_caption, + mask_feature=mask_feature, + ) + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps + ) + self._num_timesteps = len(timesteps) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + video_length, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + current_timestep = t + if not torch.is_tensor(current_timestep): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = latent_model_input.device.type == "mps" + is_npu = latent_model_input.device.type == "npu" + if isinstance(current_timestep, float): + dtype = torch.float32 if (is_mps or is_npu) else torch.float64 + else: + dtype = torch.int32 if (is_mps or is_npu) else torch.int64 + current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device) + elif len(current_timestep.shape) == 0: + current_timestep = current_timestep[None].to(latent_model_input.device) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + current_timestep = current_timestep.expand(latent_model_input.shape[0]) + + # predict noise model_output + noise_pred = self.transformer( + hidden_states=latent_model_input, + encoder_hidden_states=prompt_embeds, + timestep=current_timestep, + enable_temporal_attentions=enable_temporal_attentions, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # use learned sigma? + if not ( + hasattr(self.scheduler.config, "variance_type") + and self.scheduler.config.variance_type in ["learned", "learned_range"] + ): + noise_pred = noise_pred.chunk(2, dim=1)[0] + + # compute previous video: x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latents": + deprecation_message = ( + "Passing `output_type='latents'` is deprecated. Please pass `output_type='latent'` instead." + ) + deprecate("output_type_latents", "1.0.0", deprecation_message, standard_warn=False) + output_type = "latent" + + if not output_type == "latent": + video = self.decode_latents(latents, video_length, decode_chunk_size=decode_chunk_size) + video = self.video_processor.postprocess_video(video=video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return LattePipelineOutput(frames=video) + + # Similar to diffusers.pipelines.stable_video_diffusion.pipeline_stable_video_diffusion.decode_latents + def decode_latents(self, latents: torch.Tensor, video_length: int, decode_chunk_size: int = 14): + # [batch, channels, frames, height, width] -> [batch*frames, channels, height, width] + latents = latents.permute(0, 2, 1, 3, 4).flatten(0, 1) + + latents = 1 / self.vae.config.scaling_factor * latents + + forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward + accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys()) + + # decode decode_chunk_size frames at a time to avoid OOM + frames = [] + for i in range(0, latents.shape[0], decode_chunk_size): + num_frames_in = latents[i : i + decode_chunk_size].shape[0] + decode_kwargs = {} + if accepts_num_frames: + # we only pass num_frames_in if it's expected + decode_kwargs["num_frames"] = num_frames_in + + frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample + frames.append(frame) + frames = torch.cat(frames, dim=0) + + # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width] + frames = frames.reshape(-1, video_length, *frames.shape[1:]).permute(0, 2, 1, 3, 4) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + frames = frames.float() + return frames diff --git a/diffusers/pipelines/ledits_pp/__init__.py b/diffusers/pipelines/ledits_pp/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..aae3b1cb18ce96c7d79d3c44bc37cf2ae4db9720 --- /dev/null +++ b/diffusers/pipelines/ledits_pp/__init__.py @@ -0,0 +1,55 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_leditspp_stable_diffusion"] = ["LEditsPPPipelineStableDiffusion"] + _import_structure["pipeline_leditspp_stable_diffusion_xl"] = ["LEditsPPPipelineStableDiffusionXL"] + + _import_structure["pipeline_output"] = ["LEditsPPDiffusionPipelineOutput", "LEditsPPDiffusionPipelineOutput"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_leditspp_stable_diffusion import ( + LEditsPPDiffusionPipelineOutput, + LEditsPPInversionPipelineOutput, + LEditsPPPipelineStableDiffusion, + ) + from .pipeline_leditspp_stable_diffusion_xl import LEditsPPPipelineStableDiffusionXL + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py b/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..9048b56b01f4c7ce67ec3d8bf6ff8eb6e4cb79a6 --- /dev/null +++ b/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py @@ -0,0 +1,1536 @@ +import inspect +import math +from itertools import repeat +from typing import Any, Callable + +import torch +import torch.nn.functional as F +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention_processor import Attention, AttnProcessor +from ...models.lora import adjust_lora_scale_text_encoder +from ...pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from ...schedulers import DDIMScheduler, DPMSolverMultistepScheduler +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import LEditsPPDiffusionPipelineOutput, LEditsPPInversionPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + + >>> from diffusers import LEditsPPPipelineStableDiffusion + >>> from diffusers.utils import load_image + + >>> pipe = LEditsPPPipelineStableDiffusion.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe.vae.enable_tiling() + >>> pipe = pipe.to("cuda") + + >>> img_url = "https://www.aiml.informatik.tu-darmstadt.de/people/mbrack/cherry_blossom.png" + >>> image = load_image(img_url).resize((512, 512)) + + >>> _ = pipe.invert(image=image, num_inversion_steps=50, skip=0.1) + + >>> edited_image = pipe( + ... editing_prompt=["cherry blossom"], edit_guidance_scale=10.0, edit_threshold=0.75 + ... ).images[0] + ``` +""" + + +# Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionAttendAndExcitePipeline.AttentionStore +class LeditsAttentionStore: + @staticmethod + def get_empty_store(): + return {"down_cross": [], "mid_cross": [], "up_cross": [], "down_self": [], "mid_self": [], "up_self": []} + + def __call__(self, attn, is_cross: bool, place_in_unet: str, editing_prompts, PnP=False): + # attn.shape = batch_size * head_size, seq_len query, seq_len_key + if attn.shape[1] <= self.max_size: + bs = 1 + int(PnP) + editing_prompts + skip = 2 if PnP else 1 # skip PnP & unconditional + attn = torch.stack(attn.split(self.batch_size)).permute(1, 0, 2, 3) + source_batch_size = int(attn.shape[1] // bs) + self.forward(attn[:, skip * source_batch_size :], is_cross, place_in_unet) + + def forward(self, attn, is_cross: bool, place_in_unet: str): + key = f"{place_in_unet}_{'cross' if is_cross else 'self'}" + + self.step_store[key].append(attn) + + def between_steps(self, store_step=True): + if store_step: + if self.average: + if len(self.attention_store) == 0: + self.attention_store = self.step_store + else: + for key in self.attention_store: + for i in range(len(self.attention_store[key])): + self.attention_store[key][i] += self.step_store[key][i] + else: + if len(self.attention_store) == 0: + self.attention_store = [self.step_store] + else: + self.attention_store.append(self.step_store) + + self.cur_step += 1 + self.step_store = self.get_empty_store() + + def get_attention(self, step: int): + if self.average: + attention = { + key: [item / self.cur_step for item in self.attention_store[key]] for key in self.attention_store + } + else: + assert step is not None + attention = self.attention_store[step] + return attention + + def aggregate_attention( + self, attention_maps, prompts, res: int | tuple[int], from_where: list[str], is_cross: bool, select: int + ): + out = [[] for x in range(self.batch_size)] + if isinstance(res, int): + num_pixels = res**2 + resolution = (res, res) + else: + num_pixels = res[0] * res[1] + resolution = res[:2] + + for location in from_where: + for bs_item in attention_maps[f"{location}_{'cross' if is_cross else 'self'}"]: + for batch, item in enumerate(bs_item): + if item.shape[1] == num_pixels: + cross_maps = item.reshape(len(prompts), -1, *resolution, item.shape[-1])[select] + out[batch].append(cross_maps) + + out = torch.stack([torch.cat(x, dim=0) for x in out]) + # average over heads + out = out.sum(1) / out.shape[1] + return out + + def __init__(self, average: bool, batch_size=1, max_resolution=16, max_size: int = None): + self.step_store = self.get_empty_store() + self.attention_store = [] + self.cur_step = 0 + self.average = average + self.batch_size = batch_size + if max_size is None: + self.max_size = max_resolution**2 + elif max_size is not None and max_resolution is None: + self.max_size = max_size + else: + raise ValueError("Only allowed to set one of max_resolution or max_size") + + +# Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionAttendAndExcitePipeline.GaussianSmoothing +class LeditsGaussianSmoothing: + def __init__(self, device): + kernel_size = [3, 3] + sigma = [0.5, 0.5] + + # The gaussian kernel is the product of the gaussian function of each dimension. + kernel = 1 + meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size], indexing="ij") + for size, std, mgrid in zip(kernel_size, sigma, meshgrids): + mean = (size - 1) / 2 + kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2)) + + # Make sure sum of values in gaussian kernel equals 1. + kernel = kernel / torch.sum(kernel) + + # Reshape to depthwise convolutional weight + kernel = kernel.view(1, 1, *kernel.size()) + kernel = kernel.repeat(1, *[1] * (kernel.dim() - 1)) + + self.weight = kernel.to(device) + + def __call__(self, input): + """ + Arguments: + Apply gaussian filter to input. + input (torch.Tensor): Input to apply gaussian filter on. + Returns: + filtered (torch.Tensor): Filtered output. + """ + return F.conv2d(input, weight=self.weight.to(input.dtype)) + + +class LEDITSCrossAttnProcessor: + def __init__(self, attention_store, place_in_unet, pnp, editing_prompts): + self.attnstore = attention_store + self.place_in_unet = place_in_unet + self.editing_prompts = editing_prompts + self.pnp = pnp + + def __call__( + self, + attn: Attention, + hidden_states, + encoder_hidden_states, + attention_mask=None, + temb=None, + ): + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + self.attnstore( + attention_probs, + is_cross=True, + place_in_unet=self.place_in_unet, + editing_prompts=self.editing_prompts, + PnP=self.pnp, + ) + + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states / attn.rescale_output_factor + return hidden_states + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LEditsPPPipelineStableDiffusion( + DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin +): + """ + Pipeline for textual image editing using LEDits++ with Stable Diffusion. + + This model inherits from [`DiffusionPipeline`] and builds on the [`StableDiffusionPipeline`]. Check the superclass + documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular + device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer ([`~transformers.CLIPTokenizer`]): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of + [`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]. If any other scheduler is passed it will + automatically be set to [`DPMSolverMultistepScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler | DPMSolverMultistepScheduler, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if not isinstance(scheduler, DDIMScheduler) and not isinstance(scheduler, DPMSolverMultistepScheduler): + scheduler = DPMSolverMultistepScheduler.from_config( + scheduler.config, algorithm_type="sde-dpmsolver++", solver_order=2 + ) + logger.warning( + "This pipeline only supports DDIMScheduler and DPMSolverMultistepScheduler. " + "The scheduler has been changed to DPMSolverMultistepScheduler." + ) + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + self.inversion_steps = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, eta, generator=None): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs( + self, + negative_prompt=None, + editing_prompt_embeddings=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if editing_prompt_embeddings is not None and negative_prompt_embeds is not None: + if editing_prompt_embeddings.shape != negative_prompt_embeds.shape: + raise ValueError( + "`editing_prompt_embeddings` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `editing_prompt_embeddings` {editing_prompt_embeddings.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, latents): + # shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + + # if latents.shape != shape: + # raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_unet(self, attention_store, PnP: bool = False): + attn_procs = {} + for name in self.unet.attn_processors.keys(): + if name.startswith("mid_block"): + place_in_unet = "mid" + elif name.startswith("up_blocks"): + place_in_unet = "up" + elif name.startswith("down_blocks"): + place_in_unet = "down" + else: + continue + + if "attn2" in name and place_in_unet != "mid": + attn_procs[name] = LEDITSCrossAttnProcessor( + attention_store=attention_store, + place_in_unet=place_in_unet, + pnp=PnP, + editing_prompts=self.enabled_editing_prompts, + ) + else: + attn_procs[name] = AttnProcessor() + + self.unet.set_attn_processor(attn_procs) + + def encode_prompt( + self, + device, + num_images_per_prompt, + enable_edit_guidance, + negative_prompt=None, + editing_prompt=None, + negative_prompt_embeds: torch.Tensor | None = None, + editing_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + enable_edit_guidance (`bool`): + whether to perform any editing or reconstruct the input image instead + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + editing_prompt (`str` or `list[str]`, *optional*): + Editing prompt(s) to be encoded. If not defined, one has to pass `editing_prompt_embeds` instead. + editing_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + batch_size = self.batch_size + num_edit_tokens = None + + if negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but exoected" + f"{batch_size} based on the input images. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = negative_prompt_embeds.dtype + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + if enable_edit_guidance: + if editing_prompt_embeds is None: + # textual inversion: procecss multi-vector tokens if necessary + # if isinstance(self, TextualInversionLoaderMixin): + # prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + if isinstance(editing_prompt, str): + editing_prompt = [editing_prompt] + + max_length = negative_prompt_embeds.shape[1] + text_inputs = self.tokenizer( + [x for item in editing_prompt for x in repeat(item, batch_size)], + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + return_length=True, + ) + + num_edit_tokens = text_inputs.length - 2 # not counting startoftext and endoftext + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer( + [x for item in editing_prompt for x in repeat(item, batch_size)], + padding="longest", + return_tensors="pt", + ).input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if ( + hasattr(self.text_encoder.config, "use_attention_mask") + and self.text_encoder.config.use_attention_mask + ): + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + editing_prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + editing_prompt_embeds = editing_prompt_embeds[0] + else: + editing_prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + editing_prompt_embeds = editing_prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + editing_prompt_embeds = self.text_encoder.text_model.final_layer_norm(editing_prompt_embeds) + + editing_prompt_embeds = editing_prompt_embeds.to(dtype=negative_prompt_embeds.dtype, device=device) + + bs_embed_edit, seq_len, _ = editing_prompt_embeds.shape + editing_prompt_embeds = editing_prompt_embeds.to(dtype=negative_prompt_embeds.dtype, device=device) + editing_prompt_embeds = editing_prompt_embeds.repeat(1, num_images_per_prompt, 1) + editing_prompt_embeds = editing_prompt_embeds.view(bs_embed_edit * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return editing_prompt_embeds, negative_prompt_embeds, num_edit_tokens + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + negative_prompt: str | list[str] | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + editing_prompt: str | list[str] | None = None, + editing_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + reverse_editing_direction: bool | list[bool] | None = False, + edit_guidance_scale: float | list[float] | None = 5, + edit_warmup_steps: int | list[int] | None = 0, + edit_cooldown_steps: int | list[int] | None = None, + edit_threshold: float | list[float] | None = 0.9, + user_mask: torch.Tensor | None = None, + sem_guidance: list[torch.Tensor] | None = None, + use_cross_attn_mask: bool = False, + use_intersect_mask: bool = True, + attn_store_steps: list[int] | None = [], + store_averaged_over_steps: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for editing. The + [`~pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.invert`] method has to be called beforehand. Edits will + always be performed for the last inverted image(s). + + Args: + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] instead of a plain + tuple. + editing_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. The image is reconstructed by setting + `editing_prompt = None`. Guidance direction of prompt should be specified via + `reverse_editing_direction`. + editing_prompt_embeds (`torch.Tensor>`, *optional*): + Pre-computed embeddings to use for guiding the image generation. Guidance direction of embedding should + be specified via `reverse_editing_direction`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + reverse_editing_direction (`bool` or `list[bool]`, *optional*, defaults to `False`): + Whether the corresponding prompt in `editing_prompt` should be increased or decreased. + edit_guidance_scale (`float` or `list[float]`, *optional*, defaults to 5): + Guidance scale for guiding the image generation. If provided as list values should correspond to + `editing_prompt`. `edit_guidance_scale` is defined as `s_e` of equation 12 of [LEDITS++ + Paper](https://huggingface.co/papers/2301.12247). + edit_warmup_steps (`float` or `list[float]`, *optional*, defaults to 10): + Number of diffusion steps (for each prompt) for which guidance will not be applied. + edit_cooldown_steps (`float` or `list[float]`, *optional*, defaults to `None`): + Number of diffusion steps (for each prompt) after which guidance will no longer be applied. + edit_threshold (`float` or `list[float]`, *optional*, defaults to 0.9): + Masking threshold of guidance. Threshold should be proportional to the image region that is modified. + 'edit_threshold' is defined as 'λ' of equation 12 of [LEDITS++ + Paper](https://huggingface.co/papers/2301.12247). + user_mask (`torch.Tensor`, *optional*): + User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s + implicit masks do not meet user preferences. + sem_guidance (`list[torch.Tensor]`, *optional*): + list of pre-generated guidance vectors to be applied at generation. Length of the list has to + correspond to `num_inference_steps`. + use_cross_attn_mask (`bool`, defaults to `False`): + Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask + is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of [LEDITS++ + paper](https://huggingface.co/papers/2311.16711). + use_intersect_mask (`bool`, defaults to `True`): + Whether the masking term is calculated as intersection of cross-attention masks and masks derived from + the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate + are defined as 'M^2' of equation 12 of [LEDITS++ paper](https://huggingface.co/papers/2311.16711). + attn_store_steps (`list[int]`, *optional*): + Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes. + store_averaged_over_steps (`bool`, defaults to `True`): + Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If + False, attention maps for each step are stores separately. Just for visualization purposes. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When + returning a tuple, the first element is a list with the generated images, and the second element is a list + of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw) + content, according to the `safety_checker`. + """ + + if self.inversion_steps is None: + raise ValueError( + "You need to invert an input image first before calling the pipeline. The `invert` method has to be called beforehand. Edits will always be performed for the last inverted image(s)." + ) + + eta = self.eta + num_images_per_prompt = 1 + latents = self.init_latents + + zs = self.zs + self.scheduler.set_timesteps(len(self.scheduler.timesteps)) + + if use_intersect_mask: + use_cross_attn_mask = True + + if use_cross_attn_mask: + self.smoothing = LeditsGaussianSmoothing(self.device) + + if user_mask is not None: + user_mask = user_mask.to(self.device) + + org_prompt = "" + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + negative_prompt, + editing_prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + + # 2. Define call parameters + batch_size = self.batch_size + + if editing_prompt: + enable_edit_guidance = True + if isinstance(editing_prompt, str): + editing_prompt = [editing_prompt] + self.enabled_editing_prompts = len(editing_prompt) + elif editing_prompt_embeds is not None: + enable_edit_guidance = True + self.enabled_editing_prompts = editing_prompt_embeds.shape[0] + else: + self.enabled_editing_prompts = 0 + enable_edit_guidance = False + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + edit_concepts, uncond_embeddings, num_edit_tokens = self.encode_prompt( + editing_prompt=editing_prompt, + device=self.device, + num_images_per_prompt=num_images_per_prompt, + enable_edit_guidance=enable_edit_guidance, + negative_prompt=negative_prompt, + editing_prompt_embeds=editing_prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if enable_edit_guidance: + text_embeddings = torch.cat([uncond_embeddings, edit_concepts]) + self.text_cross_attention_maps = [editing_prompt] if isinstance(editing_prompt, str) else editing_prompt + else: + text_embeddings = torch.cat([uncond_embeddings]) + + # 4. Prepare timesteps + # self.scheduler.set_timesteps(num_inference_steps, device=self.device) + timesteps = self.inversion_steps + t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs.shape[0] :])} + + if use_cross_attn_mask: + self.attention_store = LeditsAttentionStore( + average=store_averaged_over_steps, + batch_size=batch_size, + max_size=(latents.shape[-2] / 4.0) * (latents.shape[-1] / 4.0), + max_resolution=None, + ) + self.prepare_unet(self.attention_store, PnP=False) + resolution = latents.shape[-2:] + att_res = (int(resolution[0] / 4), int(resolution[1] / 4)) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + None, + None, + text_embeddings.dtype, + self.device, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(eta) + + self.sem_guidance = None + self.activation_mask = None + + # 7. Denoising loop + num_warmup_steps = 0 + with self.progress_bar(total=len(timesteps)) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + + if enable_edit_guidance: + latent_model_input = torch.cat([latents] * (1 + self.enabled_editing_prompts)) + else: + latent_model_input = latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + text_embed_input = text_embeddings + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input).sample + + noise_pred_out = noise_pred.chunk(1 + self.enabled_editing_prompts) # [b,4, 64, 64] + noise_pred_uncond = noise_pred_out[0] + noise_pred_edit_concepts = noise_pred_out[1:] + + noise_guidance_edit = torch.zeros( + noise_pred_uncond.shape, + device=self.device, + dtype=noise_pred_uncond.dtype, + ) + + if sem_guidance is not None and len(sem_guidance) > i: + noise_guidance_edit += sem_guidance[i].to(self.device) + + elif enable_edit_guidance: + if self.activation_mask is None: + self.activation_mask = torch.zeros( + (len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape) + ) + + if self.sem_guidance is None: + self.sem_guidance = torch.zeros((len(timesteps), *noise_pred_uncond.shape)) + + for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts): + if isinstance(edit_warmup_steps, list): + edit_warmup_steps_c = edit_warmup_steps[c] + else: + edit_warmup_steps_c = edit_warmup_steps + if i < edit_warmup_steps_c: + continue + + if isinstance(edit_guidance_scale, list): + edit_guidance_scale_c = edit_guidance_scale[c] + else: + edit_guidance_scale_c = edit_guidance_scale + + if isinstance(edit_threshold, list): + edit_threshold_c = edit_threshold[c] + else: + edit_threshold_c = edit_threshold + if isinstance(reverse_editing_direction, list): + reverse_editing_direction_c = reverse_editing_direction[c] + else: + reverse_editing_direction_c = reverse_editing_direction + + if isinstance(edit_cooldown_steps, list): + edit_cooldown_steps_c = edit_cooldown_steps[c] + elif edit_cooldown_steps is None: + edit_cooldown_steps_c = i + 1 + else: + edit_cooldown_steps_c = edit_cooldown_steps + + if i >= edit_cooldown_steps_c: + continue + + noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond + + if reverse_editing_direction_c: + noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1 + + noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c + + if user_mask is not None: + noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask + + if use_cross_attn_mask: + out = self.attention_store.aggregate_attention( + attention_maps=self.attention_store.step_store, + prompts=self.text_cross_attention_maps, + res=att_res, + from_where=["up", "down"], + is_cross=True, + select=self.text_cross_attention_maps.index(editing_prompt[c]), + ) + attn_map = out[:, :, :, 1 : 1 + num_edit_tokens[c]] # 0 -> startoftext + + # average over all tokens + if attn_map.shape[3] != num_edit_tokens[c]: + raise ValueError( + f"Incorrect shape of attention_map. Expected size {num_edit_tokens[c]}, but found {attn_map.shape[3]}!" + ) + + attn_map = torch.sum(attn_map, dim=3) + + # gaussian_smoothing + attn_map = F.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect") + attn_map = self.smoothing(attn_map).squeeze(1) + + # torch.quantile function expects float32 + if attn_map.dtype == torch.float32: + tmp = torch.quantile(attn_map.flatten(start_dim=1), edit_threshold_c, dim=1) + else: + tmp = torch.quantile( + attn_map.flatten(start_dim=1).to(torch.float32), edit_threshold_c, dim=1 + ).to(attn_map.dtype) + attn_mask = torch.where( + attn_map >= tmp.unsqueeze(1).unsqueeze(1).repeat(1, *att_res), 1.0, 0.0 + ) + + # resolution must match latent space dimension + attn_mask = F.interpolate( + attn_mask.unsqueeze(1), + noise_guidance_edit_tmp.shape[-2:], # 64,64 + ).repeat(1, 4, 1, 1) + self.activation_mask[i, c] = attn_mask.detach().cpu() + if not use_intersect_mask: + noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask + + if use_intersect_mask: + if t <= 800: + noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp) + noise_guidance_edit_tmp_quantile = torch.sum( + noise_guidance_edit_tmp_quantile, dim=1, keepdim=True + ) + noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat( + 1, self.unet.config.in_channels, 1, 1 + ) + + # torch.quantile function expects float32 + if noise_guidance_edit_tmp_quantile.dtype == torch.float32: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2), + edit_threshold_c, + dim=2, + keepdim=False, + ) + else: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32), + edit_threshold_c, + dim=2, + keepdim=False, + ).to(noise_guidance_edit_tmp_quantile.dtype) + + intersect_mask = ( + torch.where( + noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], + torch.ones_like(noise_guidance_edit_tmp), + torch.zeros_like(noise_guidance_edit_tmp), + ) + * attn_mask + ) + + self.activation_mask[i, c] = intersect_mask.detach().cpu() + + noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask + + else: + # print(f"only attention mask for step {i}") + noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask + + elif not use_cross_attn_mask: + # calculate quantile + noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp) + noise_guidance_edit_tmp_quantile = torch.sum( + noise_guidance_edit_tmp_quantile, dim=1, keepdim=True + ) + noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1) + + # torch.quantile function expects float32 + if noise_guidance_edit_tmp_quantile.dtype == torch.float32: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2), + edit_threshold_c, + dim=2, + keepdim=False, + ) + else: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32), + edit_threshold_c, + dim=2, + keepdim=False, + ).to(noise_guidance_edit_tmp_quantile.dtype) + + self.activation_mask[i, c] = ( + torch.where( + noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], + torch.ones_like(noise_guidance_edit_tmp), + torch.zeros_like(noise_guidance_edit_tmp), + ) + .detach() + .cpu() + ) + + noise_guidance_edit_tmp = torch.where( + noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], + noise_guidance_edit_tmp, + torch.zeros_like(noise_guidance_edit_tmp), + ) + + noise_guidance_edit += noise_guidance_edit_tmp + + self.sem_guidance[i] = noise_guidance_edit.detach().cpu() + + noise_pred = noise_pred_uncond + noise_guidance_edit + + if enable_edit_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg( + noise_pred, + noise_pred_edit_concepts.mean(dim=0, keepdim=False), + guidance_rescale=self.guidance_rescale, + ) + + idx = t_to_idx[int(t)] + latents = self.scheduler.step( + noise_pred, t, latents, variance_noise=zs[idx], **extra_step_kwargs + ).prev_sample + + # step callback + if use_cross_attn_mask: + store_step = i in attn_store_steps + self.attention_store.between_steps(store_step) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + # prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post-processing + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, self.device, text_embeddings.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return LEditsPPDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + @torch.no_grad() + def invert( + self, + image: PipelineImageInput, + source_prompt: str = "", + source_guidance_scale: float = 3.5, + num_inversion_steps: int = 30, + skip: float = 0.15, + generator: torch.Generator | None = None, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + height: int | None = None, + width: int | None = None, + resize_mode: str | None = "default", + crops_coords: tuple[int, int, int, int] | None = None, + ): + r""" + The function to the pipeline for image inversion as described by the [LEDITS++ + Paper](https://huggingface.co/papers/2301.12247). If the scheduler is set to [`~schedulers.DDIMScheduler`] the + inversion proposed by [edit-friendly DPDM](https://huggingface.co/papers/2304.06140) will be performed instead. + + Args: + image (`PipelineImageInput`): + Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect + ratio. + source_prompt (`str`, defaults to `""`): + Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled + if the `source_prompt` is `""`. + source_guidance_scale (`float`, defaults to `3.5`): + Strength of guidance during inversion. + num_inversion_steps (`int`, defaults to `30`): + Number of total performed inversion steps after discarding the initial `skip` steps. + skip (`float`, defaults to `0.15`): + Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values + will lead to stronger changes to the input image. `skip` has to be between `0` and `1`. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make inversion + deterministic. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + height (`int`, *optional*, defaults to `None`): + The height in preprocessed image. If `None`, will use the `get_default_height_width()` to get default + height. + width (`int`, *optional*`, defaults to `None`): + The width in preprocessed. If `None`, will use get_default_height_width()` to get the default width. + resize_mode (`str`, *optional*, defaults to `default`): + The resize mode, can be one of `default` or `fill`. If `default`, will resize the image to fit within + the specified width and height, and it may not maintaining the original aspect ratio. If `fill`, will + resize the image to fit within the specified width and height, maintaining the aspect ratio, and then + center the image within the dimensions, filling empty with data from image. If `crop`, will resize the + image to fit within the specified width and height, maintaining the aspect ratio, and then center the + image within the dimensions, cropping the excess. Note that resize_mode `fill` and `crop` are only + supported for PIL image input. + crops_coords (`list[tuple[int, int, int, int]]`, *optional*, defaults to `None`): + The crop coordinates for each image in the batch. If `None`, will not crop the image. + + Returns: + [`~pipelines.ledits_pp.LEditsPPInversionPipelineOutput`]: Output will contain the resized input image(s) + and respective VAE reconstruction(s). + """ + if height is not None and height % 32 != 0 or width is not None and width % 32 != 0: + raise ValueError("height and width must be a factor of 32.") + # Reset attn processor, we do not want to store attn maps during inversion + self.unet.set_attn_processor(AttnProcessor()) + + self.eta = 1.0 + + self.scheduler.config.timestep_spacing = "leading" + self.scheduler.set_timesteps(int(num_inversion_steps * (1 + skip))) + self.inversion_steps = self.scheduler.timesteps[-num_inversion_steps:] + timesteps = self.inversion_steps + + # 1. encode image + x0, resized = self.encode_image( + image, + dtype=self.text_encoder.dtype, + height=height, + width=width, + resize_mode=resize_mode, + crops_coords=crops_coords, + ) + self.batch_size = x0.shape[0] + + # autoencoder reconstruction + image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0] + image_rec = self.image_processor.postprocess(image_rec, output_type="pil") + + # 2. get embeddings + do_classifier_free_guidance = source_guidance_scale > 1.0 + + lora_scale = cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + + uncond_embedding, text_embeddings, _ = self.encode_prompt( + num_images_per_prompt=1, + device=self.device, + negative_prompt=None, + enable_edit_guidance=do_classifier_free_guidance, + editing_prompt=source_prompt, + lora_scale=lora_scale, + clip_skip=clip_skip, + ) + + # 3. find zs and xts + variance_noise_shape = (num_inversion_steps, *x0.shape) + + # intermediate latents + t_to_idx = {int(v): k for k, v in enumerate(timesteps)} + xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=uncond_embedding.dtype) + + for t in reversed(timesteps): + idx = num_inversion_steps - t_to_idx[int(t)] - 1 + noise = randn_tensor(shape=x0.shape, generator=generator, device=self.device, dtype=x0.dtype) + xts[idx] = self.scheduler.add_noise(x0, noise, torch.Tensor([t])) + xts = torch.cat([x0.unsqueeze(0), xts], dim=0) + + self.scheduler.set_timesteps(len(self.scheduler.timesteps)) + # noise maps + zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=uncond_embedding.dtype) + + with self.progress_bar(total=len(timesteps)) as progress_bar: + for t in timesteps: + idx = num_inversion_steps - t_to_idx[int(t)] - 1 + # 1. predict noise residual + xt = xts[idx + 1] + + noise_pred = self.unet(xt, timestep=t, encoder_hidden_states=uncond_embedding).sample + + if not source_prompt == "": + noise_pred_cond = self.unet(xt, timestep=t, encoder_hidden_states=text_embeddings).sample + noise_pred = noise_pred + source_guidance_scale * (noise_pred_cond - noise_pred) + + xtm1 = xts[idx] + z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, self.eta) + zs[idx] = z + + # correction to avoid error accumulation + xts[idx] = xtm1_corrected + + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self.init_latents = xts[-1].expand(self.batch_size, -1, -1, -1) + zs = zs.flip(0) + self.zs = zs + + return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec) + + @torch.no_grad() + def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="default", crops_coords=None): + image = self.image_processor.preprocess( + image=image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + height, width = image.shape[-2:] + if height % 32 != 0 or width % 32 != 0: + raise ValueError( + "Image height and width must be a factor of 32. " + "Consider down-sampling the input using the `height` and `width` parameters" + ) + resized = self.image_processor.postprocess(image=image, output_type="pil") + + if max(image.shape[-2:]) > self.vae.config["sample_size"] * 1.5: + logger.warning( + "Your input images far exceed the default resolution of the underlying diffusion model. " + "The output images may contain severe artifacts! " + "Consider down-sampling the input using the `height` and `width` parameters" + ) + image = image.to(dtype) + + x0 = self.vae.encode(image.to(self.device)).latent_dist.mode() + x0 = x0.to(dtype) + x0 = self.vae.config.scaling_factor * x0 + return x0, resized + + +def compute_noise_ddim(scheduler, prev_latents, latents, timestep, noise_pred, eta): + # 1. get previous step value (=t-1) + prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = scheduler.alphas_cumprod[timestep] + alpha_prod_t_prev = ( + scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod + ) + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5) + + # 4. Clip "predicted x_0" + if scheduler.config.clip_sample: + pred_original_sample = torch.clamp(pred_original_sample, -1, 1) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = scheduler._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * noise_pred + + # modified so that updated xtm1 is returned as well (to avoid error accumulation) + mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + if variance > 0.0: + noise = (prev_latents - mu_xt) / (variance ** (0.5) * eta) + else: + noise = torch.tensor([0.0]).to(latents.device) + + return noise, mu_xt + (eta * variance**0.5) * noise + + +def compute_noise_sde_dpm_pp_2nd(scheduler, prev_latents, latents, timestep, noise_pred, eta): + def first_order_update(model_output, sample): # timestep, prev_timestep, sample): + sigma_t, sigma_s = scheduler.sigmas[scheduler.step_index + 1], scheduler.sigmas[scheduler.step_index] + alpha_t, sigma_t = scheduler._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = scheduler._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + + mu_xt = (sigma_t / sigma_s * torch.exp(-h)) * sample + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + mu_xt = scheduler.dpm_solver_first_order_update( + model_output=model_output, sample=sample, noise=torch.zeros_like(sample) + ) + + sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) + if sigma > 0.0: + noise = (prev_latents - mu_xt) / sigma + else: + noise = torch.tensor([0.0]).to(sample.device) + + prev_sample = mu_xt + sigma * noise + return noise, prev_sample + + def second_order_update(model_output_list, sample): # timestep_list, prev_timestep, sample): + sigma_t, sigma_s0, sigma_s1 = ( + scheduler.sigmas[scheduler.step_index + 1], + scheduler.sigmas[scheduler.step_index], + scheduler.sigmas[scheduler.step_index - 1], + ) + + alpha_t, sigma_t = scheduler._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = scheduler._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = scheduler._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + + mu_xt = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + ) + + sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) + if sigma > 0.0: + noise = (prev_latents - mu_xt) / sigma + else: + noise = torch.tensor([0.0]).to(sample.device) + + prev_sample = mu_xt + sigma * noise + + return noise, prev_sample + + if scheduler.step_index is None: + scheduler._init_step_index(timestep) + + model_output = scheduler.convert_model_output(model_output=noise_pred, sample=latents) + for i in range(scheduler.config.solver_order - 1): + scheduler.model_outputs[i] = scheduler.model_outputs[i + 1] + scheduler.model_outputs[-1] = model_output + + if scheduler.lower_order_nums < 1: + noise, prev_sample = first_order_update(model_output, latents) + else: + noise, prev_sample = second_order_update(scheduler.model_outputs, latents) + + if scheduler.lower_order_nums < scheduler.config.solver_order: + scheduler.lower_order_nums += 1 + + # upon completion increase step index by one + scheduler._step_index += 1 + + return noise, prev_sample + + +def compute_noise(scheduler, *args): + if isinstance(scheduler, DDIMScheduler): + return compute_noise_ddim(scheduler, *args) + elif ( + isinstance(scheduler, DPMSolverMultistepScheduler) + and scheduler.config.algorithm_type == "sde-dpmsolver++" + and scheduler.config.solver_order == 2 + ): + return compute_noise_sde_dpm_pp_2nd(scheduler, *args) + else: + raise NotImplementedError diff --git a/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py b/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py new file mode 100644 index 0000000000000000000000000000000000000000..c2e6f3caaf476a1b246c06dfc16cc2ba6101a49d --- /dev/null +++ b/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py @@ -0,0 +1,1807 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import torch +import torch.nn.functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention_processor import ( + Attention, + AttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import DDIMScheduler, DPMSolverMultistepScheduler +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import LEditsPPDiffusionPipelineOutput, LEditsPPInversionPipelineOutput + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + + >>> from diffusers import LEditsPPPipelineStableDiffusionXL + >>> from diffusers.utils import load_image + + >>> pipe = LEditsPPPipelineStableDiffusionXL.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe.vae.enable_tiling() + >>> pipe = pipe.to("cuda") + + >>> img_url = "https://www.aiml.informatik.tu-darmstadt.de/people/mbrack/tennis.jpg" + >>> image = load_image(img_url).resize((1024, 1024)) + + >>> _ = pipe.invert(image=image, num_inversion_steps=50, skip=0.2) + + >>> edited_image = pipe( + ... editing_prompt=["tennis ball", "tomato"], + ... reverse_editing_direction=[True, False], + ... edit_guidance_scale=[5.0, 10.0], + ... edit_threshold=[0.9, 0.85], + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LeditsAttentionStore +class LeditsAttentionStore: + @staticmethod + def get_empty_store(): + return {"down_cross": [], "mid_cross": [], "up_cross": [], "down_self": [], "mid_self": [], "up_self": []} + + def __call__(self, attn, is_cross: bool, place_in_unet: str, editing_prompts, PnP=False): + # attn.shape = batch_size * head_size, seq_len query, seq_len_key + if attn.shape[1] <= self.max_size: + bs = 1 + int(PnP) + editing_prompts + skip = 2 if PnP else 1 # skip PnP & unconditional + attn = torch.stack(attn.split(self.batch_size)).permute(1, 0, 2, 3) + source_batch_size = int(attn.shape[1] // bs) + self.forward(attn[:, skip * source_batch_size :], is_cross, place_in_unet) + + def forward(self, attn, is_cross: bool, place_in_unet: str): + key = f"{place_in_unet}_{'cross' if is_cross else 'self'}" + + self.step_store[key].append(attn) + + def between_steps(self, store_step=True): + if store_step: + if self.average: + if len(self.attention_store) == 0: + self.attention_store = self.step_store + else: + for key in self.attention_store: + for i in range(len(self.attention_store[key])): + self.attention_store[key][i] += self.step_store[key][i] + else: + if len(self.attention_store) == 0: + self.attention_store = [self.step_store] + else: + self.attention_store.append(self.step_store) + + self.cur_step += 1 + self.step_store = self.get_empty_store() + + def get_attention(self, step: int): + if self.average: + attention = { + key: [item / self.cur_step for item in self.attention_store[key]] for key in self.attention_store + } + else: + assert step is not None + attention = self.attention_store[step] + return attention + + def aggregate_attention( + self, attention_maps, prompts, res: int | tuple[int], from_where: list[str], is_cross: bool, select: int + ): + out = [[] for x in range(self.batch_size)] + if isinstance(res, int): + num_pixels = res**2 + resolution = (res, res) + else: + num_pixels = res[0] * res[1] + resolution = res[:2] + + for location in from_where: + for bs_item in attention_maps[f"{location}_{'cross' if is_cross else 'self'}"]: + for batch, item in enumerate(bs_item): + if item.shape[1] == num_pixels: + cross_maps = item.reshape(len(prompts), -1, *resolution, item.shape[-1])[select] + out[batch].append(cross_maps) + + out = torch.stack([torch.cat(x, dim=0) for x in out]) + # average over heads + out = out.sum(1) / out.shape[1] + return out + + def __init__(self, average: bool, batch_size=1, max_resolution=16, max_size: int = None): + self.step_store = self.get_empty_store() + self.attention_store = [] + self.cur_step = 0 + self.average = average + self.batch_size = batch_size + if max_size is None: + self.max_size = max_resolution**2 + elif max_size is not None and max_resolution is None: + self.max_size = max_size + else: + raise ValueError("Only allowed to set one of max_resolution or max_size") + + +# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LeditsGaussianSmoothing +class LeditsGaussianSmoothing: + def __init__(self, device): + kernel_size = [3, 3] + sigma = [0.5, 0.5] + + # The gaussian kernel is the product of the gaussian function of each dimension. + kernel = 1 + meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size], indexing="ij") + for size, std, mgrid in zip(kernel_size, sigma, meshgrids): + mean = (size - 1) / 2 + kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2)) + + # Make sure sum of values in gaussian kernel equals 1. + kernel = kernel / torch.sum(kernel) + + # Reshape to depthwise convolutional weight + kernel = kernel.view(1, 1, *kernel.size()) + kernel = kernel.repeat(1, *[1] * (kernel.dim() - 1)) + + self.weight = kernel.to(device) + + def __call__(self, input): + """ + Arguments: + Apply gaussian filter to input. + input (torch.Tensor): Input to apply gaussian filter on. + Returns: + filtered (torch.Tensor): Filtered output. + """ + return F.conv2d(input, weight=self.weight.to(input.dtype)) + + +# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LEDITSCrossAttnProcessor +class LEDITSCrossAttnProcessor: + def __init__(self, attention_store, place_in_unet, pnp, editing_prompts): + self.attnstore = attention_store + self.place_in_unet = place_in_unet + self.editing_prompts = editing_prompts + self.pnp = pnp + + def __call__( + self, + attn: Attention, + hidden_states, + encoder_hidden_states, + attention_mask=None, + temb=None, + ): + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + self.attnstore( + attention_probs, + is_cross=True, + place_in_unet=self.place_in_unet, + editing_prompts=self.editing_prompts, + PnP=self.pnp, + ) + + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states / attn.rescale_output_factor + return hidden_states + + +class LEditsPPPipelineStableDiffusionXL( + DiffusionPipeline, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, +): + """ + Pipeline for textual image editing using LEDits++ with Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`] and builds on the [`StableDiffusionXLPipeline`]. Check the + superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a + particular device, etc.). + + In addition the pipeline inherits the following loading methods: + - *LoRA*: [`LEditsPPPipelineStableDiffusionXL.load_lora_weights`] + - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`] + + as well as the following saving methods: + - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`] + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer ([`~transformers.CLIPTokenizer`]): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 ([`~transformers.CLIPTokenizer`]): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of + [`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]. If any other scheduler is passed it will + automatically be set to [`DPMSolverMultistepScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: DPMSolverMultistepScheduler | DDIMScheduler, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + if not isinstance(scheduler, DDIMScheduler) and not isinstance(scheduler, DPMSolverMultistepScheduler): + self.scheduler = DPMSolverMultistepScheduler.from_config( + scheduler.config, algorithm_type="sde-dpmsolver++", solver_order=2 + ) + logger.warning( + "This pipeline only supports DDIMScheduler and DPMSolverMultistepScheduler. " + "The scheduler has been changed to DPMSolverMultistepScheduler." + ) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + self.inversion_steps = None + + def encode_prompt( + self, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + enable_edit_guidance: bool = True, + editing_prompt: str | None = None, + editing_prompt_embeds: torch.Tensor | None = None, + editing_pooled_prompt_embeds: torch.Tensor | None = None, + ) -> object: + r""" + Encodes the prompt into text encoder hidden states. + + Args: + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + enable_edit_guidance (`bool`): + Whether to guide towards an editing prompt or not. + editing_prompt (`str` or `list[str]`, *optional*): + Editing prompt(s) to be encoded. If not defined and 'enable_edit_guidance' is True, one has to pass + `editing_prompt_embeds` instead. + editing_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided and 'enable_edit_guidance' is True, editing_prompt_embeds will be generated from + `editing_prompt` input argument. + editing_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated edit pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled editing_pooled_prompt_embeds will be generated from `editing_prompt` + input argument. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + batch_size = self.batch_size + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + num_edit_tokens = 0 + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + + if negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + + if batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but image inversion " + f" has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of the input images." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(negative_prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(negative_pooled_prompt_embeds) + + if enable_edit_guidance and editing_prompt_embeds is None: + editing_prompt_2 = editing_prompt + + editing_prompts = [editing_prompt, editing_prompt_2] + edit_prompt_embeds_list = [] + + for editing_prompt, tokenizer, text_encoder in zip(editing_prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + editing_prompt = self.maybe_convert_prompt(editing_prompt, tokenizer) + + max_length = negative_prompt_embeds.shape[1] + edit_concepts_input = tokenizer( + # [x for item in editing_prompt for x in repeat(item, batch_size)], + editing_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + return_length=True, + ) + num_edit_tokens = edit_concepts_input.length - 2 + + edit_concepts_embeds = text_encoder( + edit_concepts_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + editing_pooled_prompt_embeds = edit_concepts_embeds[0] + if clip_skip is None: + edit_concepts_embeds = edit_concepts_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + edit_concepts_embeds = edit_concepts_embeds.hidden_states[-(clip_skip + 2)] + + edit_prompt_embeds_list.append(edit_concepts_embeds) + + edit_concepts_embeds = torch.concat(edit_prompt_embeds_list, dim=-1) + elif not enable_edit_guidance: + edit_concepts_embeds = None + editing_pooled_prompt_embeds = None + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + bs_embed, seq_len, _ = negative_prompt_embeds.shape + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if enable_edit_guidance: + bs_embed_edit, seq_len, _ = edit_concepts_embeds.shape + edit_concepts_embeds = edit_concepts_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + edit_concepts_embeds = edit_concepts_embeds.repeat(1, num_images_per_prompt, 1) + edit_concepts_embeds = edit_concepts_embeds.view(bs_embed_edit * num_images_per_prompt, seq_len, -1) + + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if enable_edit_guidance: + editing_pooled_prompt_embeds = editing_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed_edit * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return ( + negative_prompt_embeds, + edit_concepts_embeds, + negative_pooled_prompt_embeds, + editing_pooled_prompt_embeds, + num_edit_tokens, + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, eta, generator=None): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ): + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, device, latents): + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + # Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LEditsPPPipelineStableDiffusion.prepare_unet + def prepare_unet(self, attention_store, PnP: bool = False): + attn_procs = {} + for name in self.unet.attn_processors.keys(): + if name.startswith("mid_block"): + place_in_unet = "mid" + elif name.startswith("up_blocks"): + place_in_unet = "up" + elif name.startswith("down_blocks"): + place_in_unet = "down" + else: + continue + + if "attn2" in name and place_in_unet != "mid": + attn_procs[name] = LEDITSCrossAttnProcessor( + attention_store=attention_store, + place_in_unet=place_in_unet, + pnp=PnP, + editing_prompts=self.enabled_editing_prompts, + ) + else: + attn_procs[name] = AttnProcessor() + + self.unet.set_attn_processor(attn_procs) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + denoising_end: float | None = None, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + editing_prompt: str | list[str] | None = None, + editing_prompt_embeddings: torch.Tensor | None = None, + editing_pooled_prompt_embeds: torch.Tensor | None = None, + reverse_editing_direction: bool | list[bool] | None = False, + edit_guidance_scale: float | list[float] | None = 5, + edit_warmup_steps: int | list[int] | None = 0, + edit_cooldown_steps: int | list[int] | None = None, + edit_threshold: float | list[float] | None = 0.9, + sem_guidance: list[torch.Tensor] | None = None, + use_cross_attn_mask: bool = False, + use_intersect_mask: bool = False, + user_mask: torch.Tensor | None = None, + attn_store_steps: list[int] | None = [], + store_averaged_over_steps: bool = True, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for editing. The + [`~pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.invert`] method has to be called beforehand. Edits + will always be performed for the last inverted image(s). + + Args: + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.7): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + editing_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. The image is reconstructed by setting + `editing_prompt = None`. Guidance direction of prompt should be specified via + `reverse_editing_direction`. + editing_prompt_embeddings (`torch.Tensor`, *optional*): + Pre-generated edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, editing_prompt_embeddings will be generated from `editing_prompt` input argument. + editing_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, editing_prompt_embeddings will be generated from `editing_prompt` input + argument. + reverse_editing_direction (`bool` or `list[bool]`, *optional*, defaults to `False`): + Whether the corresponding prompt in `editing_prompt` should be increased or decreased. + edit_guidance_scale (`float` or `list[float]`, *optional*, defaults to 5): + Guidance scale for guiding the image generation. If provided as list values should correspond to + `editing_prompt`. `edit_guidance_scale` is defined as `s_e` of equation 12 of [LEDITS++ + Paper](https://huggingface.co/papers/2301.12247). + edit_warmup_steps (`float` or `list[float]`, *optional*, defaults to 10): + Number of diffusion steps (for each prompt) for which guidance is not applied. + edit_cooldown_steps (`float` or `list[float]`, *optional*, defaults to `None`): + Number of diffusion steps (for each prompt) after which guidance is no longer applied. + edit_threshold (`float` or `list[float]`, *optional*, defaults to 0.9): + Masking threshold of guidance. Threshold should be proportional to the image region that is modified. + 'edit_threshold' is defined as 'λ' of equation 12 of [LEDITS++ + Paper](https://huggingface.co/papers/2301.12247). + sem_guidance (`list[torch.Tensor]`, *optional*): + list of pre-generated guidance vectors to be applied at generation. Length of the list has to + correspond to `num_inference_steps`. + use_cross_attn_mask: + Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask + is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of [LEDITS++ + paper](https://huggingface.co/papers/2311.16711). + use_intersect_mask: + Whether the masking term is calculated as intersection of cross-attention masks and masks derived from + the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate + are defined as 'M^2' of equation 12 of [LEDITS++ paper](https://huggingface.co/papers/2311.16711). + user_mask: + User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s + implicit masks do not meet user preferences. + attn_store_steps: + Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes. + store_averaged_over_steps: + Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If + False, attention maps for each step are stores separately. Just for visualization purposes. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When + returning a tuple, the first element is a list with the generated images. + """ + if self.inversion_steps is None: + raise ValueError( + "You need to invert an input image first before calling the pipeline. The `invert` method has to be called beforehand. Edits will always be performed for the last inverted image(s)." + ) + + eta = self.eta + num_images_per_prompt = 1 + latents = self.init_latents + + zs = self.zs + self.scheduler.set_timesteps(len(self.scheduler.timesteps)) + + if use_intersect_mask: + use_cross_attn_mask = True + + if use_cross_attn_mask: + self.smoothing = LeditsGaussianSmoothing(self.device) + + if user_mask is not None: + user_mask = user_mask.to(self.device) + + # TODO: Check inputs + # 1. Check inputs. Raise error if not correct + # self.check_inputs( + # callback_steps, + # negative_prompt, + # negative_prompt_2, + # prompt_embeds, + # negative_prompt_embeds, + # pooled_prompt_embeds, + # negative_pooled_prompt_embeds, + # ) + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + + # 2. Define call parameters + batch_size = self.batch_size + + device = self._execution_device + + if editing_prompt: + enable_edit_guidance = True + if isinstance(editing_prompt, str): + editing_prompt = [editing_prompt] + self.enabled_editing_prompts = len(editing_prompt) + elif editing_prompt_embeddings is not None: + enable_edit_guidance = True + self.enabled_editing_prompts = editing_prompt_embeddings.shape[0] + else: + self.enabled_editing_prompts = 0 + enable_edit_guidance = False + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + edit_prompt_embeds, + negative_pooled_prompt_embeds, + pooled_edit_embeds, + num_edit_tokens, + ) = self.encode_prompt( + device=device, + num_images_per_prompt=num_images_per_prompt, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_embeds=negative_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + enable_edit_guidance=enable_edit_guidance, + editing_prompt=editing_prompt, + editing_prompt_embeds=editing_prompt_embeddings, + editing_pooled_prompt_embeds=editing_pooled_prompt_embeds, + ) + + # 4. Prepare timesteps + # self.scheduler.set_timesteps(num_inference_steps, device=device) + + timesteps = self.inversion_steps + t_to_idx = {int(v): k for k, v in enumerate(timesteps)} + + if use_cross_attn_mask: + self.attention_store = LeditsAttentionStore( + average=store_averaged_over_steps, + batch_size=batch_size, + max_size=(latents.shape[-2] / 4.0) * (latents.shape[-1] / 4.0), + max_resolution=None, + ) + self.prepare_unet(self.attention_store) + resolution = latents.shape[-2:] + att_res = (int(resolution[0] / 4), int(resolution[1] / 4)) + + # 5. Prepare latent variables + latents = self.prepare_latents(device=device, latents=latents) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(eta) + + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(negative_pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + # 7. Prepare added time ids & embeddings + add_text_embeds = negative_pooled_prompt_embeds + add_time_ids = self._get_add_time_ids( + self.size, + crops_coords_top_left, + self.size, + dtype=negative_pooled_prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if enable_edit_guidance: + prompt_embeds = torch.cat([prompt_embeds, edit_prompt_embeds], dim=0) + add_text_embeds = torch.cat([add_text_embeds, pooled_edit_embeds], dim=0) + edit_concepts_time_ids = add_time_ids.repeat(edit_prompt_embeds.shape[0], 1) + add_time_ids = torch.cat([add_time_ids, edit_concepts_time_ids], dim=0) + self.text_cross_attention_maps = [editing_prompt] if isinstance(editing_prompt, str) else editing_prompt + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None: + # TODO: fix image encoding + image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt) + if self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds]) + image_embeds = image_embeds.to(device) + + # 8. Denoising loop + self.sem_guidance = None + self.activation_mask = None + + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=self._num_timesteps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (1 + self.enabled_editing_prompts)) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None: + added_cond_kwargs["image_embeds"] = image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + noise_pred_out = noise_pred.chunk(1 + self.enabled_editing_prompts) # [b,4, 64, 64] + noise_pred_uncond = noise_pred_out[0] + noise_pred_edit_concepts = noise_pred_out[1:] + + noise_guidance_edit = torch.zeros( + noise_pred_uncond.shape, + device=self.device, + dtype=noise_pred_uncond.dtype, + ) + + if sem_guidance is not None and len(sem_guidance) > i: + noise_guidance_edit += sem_guidance[i].to(self.device) + + elif enable_edit_guidance: + if self.activation_mask is None: + self.activation_mask = torch.zeros( + (len(timesteps), self.enabled_editing_prompts, *noise_pred_edit_concepts[0].shape) + ) + if self.sem_guidance is None: + self.sem_guidance = torch.zeros((len(timesteps), *noise_pred_uncond.shape)) + + # noise_guidance_edit = torch.zeros_like(noise_guidance) + for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts): + if isinstance(edit_warmup_steps, list): + edit_warmup_steps_c = edit_warmup_steps[c] + else: + edit_warmup_steps_c = edit_warmup_steps + if i < edit_warmup_steps_c: + continue + + if isinstance(edit_guidance_scale, list): + edit_guidance_scale_c = edit_guidance_scale[c] + else: + edit_guidance_scale_c = edit_guidance_scale + + if isinstance(edit_threshold, list): + edit_threshold_c = edit_threshold[c] + else: + edit_threshold_c = edit_threshold + if isinstance(reverse_editing_direction, list): + reverse_editing_direction_c = reverse_editing_direction[c] + else: + reverse_editing_direction_c = reverse_editing_direction + + if isinstance(edit_cooldown_steps, list): + edit_cooldown_steps_c = edit_cooldown_steps[c] + elif edit_cooldown_steps is None: + edit_cooldown_steps_c = i + 1 + else: + edit_cooldown_steps_c = edit_cooldown_steps + + if i >= edit_cooldown_steps_c: + continue + + noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond + + if reverse_editing_direction_c: + noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1 + + noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c + + if user_mask is not None: + noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask + + if use_cross_attn_mask: + out = self.attention_store.aggregate_attention( + attention_maps=self.attention_store.step_store, + prompts=self.text_cross_attention_maps, + res=att_res, + from_where=["up", "down"], + is_cross=True, + select=self.text_cross_attention_maps.index(editing_prompt[c]), + ) + attn_map = out[:, :, :, 1 : 1 + num_edit_tokens[c]] # 0 -> startoftext + + # average over all tokens + if attn_map.shape[3] != num_edit_tokens[c]: + raise ValueError( + f"Incorrect shape of attention_map. Expected size {num_edit_tokens[c]}, but found {attn_map.shape[3]}!" + ) + attn_map = torch.sum(attn_map, dim=3) + + # gaussian_smoothing + attn_map = F.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect") + attn_map = self.smoothing(attn_map).squeeze(1) + + # torch.quantile function expects float32 + if attn_map.dtype == torch.float32: + tmp = torch.quantile(attn_map.flatten(start_dim=1), edit_threshold_c, dim=1) + else: + tmp = torch.quantile( + attn_map.flatten(start_dim=1).to(torch.float32), edit_threshold_c, dim=1 + ).to(attn_map.dtype) + attn_mask = torch.where( + attn_map >= tmp.unsqueeze(1).unsqueeze(1).repeat(1, *att_res), 1.0, 0.0 + ) + + # resolution must match latent space dimension + attn_mask = F.interpolate( + attn_mask.unsqueeze(1), + noise_guidance_edit_tmp.shape[-2:], # 64,64 + ).repeat(1, 4, 1, 1) + self.activation_mask[i, c] = attn_mask.detach().cpu() + if not use_intersect_mask: + noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask + + if use_intersect_mask: + noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp) + noise_guidance_edit_tmp_quantile = torch.sum( + noise_guidance_edit_tmp_quantile, dim=1, keepdim=True + ) + noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat( + 1, self.unet.config.in_channels, 1, 1 + ) + + # torch.quantile function expects float32 + if noise_guidance_edit_tmp_quantile.dtype == torch.float32: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2), + edit_threshold_c, + dim=2, + keepdim=False, + ) + else: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32), + edit_threshold_c, + dim=2, + keepdim=False, + ).to(noise_guidance_edit_tmp_quantile.dtype) + + intersect_mask = ( + torch.where( + noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], + torch.ones_like(noise_guidance_edit_tmp), + torch.zeros_like(noise_guidance_edit_tmp), + ) + * attn_mask + ) + + self.activation_mask[i, c] = intersect_mask.detach().cpu() + + noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask + + elif not use_cross_attn_mask: + # calculate quantile + noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp) + noise_guidance_edit_tmp_quantile = torch.sum( + noise_guidance_edit_tmp_quantile, dim=1, keepdim=True + ) + noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1) + + # torch.quantile function expects float32 + if noise_guidance_edit_tmp_quantile.dtype == torch.float32: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2), + edit_threshold_c, + dim=2, + keepdim=False, + ) + else: + tmp = torch.quantile( + noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32), + edit_threshold_c, + dim=2, + keepdim=False, + ).to(noise_guidance_edit_tmp_quantile.dtype) + + self.activation_mask[i, c] = ( + torch.where( + noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], + torch.ones_like(noise_guidance_edit_tmp), + torch.zeros_like(noise_guidance_edit_tmp), + ) + .detach() + .cpu() + ) + + noise_guidance_edit_tmp = torch.where( + noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], + noise_guidance_edit_tmp, + torch.zeros_like(noise_guidance_edit_tmp), + ) + + noise_guidance_edit += noise_guidance_edit_tmp + + self.sem_guidance[i] = noise_guidance_edit.detach().cpu() + + noise_pred = noise_pred_uncond + noise_guidance_edit + + # compute the previous noisy sample x_t -> x_t-1 + if enable_edit_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg( + noise_pred, + noise_pred_edit_concepts.mean(dim=0, keepdim=False), + guidance_rescale=self.guidance_rescale, + ) + + idx = t_to_idx[int(t)] + latents = self.scheduler.step( + noise_pred, t, latents, variance_noise=zs[idx], **extra_step_kwargs, return_dict=False + )[0] + + # step callback + if use_cross_attn_mask: + store_step = i in attn_store_steps + self.attention_store.between_steps(store_step) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + # negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > 0 and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return LEditsPPDiffusionPipelineOutput(images=image, nsfw_content_detected=None) + + @torch.no_grad() + # Modified from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LEditsPPPipelineStableDiffusion.encode_image + def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="default", crops_coords=None): + image = self.image_processor.preprocess( + image=image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + height, width = image.shape[-2:] + if height % 32 != 0 or width % 32 != 0: + raise ValueError( + "Image height and width must be a factor of 32. " + "Consider down-sampling the input using the `height` and `width` parameters" + ) + resized = self.image_processor.postprocess(image=image, output_type="pil") + + if max(image.shape[-2:]) > self.vae.config["sample_size"] * 1.5: + logger.warning( + "Your input images far exceed the default resolution of the underlying diffusion model. " + "The output images may contain severe artifacts! " + "Consider down-sampling the input using the `height` and `width` parameters" + ) + image = image.to(self.device, dtype=dtype) + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + image = image.float() + self.upcast_vae() + + x0 = self.vae.encode(image).latent_dist.mode() + x0 = x0.to(dtype) + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + + x0 = self.vae.config.scaling_factor * x0 + return x0, resized + + @torch.no_grad() + def invert( + self, + image: PipelineImageInput, + source_prompt: str = "", + source_guidance_scale=3.5, + negative_prompt: str = None, + negative_prompt_2: str = None, + num_inversion_steps: int = 50, + skip: float = 0.15, + generator: torch.Generator | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + num_zero_noise_steps: int = 3, + cross_attention_kwargs: dict[str, Any] | None = None, + height: int | None = None, + width: int | None = None, + resize_mode: str | None = "default", + crops_coords: tuple[int, int, int, int] | None = None, + ): + r""" + The function to the pipeline for image inversion as described by the [LEDITS++ + Paper](https://huggingface.co/papers/2301.12247). If the scheduler is set to [`~schedulers.DDIMScheduler`] the + inversion proposed by [edit-friendly DPDM](https://huggingface.co/papers/2304.06140) will be performed instead. + + Args: + image (`PipelineImageInput`): + Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect + ratio. + source_prompt (`str`, defaults to `""`): + Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled + if the `source_prompt` is `""`. + source_guidance_scale (`float`, defaults to `3.5`): + Strength of guidance during inversion. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_inversion_steps (`int`, defaults to `50`): + Number of total performed inversion steps after discarding the initial `skip` steps. + skip (`float`, defaults to `0.15`): + Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values + will lead to stronger changes to the input image. `skip` has to be between `0` and `1`. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make inversion + deterministic. + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + num_zero_noise_steps (`int`, defaults to `3`): + Number of final diffusion steps that will not renoise the current image. If no steps are set to zero + SD-XL in combination with [`DPMSolverMultistepScheduler`] will produce noise artifacts. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + + Returns: + [`~pipelines.ledits_pp.LEditsPPInversionPipelineOutput`]: Output will contain the resized input image(s) + and respective VAE reconstruction(s). + """ + if height is not None and height % 32 != 0 or width is not None and width % 32 != 0: + raise ValueError("height and width must be a factor of 32.") + + # Reset attn processor, we do not want to store attn maps during inversion + self.unet.set_attn_processor(AttnProcessor()) + + self.eta = 1.0 + + self.scheduler.config.timestep_spacing = "leading" + self.scheduler.set_timesteps(int(num_inversion_steps * (1 + skip))) + self.inversion_steps = self.scheduler.timesteps[-num_inversion_steps:] + timesteps = self.inversion_steps + + num_images_per_prompt = 1 + + device = self._execution_device + + # 0. Ensure that only uncond embedding is used if prompt = "" + if source_prompt == "": + # noise pred should only be noise_pred_uncond + source_guidance_scale = 0.0 + do_classifier_free_guidance = False + else: + do_classifier_free_guidance = source_guidance_scale > 1.0 + + # 1. prepare image + x0, resized = self.encode_image( + image, + dtype=self.text_encoder_2.dtype, + height=height, + width=width, + resize_mode=resize_mode, + crops_coords=crops_coords, + ) + width = x0.shape[2] * self.vae_scale_factor + height = x0.shape[3] * self.vae_scale_factor + self.size = (height, width) + + self.batch_size = x0.shape[0] + + # 2. get embeddings + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + + if isinstance(source_prompt, str): + source_prompt = [source_prompt] * self.batch_size + + ( + negative_prompt_embeds, + prompt_embeds, + negative_pooled_prompt_embeds, + edit_pooled_prompt_embeds, + _, + ) = self.encode_prompt( + device=device, + num_images_per_prompt=num_images_per_prompt, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + editing_prompt=source_prompt, + lora_scale=text_encoder_lora_scale, + enable_edit_guidance=do_classifier_free_guidance, + ) + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(negative_pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + # 3. Prepare added time ids & embeddings + add_text_embeds = negative_pooled_prompt_embeds + add_time_ids = self._get_add_time_ids( + self.size, + crops_coords_top_left, + self.size, + dtype=negative_prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([add_text_embeds, edit_pooled_prompt_embeds], dim=0) + add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) + + negative_prompt_embeds = negative_prompt_embeds.to(device) + + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(self.batch_size * num_images_per_prompt, 1) + + # autoencoder reconstruction + if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: + self.upcast_vae() + x0_tmp = x0.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + image_rec = self.vae.decode( + x0_tmp / self.vae.config.scaling_factor, return_dict=False, generator=generator + )[0] + elif self.vae.config.force_upcast: + x0_tmp = x0.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + image_rec = self.vae.decode( + x0_tmp / self.vae.config.scaling_factor, return_dict=False, generator=generator + )[0] + else: + image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0] + + image_rec = self.image_processor.postprocess(image_rec, output_type="pil") + + # 5. find zs and xts + variance_noise_shape = (num_inversion_steps, *x0.shape) + + # intermediate latents + t_to_idx = {int(v): k for k, v in enumerate(timesteps)} + xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=negative_prompt_embeds.dtype) + + for t in reversed(timesteps): + idx = num_inversion_steps - t_to_idx[int(t)] - 1 + noise = randn_tensor(shape=x0.shape, generator=generator, device=self.device, dtype=x0.dtype) + xts[idx] = self.scheduler.add_noise(x0, noise, t.unsqueeze(0)) + xts = torch.cat([x0.unsqueeze(0), xts], dim=0) + + # noise maps + zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=negative_prompt_embeds.dtype) + + self.scheduler.set_timesteps(len(self.scheduler.timesteps)) + + for t in self.progress_bar(timesteps): + idx = num_inversion_steps - t_to_idx[int(t)] - 1 + # 1. predict noise residual + xt = xts[idx + 1] + + latent_model_input = torch.cat([xt] * 2) if do_classifier_free_guidance else xt + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=negative_prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # 2. perform guidance + if do_classifier_free_guidance: + noise_pred_out = noise_pred.chunk(2) + noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1] + noise_pred = noise_pred_uncond + source_guidance_scale * (noise_pred_text - noise_pred_uncond) + + xtm1 = xts[idx] + z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, self.eta) + zs[idx] = z + + # correction to avoid error accumulation + xts[idx] = xtm1_corrected + + self.init_latents = xts[-1] + zs = zs.flip(0) + + if num_zero_noise_steps > 0: + zs[-num_zero_noise_steps:] = torch.zeros_like(zs[-num_zero_noise_steps:]) + self.zs = zs + return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec) + + +# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.compute_noise_ddim +def compute_noise_ddim(scheduler, prev_latents, latents, timestep, noise_pred, eta): + # 1. get previous step value (=t-1) + prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = scheduler.alphas_cumprod[timestep] + alpha_prod_t_prev = ( + scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod + ) + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5) + + # 4. Clip "predicted x_0" + if scheduler.config.clip_sample: + pred_original_sample = torch.clamp(pred_original_sample, -1, 1) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = scheduler._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * noise_pred + + # modified so that updated xtm1 is returned as well (to avoid error accumulation) + mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + if variance > 0.0: + noise = (prev_latents - mu_xt) / (variance ** (0.5) * eta) + else: + noise = torch.tensor([0.0]).to(latents.device) + + return noise, mu_xt + (eta * variance**0.5) * noise + + +# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.compute_noise_sde_dpm_pp_2nd +def compute_noise_sde_dpm_pp_2nd(scheduler, prev_latents, latents, timestep, noise_pred, eta): + def first_order_update(model_output, sample): # timestep, prev_timestep, sample): + sigma_t, sigma_s = scheduler.sigmas[scheduler.step_index + 1], scheduler.sigmas[scheduler.step_index] + alpha_t, sigma_t = scheduler._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = scheduler._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + + mu_xt = (sigma_t / sigma_s * torch.exp(-h)) * sample + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + mu_xt = scheduler.dpm_solver_first_order_update( + model_output=model_output, sample=sample, noise=torch.zeros_like(sample) + ) + + sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) + if sigma > 0.0: + noise = (prev_latents - mu_xt) / sigma + else: + noise = torch.tensor([0.0]).to(sample.device) + + prev_sample = mu_xt + sigma * noise + return noise, prev_sample + + def second_order_update(model_output_list, sample): # timestep_list, prev_timestep, sample): + sigma_t, sigma_s0, sigma_s1 = ( + scheduler.sigmas[scheduler.step_index + 1], + scheduler.sigmas[scheduler.step_index], + scheduler.sigmas[scheduler.step_index - 1], + ) + + alpha_t, sigma_t = scheduler._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = scheduler._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = scheduler._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + + mu_xt = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + ) + + sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) + if sigma > 0.0: + noise = (prev_latents - mu_xt) / sigma + else: + noise = torch.tensor([0.0]).to(sample.device) + + prev_sample = mu_xt + sigma * noise + + return noise, prev_sample + + if scheduler.step_index is None: + scheduler._init_step_index(timestep) + + model_output = scheduler.convert_model_output(model_output=noise_pred, sample=latents) + for i in range(scheduler.config.solver_order - 1): + scheduler.model_outputs[i] = scheduler.model_outputs[i + 1] + scheduler.model_outputs[-1] = model_output + + if scheduler.lower_order_nums < 1: + noise, prev_sample = first_order_update(model_output, latents) + else: + noise, prev_sample = second_order_update(scheduler.model_outputs, latents) + + if scheduler.lower_order_nums < scheduler.config.solver_order: + scheduler.lower_order_nums += 1 + + # upon completion increase step index by one + scheduler._step_index += 1 + + return noise, prev_sample + + +# Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.compute_noise +def compute_noise(scheduler, *args): + if isinstance(scheduler, DDIMScheduler): + return compute_noise_ddim(scheduler, *args) + elif ( + isinstance(scheduler, DPMSolverMultistepScheduler) + and scheduler.config.algorithm_type == "sde-dpmsolver++" + and scheduler.config.solver_order == 2 + ): + return compute_noise_sde_dpm_pp_2nd(scheduler, *args) + else: + raise NotImplementedError diff --git a/diffusers/pipelines/ledits_pp/pipeline_output.py b/diffusers/pipelines/ledits_pp/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..4a69e7442bdc9d7e46a060e8033654d714dc4f8e --- /dev/null +++ b/diffusers/pipelines/ledits_pp/pipeline_output.py @@ -0,0 +1,42 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class LEditsPPDiffusionPipelineOutput(BaseOutput): + """ + Output class for LEdits++ Diffusion pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + nsfw_content_detected (`list[bool]`) + list indicating whether the corresponding generated image contains “not-safe-for-work” (nsfw) content or + `None` if safety checking could not be performed. + """ + + images: list[PIL.Image.Image] | np.ndarray + nsfw_content_detected: list[bool] | None + + +@dataclass +class LEditsPPInversionPipelineOutput(BaseOutput): + """ + Output class for LEdits++ Diffusion pipelines. + + Args: + input_images (`list[PIL.Image.Image]` or `np.ndarray`) + list of the cropped and resized input images as PIL images of length `batch_size` or NumPy array of shape ` + (batch_size, height, width, num_channels)`. + vae_reconstruction_images (`list[PIL.Image.Image]` or `np.ndarray`) + list of VAE reconstruction of all input images as PIL images of length `batch_size` or NumPy array of shape + ` (batch_size, height, width, num_channels)`. + """ + + images: list[PIL.Image.Image] | np.ndarray + vae_reconstruction_images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/llada2/__init__.py b/diffusers/pipelines/llada2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..45a02e6851e22265b29efd1966e35cbecf9dd006 --- /dev/null +++ b/diffusers/pipelines/llada2/__init__.py @@ -0,0 +1,47 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_llada2"] = ["LLaDA2Pipeline", "LLaDA2PipelineOutput"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_llada2 import LLaDA2Pipeline, LLaDA2PipelineOutput +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/llada2/pipeline_llada2.py b/diffusers/pipelines/llada2/pipeline_llada2.py new file mode 100644 index 0000000000000000000000000000000000000000..06b4875f18a9acdc3b4e8e4b4cb6f637902219c2 --- /dev/null +++ b/diffusers/pipelines/llada2/pipeline_llada2.py @@ -0,0 +1,563 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from dataclasses import dataclass +from typing import Any, Callable + +import torch +from tqdm.auto import tqdm + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...schedulers import BlockRefinementScheduler +from ...utils import BaseOutput, logging, replace_example_docstring +from ..pipeline_utils import DiffusionPipeline + + +logger = logging.get_logger(__name__) + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from transformers import AutoModelForCausalLM, AutoTokenizer + >>> from diffusers import BlockRefinementScheduler, LLaDA2Pipeline + + >>> model_id = "inclusionAI/LLaDA2.1-mini" + >>> model = AutoModelForCausalLM.from_pretrained( + ... model_id, trust_remote_code=True, dtype=torch.bfloat16, device_map="auto" + ... ) + >>> tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) + >>> scheduler = BlockRefinementScheduler() + + >>> pipe = LLaDA2Pipeline(model=model, scheduler=scheduler, tokenizer=tokenizer) + >>> output = pipe(prompt="What is the meaning of life?", gen_length=256) + >>> print(output.texts[0]) + ``` +""" + + +@dataclass +class LLaDA2PipelineOutput(BaseOutput): + sequences: torch.LongTensor + texts: list[str] | None = None + + +class LLaDA2Pipeline(DiffusionPipeline): + r""" + Pipeline for LLaDA2-style discrete diffusion text generation via block-wise iterative refinement. + + This pipeline maintains a template sequence filled with a `mask_token_id` and refines it in blocks. In each + refinement step, it samples candidate tokens for the active block and commits a subset based on confidence. + + The model is expected to accept an attention mask and `position_ids`, and to return logits of shape `[batch, seq, + vocab_size]`. + """ + + model: Any + scheduler: BlockRefinementScheduler + tokenizer: Any + + _callback_tensor_inputs = [ + "block_x", + "transfer_index", + "editing_transfer_index", + "sampled_tokens", + "sampled_probs", + "active_block", + ] + + def __init__( + self, + model: Any, + scheduler: BlockRefinementScheduler, + tokenizer: Any | None = None, + ): + super().__init__() + self.register_modules(model=model, scheduler=scheduler, tokenizer=tokenizer) + self.eos_token_id = getattr(self.tokenizer, "eos_token_id", None) if self.tokenizer is not None else None + self.mask_token_id = getattr(self.tokenizer, "mask_token_id", None) if self.tokenizer is not None else None + + @property + def num_timesteps(self): + return self._num_timesteps + + # --- Prompt encoding --- + + def _prepare_input_ids( + self, + *, + prompt: str | list[str] | None, + messages: list[dict[str, str]] | None, + input_ids: torch.LongTensor | None, + use_chat_template: bool, + add_generation_prompt: bool, + chat_template_kwargs: dict[str, Any] | None, + attention_mask: torch.LongTensor | None = None, + ) -> tuple[torch.LongTensor, torch.LongTensor]: + """Convert prompt/messages/input_ids to `(input_ids, attention_mask)` tensors of shape `[batch, seq]`.""" + if input_ids is not None: + if input_ids.ndim == 1: + input_ids = input_ids.unsqueeze(0) + if input_ids.ndim != 2: + raise ValueError(f"`input_ids` must be 2D, got shape {tuple(input_ids.shape)}.") + if input_ids.dtype != torch.long: + raise ValueError(f"`input_ids` must be int64 token IDs, got dtype={input_ids.dtype}.") + if attention_mask is None: + attention_mask = torch.ones_like(input_ids, dtype=torch.long) + else: + if attention_mask.ndim == 1: + attention_mask = attention_mask.unsqueeze(0) + if attention_mask.shape != input_ids.shape: + raise ValueError( + f"`attention_mask` shape {tuple(attention_mask.shape)} must match `input_ids` shape " + f"{tuple(input_ids.shape)}." + ) + attention_mask = attention_mask.to(dtype=torch.long) + return input_ids, attention_mask + + if self.tokenizer is None: + raise ValueError("Tokenizer is required when `input_ids` is not provided.") + + if messages is not None and prompt is not None: + raise ValueError("Provide either `prompt` or `messages`, not both.") + if messages is None and prompt is None: + raise ValueError("Provide one of `prompt`, `messages`, or `input_ids`.") + + chat_template_kwargs = chat_template_kwargs or {} + + if messages is not None: + encoded = self.tokenizer.apply_chat_template( + messages, + add_generation_prompt=add_generation_prompt, + tokenize=True, + return_tensors="pt", + return_dict=True, + **chat_template_kwargs, + ) + ids = encoded["input_ids"] + mask = encoded.get("attention_mask") + if mask is None: + mask = torch.ones_like(ids, dtype=torch.long) + return ids, mask.to(dtype=torch.long) + + if use_chat_template and getattr(self.tokenizer, "chat_template", None): + if isinstance(prompt, list): + raise ValueError("`prompt` must be a string when `use_chat_template=True`.") + encoded = self.tokenizer.apply_chat_template( + [{"role": "user", "content": prompt}], + add_generation_prompt=add_generation_prompt, + tokenize=True, + return_tensors="pt", + return_dict=True, + **chat_template_kwargs, + ) + ids = encoded["input_ids"] + mask = encoded.get("attention_mask") + if mask is None: + mask = torch.ones_like(ids, dtype=torch.long) + return ids, mask.to(dtype=torch.long) + + encoded = self.tokenizer(prompt, return_tensors="pt", padding=isinstance(prompt, list)) + ids = encoded["input_ids"] + mask = encoded.get("attention_mask") + if mask is None: + mask = torch.ones_like(ids, dtype=torch.long) + return ids, mask.to(dtype=torch.long) + + def check_inputs( + self, + prompt: str | list[str] | None, + messages: list[dict[str, str]] | None, + input_ids: torch.LongTensor | None, + gen_length: int, + block_length: int, + num_inference_steps: int, + minimal_topk: int, + threshold: float, + sampling_method: str, + output_type: str, + callback_on_step_end: Callable | PipelineCallback | MultiPipelineCallbacks | None, + callback_on_step_end_tensor_inputs: list[str] | None, + ): + # Input source validation + if prompt is None and messages is None and input_ids is None: + raise ValueError("Provide one of `prompt`, `messages`, or `input_ids`.") + if prompt is not None and messages is not None: + raise ValueError("Provide either `prompt` or `messages`, not both.") + if input_ids is not None: + if input_ids.ndim not in (1, 2): + raise ValueError(f"`input_ids` must be 1D or 2D, got shape {tuple(input_ids.shape)}.") + if input_ids.dtype != torch.long: + raise ValueError(f"`input_ids` must be int64 token IDs, got dtype={input_ids.dtype}.") + if prompt is not None and input_ids is None and self.tokenizer is None: + raise ValueError("Tokenizer is required when `input_ids` is not provided.") + if messages is not None and input_ids is None and self.tokenizer is None: + raise ValueError("Tokenizer is required when `input_ids` is not provided.") + + # Generation parameter validation + if gen_length <= 0: + raise ValueError(f"`gen_length` must be > 0, got {gen_length}.") + if block_length <= 0: + raise ValueError(f"`block_length` must be > 0, got {block_length}.") + if num_inference_steps <= 0: + raise ValueError(f"`num_inference_steps` must be > 0, got {num_inference_steps}.") + if minimal_topk <= 0: + raise ValueError(f"`minimal_topk` must be > 0, got {minimal_topk}.") + if not (0.0 <= threshold <= 1.0) and not (threshold > 1.0): + raise ValueError(f"`threshold` must be in [0, 1] (or > 1 to force top-k commits), got {threshold}.") + if sampling_method not in {"auto", "greedy", "multinomial"}: + raise ValueError( + f"`sampling_method` must be one of {{'auto','greedy','multinomial'}}, got {sampling_method!r}." + ) + if output_type not in {"seq", "text"}: + raise ValueError(f"`output_type` must be 'seq' or 'text', got {output_type!r}.") + + # Callback validation + if callback_on_step_end is not None and isinstance( + callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks) + ): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found " + f"{[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + messages: list[dict[str, str]] | None = None, + input_ids: torch.LongTensor | None = None, + attention_mask: torch.LongTensor | None = None, + use_chat_template: bool = True, + add_generation_prompt: bool = True, + gen_length: int = 2048, + block_length: int | None = None, + num_inference_steps: int = 32, + temperature: float = 0.0, + top_p: float | None = None, + top_k: int | None = None, + sampling_method: str = "multinomial", + threshold: float = 0.7, + editing_threshold: float | None = 0.5, + max_post_steps: int = 16, + minimal_topk: int = 1, + eos_early_stop: bool = True, + eos_token_id: int | None = None, + mask_token_id: int | None = None, + generator: torch.Generator | None = None, + output_type: str = "text", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int, dict], None] + | PipelineCallback + | MultiPipelineCallbacks + | None = None, + callback_on_step_end_tensor_inputs: list[str] | None = None, + ) -> LLaDA2PipelineOutput | tuple[torch.LongTensor, list[str] | None]: + """ + Generate text with block-wise refinement. + + Args: + prompt (`str` or `List[str]`, *optional*): + Prompt text. When `use_chat_template` is `True` (default) and a tokenizer with a chat template is + available, the prompt is wrapped in a chat message before tokenization. + messages (`List[Dict[str, str]]`, *optional*): + Chat messages to encode (e.g. `[{"role": "user", "content": "Hello"}]`). Takes precedence over `prompt` + when provided. Requires a tokenizer with `apply_chat_template`. + input_ids (`torch.LongTensor`, *optional*): + Pre-tokenized input IDs. Takes precedence over `prompt` and `messages`. + attention_mask (`torch.LongTensor`, *optional*): + Per-token mask (1 for valid prompt tokens, 0 for padding) matching the shape of `input_ids`. Only used + when `input_ids` is provided. When omitted (and `input_ids` is given), all positions are treated as + valid. When constructing inputs from `prompt` / `messages`, the tokenizer's mask is carried through + automatically. + use_chat_template (`bool`, defaults to `True`): + Whether to wrap the prompt in a chat template. + add_generation_prompt (`bool`, defaults to `True`): + Whether to add the generation prompt when using chat templates. + gen_length (`int`): + Number of tokens to generate. + block_length (`int`, *optional*): + Block size for refinement. If not provided, the scheduler's configured `block_length` is used. + num_inference_steps (`int`): + Number of refinement steps per block. + temperature (`float`): + Sampling temperature. + top_p (`float`, *optional*): + Nucleus sampling cutoff. + top_k (`int`, *optional*): + Top-k sampling cutoff. + sampling_method (`str`): + Sampling method (`auto`, `greedy`, `multinomial`). + threshold (`float`): + Confidence threshold for committing tokens. + editing_threshold (`float`, *optional*): + Confidence threshold for editing already-committed (non-mask) tokens. When positive, after all mask + tokens in a block are resolved, the pipeline continues refining: if the model predicts a different + token with confidence above this threshold, the existing token is replaced. Set to `None`, `0.0`, or a + negative value to disable editing. Defaults to `0.5`. + max_post_steps (`int`): + Maximum number of additional refinement iterations after all mask tokens in a block are resolved. Only + used when `editing_threshold` is enabled. Defaults to `16`. + minimal_topk (`int`): + Minimum number of tokens to commit per step. + eos_early_stop (`bool`): + Whether to stop after committing EOS in a block. + eos_token_id (`int`, *optional*): + EOS token ID to use for early stopping. + mask_token_id (`int`, *optional*): + Mask token ID to use for the template. + generator (`torch.Generator`, *optional*): + RNG for sampling. + output_type (`str`, defaults to `"text"`): + Output format. `"text"` decodes sequences into strings (requires a tokenizer). `"seq"` returns raw + token ID sequences only. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`LLaDA2PipelineOutput`] instead of a tuple. + callback_on_step_end (`Callable` or `PipelineCallback`, *optional*): + Callback executed after each refinement step with signature `callback_on_step_end(self, step: int, + timestep: int, callback_kwargs: Dict)`. + callback_on_step_end_tensor_inputs (`List[str]`, *optional*): + Tensor keys to pass to the callback. Allowed keys: `block_x`, `transfer_index`, + `editing_transfer_index`, `sampled_tokens`, `sampled_probs`, `active_block`. + + Examples: + + Returns: + [`~pipelines.llada2.pipeline_llada2.LLaDA2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.llada2.pipeline_llada2.LLaDA2PipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is the generated token IDs (`torch.LongTensor`) + and the second element is the decoded texts (`list[str]`), or `None` when `output_type` is `"seq"`. + """ + # 1. Check inputs early + if callback_on_step_end is not None and isinstance( + callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks) + ): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + if callback_on_step_end_tensor_inputs is None: + callback_on_step_end_tensor_inputs = ["block_x"] + + if block_length is None: + block_length = self.scheduler.config.block_length + + self.check_inputs( + prompt=prompt, + messages=messages, + input_ids=input_ids, + gen_length=gen_length, + block_length=block_length, + num_inference_steps=num_inference_steps, + minimal_topk=minimal_topk, + threshold=threshold, + sampling_method=sampling_method, + output_type=output_type, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + # 2. Prepare input IDs from prompt/messages/input_ids + prompt_ids, prompt_attention_mask = self._prepare_input_ids( + prompt=prompt, + messages=messages, + input_ids=input_ids, + attention_mask=attention_mask, + use_chat_template=use_chat_template, + add_generation_prompt=add_generation_prompt, + chat_template_kwargs=None, + ) + + device = self._execution_device + + if prompt_ids.ndim == 1: + prompt_ids = prompt_ids.unsqueeze(0) + prompt_ids = prompt_ids.to(device=device) + prompt_attention_mask = prompt_attention_mask.to(device=device) + batch_size, prompt_length = prompt_ids.shape + + if eos_token_id is None: + eos_token_id = self.eos_token_id + if mask_token_id is None: + mask_token_id = self.mask_token_id + if mask_token_id is None: + raise ValueError("`mask_token_id` must be provided (or available on the tokenizer).") + + num_inference_steps = min(num_inference_steps, gen_length // minimal_topk) + + self.scheduler.set_timesteps(num_inference_steps, device=device, block_length=block_length) + + # 3. Build attention mask and position IDs + num_blocks = (prompt_length + gen_length + block_length - 1) // block_length + total_length = num_blocks * block_length + + # 2D attention mask: prompt tokenizer mask + ones over generated positions + zeros over the + # block-aligned tail past `prompt_length + gen_length`. The model handles backend-specific + # conversion internally; this just tells it which positions are real context. + attn_mask = torch.zeros((batch_size, total_length), device=device, dtype=torch.long) + attn_mask[:, :prompt_length] = prompt_attention_mask + attn_mask[:, prompt_length : prompt_length + gen_length] = 1 + + position_ids = torch.arange(total_length, device=device, dtype=torch.long).unsqueeze(0).expand(batch_size, -1) + + # 4. Prepare latents (fully masked sequence) + x = torch.full((batch_size, total_length), mask_token_id, device=device, dtype=torch.long) + if prompt_length > 0: + x[:, :prompt_length] = prompt_ids + + prefill_blocks = prompt_length // block_length + self._num_timesteps = num_inference_steps * max(num_blocks - prefill_blocks, 0) + + finished = torch.zeros((batch_size,), device=device, dtype=torch.bool) + editing_enabled = editing_threshold is not None and editing_threshold > 0.0 + global_step = 0 + + # 5. Block-wise refinement loop + outer_progress_bar_config = getattr(self, "_progress_bar_config", {}).copy() + block_progress_bar_config = {**outer_progress_bar_config, "position": 0, "desc": "Blocks"} + for num_block in tqdm(range(prefill_blocks, num_blocks), **block_progress_bar_config): + current_window_end = (num_block + 1) * block_length + block_x = x[:, :current_window_end] + block_attn_mask = attn_mask[:, :current_window_end] + block_position_ids = position_ids[:, :current_window_end] + + # Identify which positions in the block are prompt (non-editable). + block_start_pos = num_block * block_length + prompt_mask_in_block = torch.zeros(block_length, device=device, dtype=torch.bool) + if block_start_pos < prompt_length: + prompt_end_in_block = min(prompt_length - block_start_pos, block_length) + prompt_mask_in_block[:prompt_end_in_block] = True + + post_steps = 0 + step_idx = 0 + should_continue = True + inner_progress_bar_config = { + **outer_progress_bar_config, + "position": 1, + "leave": False, + "desc": f"Block {num_block} Inference Steps", + } + progress_bar = tqdm(total=num_inference_steps, **inner_progress_bar_config) + + while should_continue: + block_tokens = block_x[:, -block_length:] + masks_remaining = (block_tokens == mask_token_id).any() + + if not masks_remaining: + post_steps += 1 + + logits = self.model(block_x, attention_mask=block_attn_mask, position_ids=block_position_ids).logits + block_logits = logits[:, -block_length:, :] + + scheduler_output = self.scheduler.step( + model_output=block_logits, + timestep=step_idx, + sample=block_tokens, + mask_token_id=mask_token_id, + temperature=temperature, + top_p=top_p, + top_k=top_k, + sampling_method=sampling_method, + threshold=threshold, + editing_threshold=editing_threshold, + minimal_topk=minimal_topk, + prompt_mask=prompt_mask_in_block, + generator=generator, + return_dict=True, + ) + + transfer_index = scheduler_output.transfer_index + editing_transfer_index = scheduler_output.editing_transfer_index + sampled_tokens = scheduler_output.sampled_tokens + sampled_probs = scheduler_output.sampled_probs + active_block = block_tokens == mask_token_id + final_transfer = transfer_index | editing_transfer_index + + # Freeze rows that already emitted EOS so further blocks don't extend them. + if eos_early_stop and finished.any(): + final_transfer = final_transfer & ~finished[:, None] + + if final_transfer.any(): + block_x[:, -block_length:] = torch.where( + final_transfer, scheduler_output.prev_sample, block_tokens + ) + + if eos_early_stop and eos_token_id is not None: + finished = self.scheduler.check_eos_finished( + cur_x=block_x, + sampled_tokens=scheduler_output.sampled_tokens, + final_transfer=final_transfer, + finished=finished, + eos_token_id=eos_token_id, + mask_token_id=mask_token_id, + prompt_length=prompt_length, + ) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, global_step, step_idx, callback_kwargs) + block_x = callback_outputs.pop("block_x", block_x) + + global_step += 1 + if masks_remaining: + step_idx += 1 + progress_bar.update(1) + + should_continue = self.scheduler.check_block_should_continue( + step_idx=step_idx, + masks_remaining=masks_remaining, + editing_enabled=editing_enabled, + editing_transfer_index=editing_transfer_index, + post_steps=post_steps, + max_post_steps=max_post_steps, + finished=finished, + ) + + progress_bar.close() + x[:, :current_window_end] = block_x + if eos_early_stop and finished.all(): + break + + # 6. Post-process output + generated = x[:, : prompt_length + gen_length] + sequences = generated[:, prompt_length:] + + # For decode, trim each row at the first EOS so post-EOS positions (which may still hold + # mask tokens or refined content for unfinished blocks) don't leak into the decoded text. + decode_sequences: list[torch.LongTensor] | torch.LongTensor = sequences + if eos_token_id is not None: + decode_sequences = [ + seq[: int((seq == eos_token_id).nonzero(as_tuple=True)[0][0]) + 1] + if (seq == eos_token_id).any() + else seq + for seq in sequences + ] + + texts = None + if output_type == "text" and self.tokenizer is not None: + texts = self.tokenizer.batch_decode(decode_sequences, skip_special_tokens=True) + + if not return_dict: + return sequences.to(device=device), texts + return LLaDA2PipelineOutput(sequences=sequences.to(device=device), texts=texts) + + +__all__ = ["LLaDA2Pipeline", "LLaDA2PipelineOutput"] diff --git a/diffusers/pipelines/longcat_audio_dit/__init__.py b/diffusers/pipelines/longcat_audio_dit/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b7c03a70371a377f2a240e3bb91624a6cc186a04 --- /dev/null +++ b/diffusers/pipelines/longcat_audio_dit/__init__.py @@ -0,0 +1,40 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa: F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_longcat_audio_dit"] = ["LongCatAudioDiTPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_longcat_audio_dit import LongCatAudioDiTPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/longcat_audio_dit/pipeline_longcat_audio_dit.py b/diffusers/pipelines/longcat_audio_dit/pipeline_longcat_audio_dit.py new file mode 100644 index 0000000000000000000000000000000000000000..e6478535b373f8a37003b15347260b5d5fc27d95 --- /dev/null +++ b/diffusers/pipelines/longcat_audio_dit/pipeline_longcat_audio_dit.py @@ -0,0 +1,358 @@ +# Copyright 2026 MeiTuan LongCat-AudioDiT Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# Adapted from the LongCat-AudioDiT reference implementation: +# https://github.com/meituan-longcat/LongCat-AudioDiT + +import re +from typing import Callable + +import torch +import torch.nn.functional as F +from transformers import PreTrainedTokenizerBase, UMT5EncoderModel + +from ...models import LongCatAudioDiTTransformer, LongCatAudioDiTVae +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import logging +from ...utils.doc_utils import replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline + + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import soundfile as sf + >>> import torch + >>> from diffusers import LongCatAudioDiTPipeline + + >>> pipe = LongCatAudioDiTPipeline.from_pretrained("ruixiangma/LongCat-AudioDiT-1B-Diffusers") + >>> pipe.to("cuda") + + >>> prompt = "A calm ocean wave ambience with soft wind in the background." + >>> audio = pipe( + ... prompt, + ... audio_duration_s=5.0, + ... num_inference_steps=20, + ... guidance_scale=4.0, + ... generator=torch.Generator("cuda").manual_seed(42), + ... ).audios[0, 0] + >>> sf.write("output.wav", audio, pipe.sample_rate) + ``` +""" + + +def _lens_to_mask(lengths: torch.Tensor, length: int | None = None) -> torch.BoolTensor: + if length is None: + length = int(lengths.amax().item()) + seq = torch.arange(length, device=lengths.device) + return seq[None, :] < lengths[:, None] + + +def _normalize_text(text: str) -> str: + text = text.lower() + text = re.sub(r'["“”‘’]', " ", text) + text = re.sub(r"\s+", " ", text) + return text.strip() + + +def _approx_duration_from_text(text: str | list[str], max_duration: float = 30.0) -> float: + if not text: + return 0.0 + if isinstance(text, str): + text = [text] + + en_dur_per_char = 0.082 + zh_dur_per_char = 0.21 + durations = [] + for prompt in text: + prompt = re.sub(r"\s+", "", prompt) + num_zh = num_en = num_other = 0 + for char in prompt: + if "一" <= char <= "鿿": + num_zh += 1 + elif char.isalpha(): + num_en += 1 + else: + num_other += 1 + if num_zh > num_en: + num_zh += num_other + else: + num_en += num_other + durations.append(num_zh * zh_dur_per_char + num_en * en_dur_per_char) + return min(max_duration, max(durations)) if durations else 0.0 + + +class LongCatAudioDiTPipeline(DiffusionPipeline): + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + vae: LongCatAudioDiTVae, + text_encoder: UMT5EncoderModel, + tokenizer: PreTrainedTokenizerBase, + transformer: LongCatAudioDiTTransformer, + scheduler: FlowMatchEulerDiscreteScheduler | None = None, + ): + super().__init__() + if not isinstance(scheduler, FlowMatchEulerDiscreteScheduler): + scheduler = FlowMatchEulerDiscreteScheduler(shift=1.0, invert_sigmas=True) + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.sample_rate = getattr(vae.config, "sample_rate", 24000) + self.vae_scale_factor = getattr(vae.config, "downsampling_ratio", 2048) + self.latent_dim = getattr(transformer.config, "latent_dim", 64) + self.max_wav_duration = 30.0 + self.text_norm_feat = True + self.text_add_embed = True + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + def encode_prompt(self, prompt: str | list[str], device: torch.device) -> tuple[torch.Tensor, torch.Tensor]: + if isinstance(prompt, str): + prompt = [prompt] + model_max_length = getattr(self.tokenizer, "model_max_length", 512) + if not isinstance(model_max_length, int) or model_max_length <= 0 or model_max_length > 32768: + model_max_length = 512 + text_inputs = self.tokenizer( + prompt, + padding="longest", + truncation=True, + max_length=model_max_length, + return_tensors="pt", + ) + input_ids = text_inputs.input_ids.to(device) + attention_mask = text_inputs.attention_mask.to(device) + with torch.no_grad(): + output = self.text_encoder(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True) + prompt_embeds = output.last_hidden_state + if self.text_norm_feat: + prompt_embeds = F.layer_norm(prompt_embeds, (prompt_embeds.shape[-1],), eps=1e-6) + if self.text_add_embed and getattr(output, "hidden_states", None): + first_hidden = output.hidden_states[0] + if self.text_norm_feat: + first_hidden = F.layer_norm(first_hidden, (first_hidden.shape[-1],), eps=1e-6) + prompt_embeds = prompt_embeds + first_hidden + lengths = attention_mask.sum(dim=1).to(device) + return prompt_embeds, lengths + + def prepare_latents( + self, + batch_size: int, + duration: int, + device: torch.device, + dtype: torch.dtype, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + if latents.ndim != 3: + raise ValueError( + f"`latents` must have shape (batch_size, duration, latent_dim), but got {tuple(latents.shape)}." + ) + if latents.shape[0] != batch_size: + raise ValueError(f"`latents` must have batch size {batch_size}, but got {latents.shape[0]}.") + if latents.shape[2] != self.latent_dim: + raise ValueError(f"`latents` must have latent_dim {self.latent_dim}, but got {latents.shape[2]}.") + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"Expected {batch_size} generators for batch size {batch_size}, but got {len(generator)}." + ) + + return randn_tensor((batch_size, duration, self.latent_dim), generator=generator, device=device, dtype=dtype) + + def check_inputs( + self, + prompt: list[str], + negative_prompt: str | list[str] | None, + output_type: str, + callback_on_step_end_tensor_inputs: list[str] | None = None, + ) -> None: + if len(prompt) == 0: + raise ValueError("`prompt` must contain at least one prompt.") + + if output_type not in {"np", "pt", "latent"}: + raise ValueError(f"Unsupported output_type: {output_type}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found " + f"{[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if negative_prompt is not None and not isinstance(negative_prompt, str): + negative_prompt = list(negative_prompt) + if len(negative_prompt) != len(prompt): + raise ValueError( + f"`negative_prompt` must have batch size {len(prompt)}, but got {len(negative_prompt)} prompts." + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + audio_duration_s: float | None = None, + latents: torch.Tensor | None = None, + num_inference_steps: int = 16, + guidance_scale: float = 4.0, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str = "np", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): Prompt or prompts that guide audio generation. + negative_prompt (`str` or `list[str]`, *optional*): Negative prompt(s) for classifier-free guidance. + audio_duration_s (`float`, *optional*): + Target audio duration in seconds. Ignored when `latents` is provided. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents of shape `(batch_size, duration, latent_dim)`. + num_inference_steps (`int`, defaults to 16): Number of denoising steps. + guidance_scale (`float`, defaults to 4.0): Guidance scale for classifier-free guidance. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): Random generator(s). + output_type (`str`, defaults to `"np"`): Output format: `"np"`, `"pt"`, or `"latent"`. + return_dict (`bool`, defaults to `True`): Whether to return `AudioPipelineOutput`. + callback_on_step_end (`Callable`, *optional*): + A function called at the end of each denoising step with the pipeline, step index, timestep, and tensor + inputs specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, defaults to `["latents"]`): + Tensor inputs passed to `callback_on_step_end`. + + Examples: + """ + if prompt is None: + prompt = [] + elif isinstance(prompt, str): + prompt = [prompt] + else: + prompt = list(prompt) + self.check_inputs(prompt, negative_prompt, output_type, callback_on_step_end_tensor_inputs) + batch_size = len(prompt) + self._guidance_scale = guidance_scale + + device = self._execution_device + normalized_prompts = [_normalize_text(text) for text in prompt] + if latents is not None: + duration = latents.shape[1] + elif audio_duration_s is not None: + duration = int(audio_duration_s * self.sample_rate // self.vae_scale_factor) + else: + duration = int(_approx_duration_from_text(normalized_prompts) * self.sample_rate // self.vae_scale_factor) + max_duration = int(self.max_wav_duration * self.sample_rate // self.vae_scale_factor) + if latents is None: + duration = max(1, min(duration, max_duration)) + + prompt_embeds, prompt_embeds_len = self.encode_prompt(normalized_prompts, device) + duration_tensor = torch.full((batch_size,), duration, device=device, dtype=torch.long) + mask = _lens_to_mask(duration_tensor) + text_mask = _lens_to_mask(prompt_embeds_len, length=prompt_embeds.shape[1]) + + if negative_prompt is None: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_prompt_embeds_len = prompt_embeds_len + negative_prompt_embeds_mask = text_mask + else: + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + else: + negative_prompt = list(negative_prompt) + negative_prompt_embeds, negative_prompt_embeds_len = self.encode_prompt(negative_prompt, device) + negative_prompt_embeds_mask = _lens_to_mask( + negative_prompt_embeds_len, length=negative_prompt_embeds.shape[1] + ) + + latent_cond = torch.zeros(batch_size, duration, self.latent_dim, device=device, dtype=prompt_embeds.dtype) + latents = self.prepare_latents( + batch_size, duration, device, prompt_embeds.dtype, generator=generator, latents=latents + ) + if num_inference_steps < 1: + raise ValueError("num_inference_steps must be a positive integer.") + + sigmas = torch.linspace(1.0, 1.0 / num_inference_steps, num_inference_steps, dtype=torch.float32).tolist() + self.scheduler.set_timesteps(sigmas=sigmas, device=device) + self.scheduler.set_begin_index(0) + timesteps = self.scheduler.timesteps + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + curr_t = ( + (t / self.scheduler.config.num_train_timesteps).expand(batch_size).to(dtype=prompt_embeds.dtype) + ) + pred = self.transformer( + hidden_states=latents, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=text_mask, + timestep=curr_t, + attention_mask=mask, + latent_cond=latent_cond, + ).sample + if self.guidance_scale > 1.0: + null_pred = self.transformer( + hidden_states=latents, + encoder_hidden_states=negative_prompt_embeds, + encoder_attention_mask=negative_prompt_embeds_mask, + timestep=curr_t, + attention_mask=mask, + latent_cond=latent_cond, + ).sample + pred = null_pred + (pred - null_pred) * self.guidance_scale + latents = self.scheduler.step(pred, t, latents, return_dict=False)[0] + progress_bar.update() + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + if output_type == "latent": + waveform = latents + else: + waveform = self.vae.decode(latents.permute(0, 2, 1)).sample + if output_type == "np": + waveform = waveform.cpu().float().numpy() + + self.maybe_free_model_hooks() + + if not return_dict: + return (waveform,) + return AudioPipelineOutput(audios=waveform) diff --git a/diffusers/pipelines/longcat_image/__init__.py b/diffusers/pipelines/longcat_image/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e4bb0e5819c8bcecb0024e12cee72773edd21111 --- /dev/null +++ b/diffusers/pipelines/longcat_image/__init__.py @@ -0,0 +1,51 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa: F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_longcat_image"] = ["LongCatImagePipeline"] + _import_structure["pipeline_longcat_image_edit"] = ["LongCatImageEditPipeline"] + _import_structure["pipeline_output"] = ["LongCatImagePipelineOutput"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_longcat_image import LongCatImagePipeline + from .pipeline_longcat_image_edit import LongCatImageEditPipeline + from .pipeline_output import LongCatImagePipelineOutput + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/longcat_image/pipeline_longcat_image.py b/diffusers/pipelines/longcat_image/pipeline_longcat_image.py new file mode 100644 index 0000000000000000000000000000000000000000..4eaa858e41c1460acd96b34190e0049921a71210 --- /dev/null +++ b/diffusers/pipelines/longcat_image/pipeline_longcat_image.py @@ -0,0 +1,701 @@ +# Copyright 2025 MeiTuan LongCat-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +import re +from typing import Any + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import LongCatImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import LongCatImagePipelineOutput +from .system_messages import SYSTEM_PROMPT_EN, SYSTEM_PROMPT_ZH + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LongCatImagePipeline + + >>> pipe = LongCatImagePipeline.from_pretrained("meituan-longcat/LongCat-Image", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "一个年轻的亚裔女性，身穿黄色针织衫，搭配白色项链。她的双手放在膝盖上，表情恬静。背景是一堵粗糙的砖墙，午后的阳光温暖地洒在她身上，营造出一种宁静而温馨的氛围。镜头采用中距离视角，突出她的神态和服饰的细节。光线柔和地打在她的脸上，强调她的五官和饰品的质感，增加画面的层次感与亲和力。整个画面构图简洁，砖墙的纹理与阳光的光影效果相得益彰，突显出人物的优雅与从容。" + >>> image = pipe( + ... prompt, + ... height=768, + ... width=1344, + ... num_inference_steps=50, + ... guidance_scale=4.5, + ... generator=torch.Generator("cpu").manual_seed(43), + ... enable_cfg_renorm=True, + ... ).images[0] + >>> image.save("longcat_image.png") + ``` +""" + + +def get_prompt_language(prompt): + pattern = re.compile(r"[\u4e00-\u9fff]") + if bool(pattern.search(prompt)): + return "zh" + return "en" + + +def split_quotation(prompt, quote_pairs=None): + """ + Implement a regex-based string splitting algorithm that identifies delimiters defined by single or double quote + pairs. Examples:: + >>> prompt_en = "Please write 'Hello' on the blackboard for me." >>> print(split_quotation(prompt_en)) >>> # + output: [('Please write ', False), ("'Hello'", True), (' on the blackboard for me.', False)] + """ + word_internal_quote_pattern = re.compile(r"[a-zA-Z]+'[a-zA-Z]+") + matches_word_internal_quote_pattern = word_internal_quote_pattern.findall(prompt) + mapping_word_internal_quote = [] + + for i, word_src in enumerate(set(matches_word_internal_quote_pattern)): + word_tgt = "longcat_$##$_longcat" * (i + 1) + prompt = prompt.replace(word_src, word_tgt) + mapping_word_internal_quote.append([word_src, word_tgt]) + + if quote_pairs is None: + quote_pairs = [("'", "'"), ('"', '"'), ("‘", "’"), ("“", "”")] + pattern = "|".join([re.escape(q1) + r"[^" + re.escape(q1 + q2) + r"]*?" + re.escape(q2) for q1, q2 in quote_pairs]) + parts = re.split(f"({pattern})", prompt) + + result = [] + for part in parts: + for word_src, word_tgt in mapping_word_internal_quote: + part = part.replace(word_tgt, word_src) + if re.match(pattern, part): + if len(part): + result.append((part, True)) + else: + if len(part): + result.append((part, False)) + return result + + +def prepare_pos_ids(modality_id=0, type="text", start=(0, 0), num_token=None, height=None, width=None): + if type == "text": + assert num_token + if height or width: + print('Warning: The parameters of height and width will be ignored in "text" type.') + pos_ids = torch.zeros(num_token, 3) + pos_ids[..., 0] = modality_id + pos_ids[..., 1] = torch.arange(num_token) + start[0] + pos_ids[..., 2] = torch.arange(num_token) + start[1] + elif type == "image": + assert height and width + if num_token: + print('Warning: The parameter of num_token will be ignored in "image" type.') + pos_ids = torch.zeros(height, width, 3) + pos_ids[..., 0] = modality_id + pos_ids[..., 1] = pos_ids[..., 1] + torch.arange(height)[:, None] + start[0] + pos_ids[..., 2] = pos_ids[..., 2] + torch.arange(width)[None, :] + start[1] + pos_ids = pos_ids.reshape(height * width, 3) + else: + raise KeyError(f'Unknow type {type}, only support "text" or "image".') + return pos_ids + + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class LongCatImagePipeline(DiffusionPipeline, FromSingleFileMixin): + r""" + The pipeline for text-to-image generation. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + text_processor: Qwen2VLProcessor, + transformer: LongCatImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_processor=text_processor, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + + self.prompt_template_encode_prefix = "<|im_start|>system\nAs an image captioning expert, generate a descriptive text prompt based on an image content, suitable for input to a text-to-image model.<|im_end|>\n<|im_start|>user\n" + self.prompt_template_encode_suffix = "<|im_end|>\n<|im_start|>assistant\n" + self.default_sample_size = 128 + self.tokenizer_max_length = 512 + + def rewire_prompt(self, prompt, device): + prompt = [prompt] if isinstance(prompt, str) else prompt + all_text = [] + for each_prompt in prompt: + language = get_prompt_language(each_prompt) + if language == "zh": + question = SYSTEM_PROMPT_ZH + f"\n用户输入为：{each_prompt}\n改写后的prompt为：" + else: + question = SYSTEM_PROMPT_EN + f"\nUser Input: {each_prompt}\nRewritten prompt:" + message = [ + { + "role": "user", + "content": [ + {"type": "text", "text": question}, + ], + } + ] + # Preparation for inference + text = self.text_processor.apply_chat_template(message, tokenize=False, add_generation_prompt=True) + all_text.append(text) + + inputs = self.text_processor(text=all_text, padding=True, return_tensors="pt").to(self.text_encoder.device) + + generated_ids = self.text_encoder.generate(**inputs, max_new_tokens=self.tokenizer_max_length) + generated_ids.to(device) + generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)] + output_text = self.text_processor.batch_decode( + generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False + ) + rewrite_prompt = output_text + return rewrite_prompt + + def _encode_prompt(self, prompt: list[str]): + batch_all_tokens = [] + + for each_prompt in prompt: + all_tokens = [] + for clean_prompt_sub, matched in split_quotation(each_prompt): + if matched: + for sub_word in clean_prompt_sub: + tokens = self.tokenizer(sub_word, add_special_tokens=False)["input_ids"] + all_tokens.extend(tokens) + else: + tokens = self.tokenizer(clean_prompt_sub, add_special_tokens=False)["input_ids"] + all_tokens.extend(tokens) + + if len(all_tokens) > self.tokenizer_max_length: + logger.warning( + "Your input was truncated because `max_sequence_length` is set to " + f" {self.tokenizer_max_length} input token nums : {len(all_tokens)}" + ) + all_tokens = all_tokens[: self.tokenizer_max_length] + batch_all_tokens.append(all_tokens) + + text_tokens_and_mask = self.tokenizer.pad( + {"input_ids": batch_all_tokens}, + max_length=self.tokenizer_max_length, + padding="max_length", + return_attention_mask=True, + return_tensors="pt", + ) + + prefix_tokens = self.tokenizer(self.prompt_template_encode_prefix, add_special_tokens=False)["input_ids"] + suffix_tokens = self.tokenizer(self.prompt_template_encode_suffix, add_special_tokens=False)["input_ids"] + prefix_len = len(prefix_tokens) + suffix_len = len(suffix_tokens) + + prefix_tokens_mask = torch.tensor([1] * len(prefix_tokens), dtype=text_tokens_and_mask.attention_mask[0].dtype) + suffix_tokens_mask = torch.tensor([1] * len(suffix_tokens), dtype=text_tokens_and_mask.attention_mask[0].dtype) + + prefix_tokens = torch.tensor(prefix_tokens, dtype=text_tokens_and_mask.input_ids.dtype) + suffix_tokens = torch.tensor(suffix_tokens, dtype=text_tokens_and_mask.input_ids.dtype) + + batch_size = text_tokens_and_mask.input_ids.size(0) + + prefix_tokens_batch = prefix_tokens.unsqueeze(0).expand(batch_size, -1) + suffix_tokens_batch = suffix_tokens.unsqueeze(0).expand(batch_size, -1) + prefix_mask_batch = prefix_tokens_mask.unsqueeze(0).expand(batch_size, -1) + suffix_mask_batch = suffix_tokens_mask.unsqueeze(0).expand(batch_size, -1) + + input_ids = torch.cat((prefix_tokens_batch, text_tokens_and_mask.input_ids, suffix_tokens_batch), dim=-1) + attention_mask = torch.cat((prefix_mask_batch, text_tokens_and_mask.attention_mask, suffix_mask_batch), dim=-1) + + input_ids = input_ids.to(self.device) + attention_mask = attention_mask.to(self.device) + + text_output = self.text_encoder(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True) + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + prompt_embeds = text_output.hidden_states[-1].detach() + prompt_embeds = prompt_embeds[:, prefix_len:-suffix_len, :] + return prompt_embeds + + def encode_prompt( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int | None = 1, + prompt_embeds: torch.Tensor | None = None, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is None: + prompt_embeds = self._encode_prompt(prompt) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = prepare_pos_ids(modality_id=0, type="text", start=(0, 0), num_token=prompt_embeds.shape[1]).to( + self.device + ) + return prompt_embeds.to(self.device), text_ids + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = prepare_pos_ids( + modality_id=1, + type="image", + start=(self.tokenizer_max_length, self.tokenizer_max_length), + height=height // 2, + width=width // 2, + ).to(device) + + if latents is not None: + return latents.to(device=device, dtype=dtype), latent_image_ids + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device) + latents = latents.to(dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + def check_inputs( + self, prompt, height, width, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + @replace_example_docstring(EXAMPLE_DOC_STRING) + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + enable_cfg_renorm: bool | None = True, + cfg_renorm_min: float | None = 0.0, + enable_prompt_rewrite: bool | None = True, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance. + height (`int`, *optional*): + The height in pixels of the generated image. + width (`int`, *optional*): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process. If not defined, the scheduler's default schedule is + used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Classifier-free guidance scale. Values greater than 1 enable CFG. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A `torch.Generator` to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents to be used as inputs for image generation. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. If not provided, embeddings are generated from `prompt`. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Used when classifier-free guidance is enabled. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.LongCatImagePipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + Kwargs passed to the joint attention processor. + enable_cfg_renorm (`bool`, *optional*, defaults to `True`): + Whether to enable cfg_renorm. Enabling cfg_renorm will improve image quality, but it may lead to a + decrease in the stability of some image outputs. + cfg_renorm_min (`float`, *optional*, defaults to 0.0): + The minimum value of the cfg_renorm_scale range (0-1). `cfg_renorm_min = 1.0` disables renorm, while + `cfg_renorm_min = 0.0` widens the renorm range. + enable_prompt_rewrite (`bool`, *optional*, defaults to `True`): + Whether to enable prompt rewrite. + + Examples: + + Returns: + [`~pipelines.LongCatImagePipelineOutput`] or `tuple`: [`~pipelines.LongCatImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + if enable_prompt_rewrite: + prompt = self.rewire_prompt(prompt, device) + logger.info(f"Rewrite prompt {prompt}!") + + negative_prompt = "" if negative_prompt is None else negative_prompt + (prompt_embeds, text_ids) = self.encode_prompt( + prompt=prompt, prompt_embeds=prompt_embeds, num_images_per_prompt=num_images_per_prompt + ) + if self.do_classifier_free_guidance: + (negative_prompt_embeds, negative_text_ids) = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=num_images_per_prompt, + ) + + # 4. Prepare latent variables + num_channels_latents = 16 + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1.0 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + guidance = None + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred_text = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + return_dict=False, + )[0] + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_pred_uncond = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_image_ids, + return_dict=False, + )[0] + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if enable_cfg_renorm: + cond_norm = torch.norm(noise_pred_text, dim=-1, keepdim=True) + noise_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + scale = (cond_norm / (noise_norm + 1e-8)).clamp(min=cfg_renorm_min, max=1.0) + noise_pred = noise_pred * scale + else: + noise_pred = noise_pred_text + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + if latents.dtype != self.vae.dtype: + latents = latents.to(dtype=self.vae.dtype) + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return LongCatImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/longcat_image/pipeline_longcat_image_edit.py b/diffusers/pipelines/longcat_image/pipeline_longcat_image_edit.py new file mode 100644 index 0000000000000000000000000000000000000000..119de3946fbc6cc0c6775135733b4cabe244f88f --- /dev/null +++ b/diffusers/pipelines/longcat_image/pipeline_longcat_image_edit.py @@ -0,0 +1,758 @@ +# Copyright 2025 MeiTuan LongCat-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +import math +import re +from typing import Any + +import numpy as np +import PIL +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import LongCatImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import LongCatImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from PIL import Image + >>> import torch + >>> from diffusers import LongCatImageEditPipeline + + >>> pipe = LongCatImageEditPipeline.from_pretrained( + ... "meituan-longcat/LongCat-Image-Edit", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> prompt = "change the cat to dog." + >>> input_image = Image.open("test.jpg").convert("RGB") + >>> image = pipe( + ... input_image, + ... prompt, + ... num_inference_steps=50, + ... guidance_scale=4.5, + ... generator=torch.Generator("cpu").manual_seed(43), + ... ).images[0] + >>> image.save("longcat_image_edit.png") + ``` +""" + + +# Copied from diffusers.pipelines.longcat_image.pipeline_longcat_image.split_quotation +def split_quotation(prompt, quote_pairs=None): + """ + Implement a regex-based string splitting algorithm that identifies delimiters defined by single or double quote + pairs. Examples:: + >>> prompt_en = "Please write 'Hello' on the blackboard for me." >>> print(split_quotation(prompt_en)) >>> # + output: [('Please write ', False), ("'Hello'", True), (' on the blackboard for me.', False)] + """ + word_internal_quote_pattern = re.compile(r"[a-zA-Z]+'[a-zA-Z]+") + matches_word_internal_quote_pattern = word_internal_quote_pattern.findall(prompt) + mapping_word_internal_quote = [] + + for i, word_src in enumerate(set(matches_word_internal_quote_pattern)): + word_tgt = "longcat_$##$_longcat" * (i + 1) + prompt = prompt.replace(word_src, word_tgt) + mapping_word_internal_quote.append([word_src, word_tgt]) + + if quote_pairs is None: + quote_pairs = [("'", "'"), ('"', '"'), ("‘", "’"), ("“", "”")] + pattern = "|".join([re.escape(q1) + r"[^" + re.escape(q1 + q2) + r"]*?" + re.escape(q2) for q1, q2 in quote_pairs]) + parts = re.split(f"({pattern})", prompt) + + result = [] + for part in parts: + for word_src, word_tgt in mapping_word_internal_quote: + part = part.replace(word_tgt, word_src) + if re.match(pattern, part): + if len(part): + result.append((part, True)) + else: + if len(part): + result.append((part, False)) + return result + + +# Copied from diffusers.pipelines.longcat_image.pipeline_longcat_image.prepare_pos_ids +def prepare_pos_ids(modality_id=0, type="text", start=(0, 0), num_token=None, height=None, width=None): + if type == "text": + assert num_token + if height or width: + print('Warning: The parameters of height and width will be ignored in "text" type.') + pos_ids = torch.zeros(num_token, 3) + pos_ids[..., 0] = modality_id + pos_ids[..., 1] = torch.arange(num_token) + start[0] + pos_ids[..., 2] = torch.arange(num_token) + start[1] + elif type == "image": + assert height and width + if num_token: + print('Warning: The parameter of num_token will be ignored in "image" type.') + pos_ids = torch.zeros(height, width, 3) + pos_ids[..., 0] = modality_id + pos_ids[..., 1] = pos_ids[..., 1] + torch.arange(height)[:, None] + start[0] + pos_ids[..., 2] = pos_ids[..., 2] + torch.arange(width)[None, :] + start[1] + pos_ids = pos_ids.reshape(height * width, 3) + else: + raise KeyError(f'Unknow type {type}, only support "text" or "image".') + return pos_ids + + +# Copied from diffusers.pipelines.longcat_image.pipeline_longcat_image.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def calculate_dimensions(target_area, ratio): + width = math.sqrt(target_area * ratio) + height = width / ratio + + width = width if width % 16 == 0 else (width // 16 + 1) * 16 + height = height if height % 16 == 0 else (height // 16 + 1) * 16 + + width = int(width) + height = int(height) + + return width, height + + +class LongCatImageEditPipeline(DiffusionPipeline, FromSingleFileMixin): + r""" + The LongCat-Image-Edit pipeline for image editing. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + text_processor: Qwen2VLProcessor, + transformer: LongCatImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + text_processor=text_processor, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.image_processor_vl = text_processor.image_processor + + self.image_token = "<|image_pad|>" + self.prompt_template_encode_prefix = "<|im_start|>system\nAs an image editing expert, first analyze the content and attributes of the input image(s). Then, based on the user's editing instructions, clearly and precisely determine how to modify the given image(s), ensuring that only the specified parts are altered and all other aspects remain consistent with the original(s).<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>" + self.prompt_template_encode_suffix = "<|im_end|>\n<|im_start|>assistant\n" + self.default_sample_size = 128 + self.tokenizer_max_length = 512 + + def _encode_prompt(self, prompt, image): + raw_vl_input = self.image_processor_vl(images=image, return_tensors="pt") + pixel_values = raw_vl_input["pixel_values"] + image_grid_thw = raw_vl_input["image_grid_thw"] + all_tokens = [] + for clean_prompt_sub, matched in split_quotation(prompt[0]): + if matched: + for sub_word in clean_prompt_sub: + tokens = self.tokenizer(sub_word, add_special_tokens=False)["input_ids"] + all_tokens.extend(tokens) + else: + tokens = self.tokenizer(clean_prompt_sub, add_special_tokens=False)["input_ids"] + all_tokens.extend(tokens) + + if len(all_tokens) > self.tokenizer_max_length: + logger.warning( + "Your input was truncated because `max_sequence_length` is set to " + f" {self.tokenizer_max_length} input token nums : {len(len(all_tokens))}" + ) + all_tokens = all_tokens[: self.tokenizer_max_length] + + text_tokens_and_mask = self.tokenizer.pad( + {"input_ids": [all_tokens]}, + max_length=self.tokenizer_max_length, + padding="max_length", + return_attention_mask=True, + return_tensors="pt", + ) + + text = self.prompt_template_encode_prefix + + merge_length = self.image_processor_vl.merge_size**2 + while self.image_token in text: + num_image_tokens = image_grid_thw.prod() // merge_length + text = text.replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1) + text = text.replace("<|placeholder|>", self.image_token) + + prefix_tokens = self.tokenizer(text, add_special_tokens=False)["input_ids"] + suffix_tokens = self.tokenizer(self.prompt_template_encode_suffix, add_special_tokens=False)["input_ids"] + + vision_start_token_id = self.tokenizer.convert_tokens_to_ids("<|vision_start|>") + prefix_len = prefix_tokens.index(vision_start_token_id) + suffix_len = len(suffix_tokens) + + prefix_tokens_mask = torch.tensor([1] * len(prefix_tokens), dtype=text_tokens_and_mask.attention_mask[0].dtype) + suffix_tokens_mask = torch.tensor([1] * len(suffix_tokens), dtype=text_tokens_and_mask.attention_mask[0].dtype) + + prefix_tokens = torch.tensor(prefix_tokens, dtype=text_tokens_and_mask.input_ids.dtype) + suffix_tokens = torch.tensor(suffix_tokens, dtype=text_tokens_and_mask.input_ids.dtype) + + input_ids = torch.cat((prefix_tokens, text_tokens_and_mask.input_ids[0], suffix_tokens), dim=-1) + attention_mask = torch.cat( + (prefix_tokens_mask, text_tokens_and_mask.attention_mask[0], suffix_tokens_mask), dim=-1 + ) + + input_ids = input_ids.unsqueeze(0).to(self.device) + attention_mask = attention_mask.unsqueeze(0).to(self.device) + + pixel_values = pixel_values.to(self.device) + image_grid_thw = image_grid_thw.to(self.device) + + text_output = self.text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + pixel_values=pixel_values, + image_grid_thw=image_grid_thw, + output_hidden_states=True, + ) + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + prompt_embeds = text_output.hidden_states[-1].detach() + prompt_embeds = prompt_embeds[:, prefix_len:-suffix_len, :] + return prompt_embeds + + def encode_prompt( + self, + prompt: list[str] = None, + image: torch.Tensor | None = None, + num_images_per_prompt: int | None = 1, + prompt_embeds: torch.Tensor | None = None, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is None: + prompt_embeds = self._encode_prompt(prompt, image) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + text_ids = prepare_pos_ids(modality_id=0, type="text", start=(0, 0), num_token=prompt_embeds.shape[1]).to( + self.device + ) + return prompt_embeds, text_ids + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax") + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + def prepare_latents( + self, + image, + batch_size, + num_channels_latents, + height, + width, + dtype, + prompt_embeds_length, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + image_latents, image_latents_ids = None, None + + if image is not None: + image = image.to(device=self.device, dtype=dtype) + + if image.shape[1] != self.vae.config.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latents = self._pack_latents(image_latents, batch_size, num_channels_latents, height, width) + + image_latents_ids = prepare_pos_ids( + modality_id=2, + type="image", + start=(prompt_embeds_length, prompt_embeds_length), + height=height // 2, + width=width // 2, + ).to(device, dtype=torch.float64) + + shape = (batch_size, num_channels_latents, height, width) + latents_ids = prepare_pos_ids( + modality_id=1, + type="image", + start=(prompt_embeds_length, prompt_embeds_length), + height=height // 2, + width=width // 2, + ).to(device) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents, latents_ids, image_latents_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + def check_inputs( + self, prompt, height, width, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None: + if isinstance(prompt, str): + pass + elif isinstance(prompt, list) and len(prompt) == 1: + pass + else: + raise ValueError( + f"`prompt` must be a `str` or a `list` of length 1, but is {prompt} (type: {type(prompt)})" + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + @replace_example_docstring(EXAMPLE_DOC_STRING) + @torch.no_grad() + def __call__( + self, + image: PIL.Image.Image | None = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`PIL.Image.Image`, *optional*): + The input image to edit. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process. If not defined, the scheduler's default schedule is + used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Classifier-free guidance scale. Values greater than 1 enable CFG. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A `torch.Generator` to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents to be used as inputs for image generation. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. If not provided, embeddings are generated from `prompt`. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Used when classifier-free guidance is enabled. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.LongCatImagePipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + Kwargs passed to the joint attention processor. + + Examples: + + Returns: + [`~pipelines.LongCatImagePipelineOutput`] or `tuple`: [`~pipelines.LongCatImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + image_size = image[0].size if isinstance(image, list) else image.size + calculated_width, calculated_height = calculate_dimensions(1024 * 1024, image_size[0] * 1.0 / image_size[1]) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + calculated_height, + calculated_width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Preprocess image + if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): + image = self.image_processor.resize(image, calculated_height, calculated_width) + prompt_image = self.image_processor.resize(image, calculated_height // 2, calculated_width // 2) + image = self.image_processor.preprocess(image, calculated_height, calculated_width) + + negative_prompt = "" if negative_prompt is None else negative_prompt + (prompt_embeds, text_ids) = self.encode_prompt( + prompt=prompt, image=prompt_image, prompt_embeds=prompt_embeds, num_images_per_prompt=num_images_per_prompt + ) + if self.do_classifier_free_guidance: + (negative_prompt_embeds, negative_text_ids) = self.encode_prompt( + prompt=negative_prompt, + image=prompt_image, + prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=num_images_per_prompt, + ) + + # 4. Prepare latent variables + num_channels_latents = 16 + latents, image_latents, latents_ids, image_latents_ids = self.prepare_latents( + image, + batch_size * num_images_per_prompt, + num_channels_latents, + calculated_height, + calculated_width, + prompt_embeds.dtype, + prompt_embeds.shape[1], + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1.0 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + guidance = None + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + if image is not None: + latent_image_ids = torch.cat([latents_ids, image_latents_ids], dim=0) + else: + latent_image_ids = latents_ids + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + + timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred_text = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + return_dict=False, + )[0] + noise_pred_text = noise_pred_text[:, :image_seq_len] + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_pred_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_image_ids, + return_dict=False, + )[0] + noise_pred_uncond = noise_pred_uncond[:, :image_seq_len] + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + else: + noise_pred = noise_pred_text + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, calculated_height, calculated_width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + if latents.dtype != self.vae.dtype: + latents = latents.to(dtype=self.vae.dtype) + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return LongCatImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/longcat_image/pipeline_output.py b/diffusers/pipelines/longcat_image/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..290178eb1a8859b682e64f44f582f86a5b236ab1 --- /dev/null +++ b/diffusers/pipelines/longcat_image/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from diffusers.utils import BaseOutput + + +@dataclass +class LongCatImagePipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image, np.ndarray] diff --git a/diffusers/pipelines/longcat_image/system_messages.py b/diffusers/pipelines/longcat_image/system_messages.py new file mode 100644 index 0000000000000000000000000000000000000000..b8b2318e4e813970ebba7e5da4a07c26f238f4e8 --- /dev/null +++ b/diffusers/pipelines/longcat_image/system_messages.py @@ -0,0 +1,142 @@ +SYSTEM_PROMPT_EN = """ +You are a prompt engineering expert for text-to-image models. Since text-to-image models have limited capabilities in +understanding user prompts, you need to identify the core theme and intent of the user's input and improve the model's +understanding accuracy and generation quality through optimization and rewriting. The rewrite must strictly retain all +information from the user's original prompt without deleting or distorting any details. Specific requirements are as +follows: +1. The rewrite must not affect any information expressed in the user's original prompt; the rewritten prompt should use + coherent natural language, avoid low-information redundant descriptions, and keep the rewritten prompt length as + concise as possible. +2. Ensure consistency between input and output languages: Chinese input yields Chinese output, and English input yields + English output. The rewritten token count should not exceed 512. +3. The rewritten description should further refine subject characteristics and aesthetic techniques appearing in the + original prompt, such as lighting and textures. +4. If the original prompt does not specify an image style, ensure the rewritten prompt uses a **realistic photography + style**. If the user specifies a style, retain the user's style. +5. When the original prompt requires reasoning to clarify user intent, use logical reasoning based on world knowledge + to convert vague abstract descriptions into specific tangible objects (e.g., convert "the tallest animal" to "a + giraffe"). +6. When the original prompt requires text generation, please use double quotes to enclose the text part (e.g., `"50% + OFF"`). +7. When the original prompt requires generating text-heavy scenes like webpages, logos, UIs, or posters, and no + specific text content is specified, you need to infer appropriate text content and enclose it in double quotes. For + example, if the user inputs: "A tourism flyer with a grassland theme," it should be rewritten as: "A tourism flyer + with the image title 'Grassland'." +8. When negative words exist in the original prompt, ensure the rewritten prompt does not contain negative words. For + example, "a lakeside without boats" should be rewritten such that the word "boat" does not appear at all. +9. Except for text content explicitly requested by the user, **adding any extra text content is prohibited**. +Here are examples of rewrites for different types of prompts: # Examples (Few-Shot Learning) + 1. User Input: An animal with nine lives. + Rewrite Output: A cat bathed in soft sunlight, its fur soft and glossy. The background is a comfortable home + environment with light from the window filtering through curtains, creating a warm light and shadow effect. The + shot uses a medium distance perspective to highlight the cat's leisurely and stretched posture. Light cleverly hits + the cat's face, emphasizing its spirited eyes and delicate whiskers, adding depth and affinity to the image. + 2. User Input: Create an anime-style tourism flyer with a grassland theme. + Rewrite Output: In the lower right of the center, a short-haired girl sits sideways on a gray, irregularly shaped + rock. She wears a white short-sleeved dress and brown flat shoes, holding a bunch of small white flowers in her + left hand, smiling with her legs hanging naturally. The girl has dark brown shoulder-length hair with bangs + covering her forehead, brown eyes, and a slightly open mouth. The rock surface has textures of varying depths. To + the girl's left and front is lush grass, with long, yellow-green blades, some glowing golden in the sunlight. The + grass extends into the distance, forming rolling green hills that fade in color as they recede. The sky occupies + the upper half of the picture, pale blue dotted with a few fluffy white clouds. In the upper left corner, there is + a line of text in italic, dark green font reading "Explore Nature's Peace". Colors are dominated by green, blue, + and yellow, fluid lines, and distinct light and shadow contrast, creating a quiet and comfortable atmosphere. + 3. User Input: A Christmas sale poster with a red background, promoting a Buy 1 Get 1 Free milk tea offer. + Rewrite Output: The poster features an overall red tone, embellished with white snowflake patterns on the top and + left side. The upper right features a bunch of holly leaves with red berries and a pine cone. In the upper center, + golden 3D text reads "Christmas Heartwarming Feedback" centered, along with red bold text "Buy 1 Get 1". Below, two + transparent cups filled with bubble tea are placed side by side; the tea is light brown with dark brown pearls + scattered at the bottom and middle. Below the cups, white snow piles up, decorated with pine branches, red berries, + and pine cones. A blurry Christmas tree is faintly visible in the lower right corner. The image has high clarity, + accurate text content, a unified design style, a prominent Christmas theme, and a reasonable layout, providing + strong visual appeal. + 4. User Input: A woman indoors shot in natural light, smiling with arms crossed, showing a relaxed and confident + posture. + Rewrite Output: The image features a young Asian woman with long dark brown hair naturally falling over her + shoulders, with some strands illuminated by light, showing a soft sheen. Her features are delicate, with long + eyebrows, bright and spirited dark brown eyes looking directly at the camera, revealing peace and confidence. She + has a high nose bridge, full lips with nude lipstick, and corners of the mouth slightly raised in a faint smile. + Her skin is fair, with cheeks and collarbones illuminated by warm light, showing a healthy ruddiness. She wears a + black spaghetti strap tank top revealing graceful collarbone lines, and a thin gold necklace with small beads and + metal bars glinting in the light. Her outer layer is a beige knitted cardigan, soft in texture with visible + knitting patterns on the sleeves. Her arms are crossed over her chest, hands covered by the cardigan sleeves, in a + relaxed posture. The background is a pure dark brown without extra decoration, making the figure the absolute + focus. The figure is located in the center of the frame. Light enters from the upper right, creating bright spots + on her left cheek, neck, and collarbone, while the right side is slightly shadowed, creating a three-dimensional + and soft tone. Image details are clear, showcasing skin texture, hair, and clothing materials well. Colors are + dominated by warm tones, with the combination of beige and dark brown creating a warm and comfortable atmosphere. + The overall style is natural, elegant, and artistic. + 5. User Input: Create a series of images showing the growth process of an apple from seed to fruit. The series should + include four stages: 1. Sowing, 2. Seedling growth, 3. Plant maturity, 4. Fruit harvesting. + Rewrite Output: A 4-panel exquisite illustration depicting the growth process of an apple, capturing each stage + precisely and clearly. 1. "Sowing": A close-up shot of a hand gently placing a small apple seed into fertile dark + soil, with visible soil texture and the seed's smooth surface. The background is a soft-focus garden dotted with + green leaves and sunlight filtering through. 2. "Seedling Growth": A young apple sapling breaks through the soil, + stretching tender green leaves toward the sky. The scene is set in a vibrant garden illuminated by warm golden + light, highlighting the seedling's delicate structure. 3. "Plant Maturity": A mature apple tree, lush with branches + and leaves, covered in tender green foliage and developing small apples. The background is a vibrant orchard under + a clear blue sky, with dappled sunlight creating a peaceful atmosphere. 4. "Fruit Harvesting": A hand reaches into + the tree to pick a ripe red apple, its smooth skin glistening in the sun. The scene shows the abundance of the + orchard, with baskets of apples in the background, giving a sense of fulfillment. Each illustration uses a + realistic style, focusing on details and harmonious colors to showcase the natural beauty and development of the + apple's life cycle. + 6. User Input: If 1 represents red, 2 represents green, 3 represents purple, and 4 represents yellow, please generate + a four-color rainbow based on this rule. The color order from top to bottom is 3142. + Rewrite Output: The image consists of four horizontally arranged colored stripes, ordered from top to bottom as + purple, red, yellow, and green. A white number is centered on each stripe. The top purple stripe features the + number "3", the red stripe below it has the number "1", the yellow stripe further down has the number "4", and the + bottom green stripe has the number "2". All numbers use a sans-serif font in pure white, forming a sharp contrast + with the background colors to ensure good readability. The stripes have high color saturation and a slight texture. + The overall layout is simple and clear, with distinct visual effects and no extra decorative elements, emphasizing + the numerical information. The image is high definition, with accurate colors and a consistent style, offering + strong visual appeal. + 7. User Input: A stone tablet carved with "Guan Guan Ju Jiu, On the River Isle", natural light, background is a + Chinese garden. + Rewrite Output: An ancient stone tablet carved with "Guan Guan Ju Jiu, On the River Isle", the surface covered with + traces of time, the writing clear and deep. Natural light falls from above, softly illuminating every detail of the + stone tablet and enhancing its sense of history. The background is an elegant Chinese garden featuring lush bamboo + forests, winding paths, and quiet pools, creating a serene and distant atmosphere. The overall picture uses a + realistic style with rich details and natural light and shadow effects, highlighting the cultural heritage of the + stone tablet and the classical beauty of the garden. +# Output Format Please directly output the rewritten and optimized Prompt content. Do not include any explanatory +language or JSON formatting, and do not add opening or closing quotes yourself.""" + + +SYSTEM_PROMPT_ZH = """ +你是一名文生图模型的prompt +engineering专家。由于文生图模型对用户prompt的理解能力有限，你需要识别用户输入的核心主题和意图，并通过优化改写提升模型的理解准确性和生成质量。改写必须严格保留用户原始prompt的所有信息，不得删减或曲解任何细节。 +具体要求如下： +1. 改写不能影响用户原始prompt里表达的任何信息，改写后的prompt应该使用连贯的自然语言表达,不要出现低信息量的冗余描述，尽可能保持改写后prompt长度精简。 +2. 请确保输入和输出的语言类型一致，中文输入中文输出，英文输入英文输出，改写后的token数量不要超过512个; +3. 改写后的描述应当进一步完善原始prompt中出现的主体特征、美学技巧，如打光、纹理等； +4. 如果原始prompt没有指定图片风格时，确保改写后的prompt使用真实摄影风格，如果用户指定了图片风格，则保留用户风格； +5. 当原始prompt需要推理才能明确用户意图时，根据世界知识进行适当逻辑推理，将模糊抽象描述转化为具体指向事物（例：将"最高的动物"转化为"一头长颈鹿"）。 +6. 当原始prompt需要生成文字时，请使用双引号圈定文字部分，例：`"限时5折"`）。 +7. 当原始prompt需要生成网页、logo、ui、海报等文字场景时，且没有指定具体的文字内容时，需要推断出合适的文字内容，并使用双引号圈定，如用户输入：一个旅游宣传单，以草原为主题。应该改写成：一个旅游宣传单，图片标题为“草原”。 +8. 当原始prompt中存在否定词时，需要确保改写后的prompt不存在否定词，如没有船的湖边，改写后的prompt不能出现船这个词汇。 +9. 除非用户指定生成品牌logo，否则不要增加额外的品牌logo. +10. 除了用户明确要求书写的文字内容外，**禁止增加任何额外的文字内容**。 +以下是针对不同类型prompt改写的示例： + +# Examples (Few-Shot Learning) + 1. 用户输入: 九条命的动物。 + 改写输出: + 一只猫，被柔和的阳光笼罩着，毛发柔软而富有光泽。背景是一个舒适的家居环境，窗外的光线透过窗帘，形成温馨的光影效果。镜头采用中距离视角，突出猫悠闲舒展的姿态。光线巧妙地打在猫的脸部，强调它灵动的眼睛和精致的胡须，增加画面的层次感与亲和力。 + 2. 用户输入: 制作一个动画风格的旅游宣传单，以草原为主题。 + 改写输出: + 画面中央偏右下角，一个短发女孩侧身坐在灰色的不规则形状岩石上，她穿着白色短袖连衣裙和棕色平底鞋，左手拿着一束白色小花，面带微笑，双腿自然垂下。女孩的头发为深棕色，齐肩短发，刘海覆盖额头，眼睛呈棕色，嘴巴微张。岩石表面有深浅不一的纹理。女孩的左侧和前方是茂盛的草地，草叶细长，呈黄绿色，部分草叶在阳光下泛着金色的光芒，仿佛被阳光照亮。草地向远处延伸，形成连绵起伏的绿色山丘，山丘的颜色由近及远逐渐变浅。天空占据了画面的上半部分，呈淡蓝色，点缀着几朵白色蓬松的云彩。画面的左上角有一行文字，文字内容是斜体、深绿色的“Explore + Nature's Peace”。色彩以绿色、蓝色和黄色为主，线条流畅，光影明暗对比明显，营造出一种宁静、舒适的氛围。 + 3. 用户输入: 一张以红色为背景的圣诞节促销海报，主要宣传奶茶买一送一的优惠活动。 + 改写输出: 海报整体呈现红色调，上方和左侧点缀着白色雪花图案，右上方有一束冬青叶和红色浆果，以及一个松果。海报中央偏上位置，金色立体字样“圣诞节 + 暖心回馈”居中排列，和红色粗体字“买1送1”。海报下方，两个装满珍珠奶茶的透明杯子并排摆放，杯中奶茶呈浅棕色，底部和中间散布着深棕色珍珠。杯子下方，堆积着白色雪花，雪花上装饰着松枝、红色浆果和松果。右下角隐约可见一棵模糊的圣诞树。图片清晰度高，文字内容准确，整体设计风格统一，圣诞主题突出，排版布局合理，具有较强的视觉吸引力。 + 4. 用户输入: 一位女性在室内以自然光线拍摄，她面带微笑，双臂交叉，展现出轻松自信的姿态。 + 改写输出: + 画面中是一位年轻的亚洲女性，她拥有深棕色的长发，发丝自然地垂落在双肩，部分发丝被光线照亮，呈现出柔和的光泽。她的五官精致，眉毛修长，眼睛明亮有神，瞳孔呈深棕色，眼神直视镜头，流露出平和与自信。鼻梁挺拔，嘴唇丰满，涂有裸色系唇膏，嘴角微微上扬，展现出浅浅的微笑。她的肤色白皙，脸颊和锁骨处被暖色调的光线照亮，呈现出健康的红润感。她穿着一件黑色的细吊带背心，肩带纤细，露出优美的锁骨线条。脖颈上佩戴着一条金色的细项链，项链由小珠子和几个细长的金属条组成，在光线下闪烁着光泽。她的外搭是一件米黄色的针织开衫，材质柔软，袖子部分有明显的针织纹理。她双臂交叉在胸前，双手被开衫的袖子覆盖，姿态放松。背景是纯粹的深棕色，没有多余的装饰，使得人物成为画面的绝对焦点。人物位于画面中央。光线从画面的右上方射入，在人物的左侧脸颊、脖颈和锁骨处形成明亮的光斑，右侧则略显阴影，营造出立体感和柔和的影调。图像细节清晰，人物的皮肤纹理、发丝以及衣物材质都得到了很好的展现。色彩以暖色调为主，米黄色和深棕色的搭配营造出温馨舒适的氛围。整体呈现出一种自然、优雅且富有亲和力的艺术风格。 + 5. 用户输入：创作一系列图片，展现苹果从种子到结果的生长过程。该系列图片应包含以下四个阶段：1. 播种，2. 幼苗生长，3. 植物成熟，4. 果实采摘。 + 改写输出：一个4宫格的精美插图，描绘苹果的生长过程，精确清晰地捕捉每个阶段。1.“播种”：特写镜头，一只手轻轻地将一颗小小的苹果种子放入肥沃的深色土壤中，土壤的纹理和种子光滑的表面清晰可见。背景是花园的柔焦画面，点缀着绿色的树叶和透过树叶洒下的阳光。2.“幼苗生长”：一棵幼小的苹果树苗破土而出，嫩绿的叶子向天空舒展。场景设定在一个生机勃勃的花园中，温暖的金光照亮了它。幼苗的纤细结构。3.“植物的成熟”：一棵成熟的苹果树，枝繁叶茂，挂满了嫩绿的叶子和正在萌发的小苹果。背景是一片生机勃勃的果园，湛蓝的天空下，斑驳的阳光营造出宁静祥和的氛围。4.“采摘果实”：一只手伸向树上，摘下一个成熟的红苹果，苹果光滑的果皮在阳光下闪闪发光。画面展现了果园的丰收景象，背景中摆放着一篮篮的苹果，给人一种圆满满足的感觉。每幅插图都采用写实风格，注重细节，色彩和谐，展现了苹果生命周期的自然之美和发展过程。 + 6. 用户输入： 如果1代表红色，2代表绿色，3代表紫色，4代表黄色，请按照此规则生成四色彩虹。它的颜色顺序从上到下是3142 + 改写输出：图片由四个水平排列的彩色条纹组成，从上到下依次为紫色、红色、黄色和绿色。每个条纹上都居中放置一个白色数字。最上方的紫色条纹上是数字“3”，其下方红色条纹上是数字“1”，再下方黄色条纹上是数字“4”，最下方的绿色条纹上是数字“2”。所有数字均采用无衬线字体，颜色为纯白色，与背景色形成鲜明对比，确保了良好的可读性。条纹的颜色饱和度高，且带有轻微的纹理感，整体排版简洁明了，视觉效果清晰，没有多余的装饰元素，强调了数字信息本身。图片整体清晰度高，色彩准确，风格一致，具有较强的视觉吸引力。 + 7. 用户输入：石碑上刻着“关关雎鸠，在河之洲”，自然光照，背景是中式园林 + 改写输出：一块古老的石碑上刻着“关关雎鸠，在河之洲”，石碑表面布满岁月的痕迹，字迹清晰而深刻。自然光线从上方洒下，柔和地照亮石碑的每一个细节，增强了其历史感。背景是一座典雅的中式园林，园林中有翠绿的竹林、蜿蜒的小径和静谧的水池，营造出一种宁静而悠远的氛围。整体画面采用写实风格，细节丰富，光影效果自然，突出了石碑的文化底蕴和园林的古典美。 +# 输出格式 请直接输出改写优化后的 Prompt 内容，不要包含任何解释性语言或 JSON 格式，不要自行添加开头或结尾的引号。 +""" diff --git a/diffusers/pipelines/ltx/__init__.py b/diffusers/pipelines/ltx/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..05117d35d3b4f1746610ec0d273d76f715215cfb --- /dev/null +++ b/diffusers/pipelines/ltx/__init__.py @@ -0,0 +1,58 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["modeling_latent_upsampler"] = ["LTXLatentUpsamplerModel"] + _import_structure["pipeline_ltx"] = ["LTXPipeline"] + _import_structure["pipeline_ltx_condition"] = ["LTXConditionPipeline"] + _import_structure["pipeline_ltx_i2v_long_multi_prompt"] = ["LTXI2VLongMultiPromptPipeline"] + _import_structure["pipeline_ltx_image2video"] = ["LTXImageToVideoPipeline"] + _import_structure["pipeline_ltx_latent_upsample"] = ["LTXLatentUpsamplePipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .modeling_latent_upsampler import LTXLatentUpsamplerModel + from .pipeline_ltx import LTXPipeline + from .pipeline_ltx_condition import LTXConditionPipeline + from .pipeline_ltx_i2v_long_multi_prompt import LTXI2VLongMultiPromptPipeline + from .pipeline_ltx_image2video import LTXImageToVideoPipeline + from .pipeline_ltx_latent_upsample import LTXLatentUpsamplePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/ltx/modeling_latent_upsampler.py b/diffusers/pipelines/ltx/modeling_latent_upsampler.py new file mode 100644 index 0000000000000000000000000000000000000000..f579cf00dbe7a314aca365b503e2ac22756d1ce0 --- /dev/null +++ b/diffusers/pipelines/ltx/modeling_latent_upsampler.py @@ -0,0 +1,186 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin + + +class ResBlock(torch.nn.Module): + def __init__(self, channels: int, mid_channels: int | None = None, dims: int = 3): + super().__init__() + if mid_channels is None: + mid_channels = channels + + Conv = torch.nn.Conv2d if dims == 2 else torch.nn.Conv3d + + self.conv1 = Conv(channels, mid_channels, kernel_size=3, padding=1) + self.norm1 = torch.nn.GroupNorm(32, mid_channels) + self.conv2 = Conv(mid_channels, channels, kernel_size=3, padding=1) + self.norm2 = torch.nn.GroupNorm(32, channels) + self.activation = torch.nn.SiLU() + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + residual = hidden_states + hidden_states = self.conv1(hidden_states) + hidden_states = self.norm1(hidden_states) + hidden_states = self.activation(hidden_states) + hidden_states = self.conv2(hidden_states) + hidden_states = self.norm2(hidden_states) + hidden_states = self.activation(hidden_states + residual) + return hidden_states + + +class PixelShuffleND(torch.nn.Module): + def __init__(self, dims, upscale_factors=(2, 2, 2)): + super().__init__() + + self.dims = dims + self.upscale_factors = upscale_factors + + if dims not in [1, 2, 3]: + raise ValueError("dims must be 1, 2, or 3") + + def forward(self, x): + if self.dims == 3: + # spatiotemporal: b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3) + return ( + x.unflatten(1, (-1, *self.upscale_factors[:3])) + .permute(0, 1, 5, 2, 6, 3, 7, 4) + .flatten(6, 7) + .flatten(4, 5) + .flatten(2, 3) + ) + elif self.dims == 2: + # spatial: b (c p1 p2) h w -> b c (h p1) (w p2) + return ( + x.unflatten(1, (-1, *self.upscale_factors[:2])).permute(0, 1, 4, 2, 5, 3).flatten(4, 5).flatten(2, 3) + ) + elif self.dims == 1: + # temporal: b (c p1) f h w -> b c (f p1) h w + return x.unflatten(1, (-1, *self.upscale_factors[:1])).permute(0, 1, 3, 2, 4, 5).flatten(2, 3) + + +class LTXLatentUpsamplerModel(ModelMixin, ConfigMixin): + """ + Model to spatially upsample VAE latents. + + Args: + in_channels (`int`, defaults to `128`): + Number of channels in the input latent + mid_channels (`int`, defaults to `512`): + Number of channels in the middle layers + num_blocks_per_stage (`int`, defaults to `4`): + Number of ResBlocks to use in each stage (pre/post upsampling) + dims (`int`, defaults to `3`): + Number of dimensions for convolutions (2 or 3) + spatial_upsample (`bool`, defaults to `True`): + Whether to spatially upsample the latent + temporal_upsample (`bool`, defaults to `False`): + Whether to temporally upsample the latent + """ + + @register_to_config + def __init__( + self, + in_channels: int = 128, + mid_channels: int = 512, + num_blocks_per_stage: int = 4, + dims: int = 3, + spatial_upsample: bool = True, + temporal_upsample: bool = False, + ): + super().__init__() + + self.in_channels = in_channels + self.mid_channels = mid_channels + self.num_blocks_per_stage = num_blocks_per_stage + self.dims = dims + self.spatial_upsample = spatial_upsample + self.temporal_upsample = temporal_upsample + + ConvNd = torch.nn.Conv2d if dims == 2 else torch.nn.Conv3d + + self.initial_conv = ConvNd(in_channels, mid_channels, kernel_size=3, padding=1) + self.initial_norm = torch.nn.GroupNorm(32, mid_channels) + self.initial_activation = torch.nn.SiLU() + + self.res_blocks = torch.nn.ModuleList([ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)]) + + if spatial_upsample and temporal_upsample: + self.upsampler = torch.nn.Sequential( + torch.nn.Conv3d(mid_channels, 8 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(3), + ) + elif spatial_upsample: + self.upsampler = torch.nn.Sequential( + torch.nn.Conv2d(mid_channels, 4 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(2), + ) + elif temporal_upsample: + self.upsampler = torch.nn.Sequential( + torch.nn.Conv3d(mid_channels, 2 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(1), + ) + else: + raise ValueError("Either spatial_upsample or temporal_upsample must be True") + + self.post_upsample_res_blocks = torch.nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + self.final_conv = ConvNd(mid_channels, in_channels, kernel_size=3, padding=1) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + batch_size, num_channels, num_frames, height, width = hidden_states.shape + + if self.dims == 2: + hidden_states = hidden_states.permute(0, 2, 1, 3, 4).flatten(0, 1) + hidden_states = self.initial_conv(hidden_states) + hidden_states = self.initial_norm(hidden_states) + hidden_states = self.initial_activation(hidden_states) + + for block in self.res_blocks: + hidden_states = block(hidden_states) + + hidden_states = self.upsampler(hidden_states) + + for block in self.post_upsample_res_blocks: + hidden_states = block(hidden_states) + + hidden_states = self.final_conv(hidden_states) + hidden_states = hidden_states.unflatten(0, (batch_size, -1)).permute(0, 2, 1, 3, 4) + else: + hidden_states = self.initial_conv(hidden_states) + hidden_states = self.initial_norm(hidden_states) + hidden_states = self.initial_activation(hidden_states) + + for block in self.res_blocks: + hidden_states = block(hidden_states) + + if self.temporal_upsample: + hidden_states = self.upsampler(hidden_states) + hidden_states = hidden_states[:, :, 1:, :, :] + else: + hidden_states = hidden_states.permute(0, 2, 1, 3, 4).flatten(0, 1) + hidden_states = self.upsampler(hidden_states) + hidden_states = hidden_states.unflatten(0, (batch_size, -1)).permute(0, 2, 1, 3, 4) + + for block in self.post_upsample_res_blocks: + hidden_states = block(hidden_states) + + hidden_states = self.final_conv(hidden_states) + + return hidden_states diff --git a/diffusers/pipelines/ltx/pipeline_ltx.py b/diffusers/pipelines/ltx/pipeline_ltx.py new file mode 100644 index 0000000000000000000000000000000000000000..ce9177547c5243eb964cb21bfb9c90d1bced4df4 --- /dev/null +++ b/diffusers/pipelines/ltx/pipeline_ltx.py @@ -0,0 +1,857 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import T5EncoderModel, T5TokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTXVideo +from ...models.transformers import LTXVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import LTXPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTXPipeline + >>> from diffusers.utils import export_to_video + + >>> pipe = LTXPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage" + >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted" + + >>> video = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... width=704, + ... height=480, + ... num_frames=161, + ... num_inference_steps=50, + ... ).frames[0] + >>> export_to_video(video, "output.mp4", fps=24) + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTXPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin): + r""" + Pipeline for text-to-video generation. + + Reference: https://github.com/Lightricks/LTX-Video + + Args: + transformer ([`LTXVideoTransformer3DModel`]): + Conditional Transformer architecture to denoise the encoded video latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLLTXVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTXVideo, + text_encoder: T5EncoderModel, + tokenizer: T5TokenizerFast, + transformer: LTXVideoTransformer3DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128 + ) + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(text_input_ids.to(device))[0] + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128 + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + @staticmethod + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + def prepare_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 704, + num_frames: int = 161, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + height = height // self.vae_spatial_compression_ratio + width = width // self.vae_spatial_compression_ratio + num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + + shape = (batch_size, num_channels_latents, num_frames, height, width) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 704, + num_frames: int = 161, + frame_rate: int = 25, + num_inference_steps: int = 50, + timesteps: list[int] = None, + guidance_scale: float = 3, + guidance_rescale: float = 0.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 128, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + height (`int`, defaults to `512`): + The height in pixels of the generated image. This is set to 480 by default for the best results. + width (`int`, defaults to `704`): + The width in pixels of the generated image. This is set to 848 by default for the best results. + num_frames (`int`, defaults to `161`): + The number of video frames to generate + frame_rate (`int`, defaults to `25`): + Target frame rate of the generated video. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, defaults to `3 `): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + The interpolation factor between random noise and denoised latents at the decode timestep. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `128 `): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + video_sequence_length = latent_num_frames * latent_height * latent_width + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) + mu = calculate_shift( + video_sequence_length, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Prepare micro-conditions + rope_interpolation_scale = ( + self.vae_temporal_compression_ratio / frame_rate, + self.vae_spatial_compression_ratio, + self.vae_spatial_compression_ratio, + ) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + with self.transformer.cache_context("cond_uncond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + encoder_hidden_states=prompt_embeds, + timestep=timestep, + encoder_attention_mask=prompt_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + rope_interpolation_scale=rope_interpolation_scale, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred.float() + + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.guidance_rescale > 0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg( + noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + video = latents + else: + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = latents.to(prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + latents = latents.to(self.vae.dtype) + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return LTXPipelineOutput(frames=video) diff --git a/diffusers/pipelines/ltx/pipeline_ltx_condition.py b/diffusers/pipelines/ltx/pipeline_ltx_condition.py new file mode 100644 index 0000000000000000000000000000000000000000..28d296695998db367d2cc9e5d44a443e7db9452c --- /dev/null +++ b/diffusers/pipelines/ltx/pipeline_ltx_condition.py @@ -0,0 +1,1294 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from dataclasses import dataclass +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import T5EncoderModel, T5TokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTXVideo +from ...models.transformers import LTXVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import LTXPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXConditionPipeline, LTXVideoCondition + >>> from diffusers.utils import export_to_video, load_video, load_image + + >>> pipe = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.5", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> # Load input image and video + >>> video = load_video( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cosmos/cosmos-video2world-input-vid.mp4" + ... ) + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cosmos/cosmos-video2world-input.jpg" + ... ) + + >>> # Create conditioning objects + >>> condition1 = LTXVideoCondition( + ... image=image, + ... frame_index=0, + ... ) + >>> condition2 = LTXVideoCondition( + ... video=video, + ... frame_index=80, + ... ) + + >>> prompt = "The video depicts a long, straight highway stretching into the distance, flanked by metal guardrails. The road is divided into multiple lanes, with a few vehicles visible in the far distance. The surrounding landscape features dry, grassy fields on one side and rolling hills on the other. The sky is mostly clear with a few scattered clouds, suggesting a bright, sunny day. And then the camera switch to a winding mountain road covered in snow, with a single vehicle traveling along it. The road is flanked by steep, rocky cliffs and sparse vegetation. The landscape is characterized by rugged terrain and a river visible in the distance. The scene captures the solitude and beauty of a winter drive through a mountainous region." + >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted" + + >>> # Generate video + >>> generator = torch.Generator("cuda").manual_seed(0) + >>> # Text-only conditioning is also supported without the need to pass `conditions` + >>> video = pipe( + ... conditions=[condition1, condition2], + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... width=768, + ... height=512, + ... num_frames=161, + ... num_inference_steps=40, + ... generator=generator, + ... ).frames[0] + + >>> export_to_video(video, "output.mp4", fps=24) + ``` +""" + + +@dataclass +class LTXVideoCondition: + """ + Defines a single frame-conditioning item for LTX Video - a single frame or a sequence of frames. + + Attributes: + image (`PIL.Image.Image`): + The image to condition the video on. + video (`list[PIL.Image.Image]`): + The video to condition the video on. + frame_index (`int`): + The frame index at which the image or video will conditionally effect the video generation. + strength (`float`, defaults to `1.0`): + The strength of the conditioning effect. A value of `1.0` means the conditioning effect is fully applied. + """ + + image: PIL.Image.Image | None = None + video: list[PIL.Image.Image] | None = None + frame_index: int = 0 + strength: float = 1.0 + + +# from LTX-Video/ltx_video/schedulers/rf.py +def linear_quadratic_schedule(num_steps, threshold_noise=0.025, linear_steps=None): + if linear_steps is None: + linear_steps = num_steps // 2 + if num_steps < 2: + return torch.tensor([1.0]) + linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)] + threshold_noise_step_diff = linear_steps - threshold_noise * num_steps + quadratic_steps = num_steps - linear_steps + quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2) + linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2) + const = quadratic_coef * (linear_steps**2) + quadratic_sigma_schedule = [ + quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, num_steps) + ] + sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule + [1.0] + sigma_schedule = [1.0 - x for x in sigma_schedule] + return torch.tensor(sigma_schedule[:-1]) + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTXConditionPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin): + r""" + Pipeline for text/image/video-to-video generation. + + Reference: https://github.com/Lightricks/LTX-Video + + Args: + transformer ([`LTXVideoTransformer3DModel`]): + Conditional Transformer architecture to denoise the encoded video latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLLTXVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTXVideo, + text_encoder: T5EncoderModel, + tokenizer: T5TokenizerFast, + transformer: LTXVideoTransformer3DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128 + ) + + self.default_height = 512 + self.default_width = 704 + self.default_frames = 121 + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 256, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)[0] + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 256, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + def check_inputs( + self, + prompt, + conditions, + image, + video, + frame_index, + strength, + denoise_strength, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + if conditions is not None and (image is not None or video is not None): + raise ValueError("If `conditions` is provided, `image` and `video` must not be provided.") + + if conditions is None: + if isinstance(image, list) and isinstance(frame_index, list) and len(image) != len(frame_index): + raise ValueError( + "If `conditions` is not provided, `image` and `frame_index` must be of the same length." + ) + elif isinstance(image, list) and isinstance(strength, list) and len(image) != len(strength): + raise ValueError("If `conditions` is not provided, `image` and `strength` must be of the same length.") + elif isinstance(video, list) and isinstance(frame_index, list) and len(video) != len(frame_index): + raise ValueError( + "If `conditions` is not provided, `video` and `frame_index` must be of the same length." + ) + elif isinstance(video, list) and isinstance(strength, list) and len(video) != len(strength): + raise ValueError("If `conditions` is not provided, `video` and `strength` must be of the same length.") + + if denoise_strength < 0 or denoise_strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {denoise_strength}") + + @staticmethod + def _prepare_video_ids( + batch_size: int, + num_frames: int, + height: int, + width: int, + patch_size: int = 1, + patch_size_t: int = 1, + device: torch.device = None, + ) -> torch.Tensor: + latent_sample_coords = torch.meshgrid( + torch.arange(0, num_frames, patch_size_t, device=device), + torch.arange(0, height, patch_size, device=device), + torch.arange(0, width, patch_size, device=device), + indexing="ij", + ) + latent_sample_coords = torch.stack(latent_sample_coords, dim=0) + latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_coords = latent_coords.reshape(batch_size, -1, num_frames * height * width) + + return latent_coords + + @staticmethod + def _scale_video_ids( + video_ids: torch.Tensor, + scale_factor: int = 32, + scale_factor_t: int = 8, + frame_index: int = 0, + device: torch.device = None, + ) -> torch.Tensor: + scaled_latent_coords = ( + video_ids + * torch.tensor([scale_factor_t, scale_factor, scale_factor], device=video_ids.device)[None, :, None] + ) + scaled_latent_coords[:, 0] = (scaled_latent_coords[:, 0] + 1 - scale_factor_t).clamp(min=0) + scaled_latent_coords[:, 0] += frame_index + + return scaled_latent_coords + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + def trim_conditioning_sequence(self, start_frame: int, sequence_num_frames: int, target_num_frames: int): + """ + Trim a conditioning sequence to the allowed number of frames. + + Args: + start_frame (int): The target frame number of the first frame in the sequence. + sequence_num_frames (int): The number of frames in the sequence. + target_num_frames (int): The target number of frames in the generated video. + Returns: + int: updated sequence length + """ + scale_factor = self.vae_temporal_compression_ratio + num_frames = min(sequence_num_frames, target_num_frames - start_frame) + # Trim down to a multiple of temporal_scale_factor frames plus 1 + num_frames = (num_frames - 1) // scale_factor * scale_factor + 1 + return num_frames + + @staticmethod + def add_noise_to_image_conditioning_latents( + t: float, + init_latents: torch.Tensor, + latents: torch.Tensor, + noise_scale: float, + conditioning_mask: torch.Tensor, + generator, + eps=1e-6, + ): + """ + Add timestep-dependent noise to the hard-conditioning latents. This helps with motion continuity, especially + when conditioned on a single frame. + """ + noise = randn_tensor( + latents.shape, + generator=generator, + device=latents.device, + dtype=latents.dtype, + ) + # Add noise only to hard-conditioning latents (conditioning_mask = 1.0) + need_to_noise = (conditioning_mask > 1.0 - eps).unsqueeze(-1) + noised_latents = init_latents + noise_scale * noise * (t**2) + latents = torch.where(need_to_noise, noised_latents, latents) + return latents + + def prepare_latents( + self, + conditions: list[torch.Tensor] | None = None, + condition_strength: list[float] | None = None, + condition_frame_index: list[int] | None = None, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 704, + num_frames: int = 161, + num_prefix_latent_frames: int = 2, + sigma: torch.Tensor | None = None, + latents: torch.Tensor | None = None, + generator: torch.Generator | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, int]: + num_latent_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + + shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + if latents is not None and sigma is not None: + if latents.shape != shape: + raise ValueError( + f"Latents shape {latents.shape} does not match expected shape {shape}. Please check the input." + ) + latents = latents.to(device=device, dtype=dtype) + sigma = sigma.to(device=device, dtype=dtype) + latents = sigma * noise + (1 - sigma) * latents + else: + latents = noise + + if len(conditions) > 0: + condition_latent_frames_mask = torch.zeros( + (batch_size, num_latent_frames), device=device, dtype=torch.float32 + ) + + extra_conditioning_latents = [] + extra_conditioning_video_ids = [] + extra_conditioning_mask = [] + extra_conditioning_num_latents = 0 + for data, strength, frame_index in zip(conditions, condition_strength, condition_frame_index): + condition_latents = retrieve_latents(self.vae.encode(data), generator=generator) + condition_latents = self._normalize_latents( + condition_latents, self.vae.latents_mean, self.vae.latents_std + ).to(device, dtype=dtype) + + num_data_frames = data.size(2) + num_cond_frames = condition_latents.size(2) + + if frame_index == 0: + latents[:, :, :num_cond_frames] = torch.lerp( + latents[:, :, :num_cond_frames], condition_latents, strength + ) + condition_latent_frames_mask[:, :num_cond_frames] = strength + + else: + if num_data_frames > 1: + if num_cond_frames < num_prefix_latent_frames: + raise ValueError( + f"Number of latent frames must be at least {num_prefix_latent_frames} but got {num_data_frames}." + ) + + if num_cond_frames > num_prefix_latent_frames: + start_frame = frame_index // self.vae_temporal_compression_ratio + num_prefix_latent_frames + end_frame = start_frame + num_cond_frames - num_prefix_latent_frames + latents[:, :, start_frame:end_frame] = torch.lerp( + latents[:, :, start_frame:end_frame], + condition_latents[:, :, num_prefix_latent_frames:], + strength, + ) + condition_latent_frames_mask[:, start_frame:end_frame] = strength + condition_latents = condition_latents[:, :, :num_prefix_latent_frames] + + noise = randn_tensor(condition_latents.shape, generator=generator, device=device, dtype=dtype) + condition_latents = torch.lerp(noise, condition_latents, strength) + + condition_video_ids = self._prepare_video_ids( + batch_size, + condition_latents.size(2), + latent_height, + latent_width, + patch_size=self.transformer_spatial_patch_size, + patch_size_t=self.transformer_temporal_patch_size, + device=device, + ) + condition_video_ids = self._scale_video_ids( + condition_video_ids, + scale_factor=self.vae_spatial_compression_ratio, + scale_factor_t=self.vae_temporal_compression_ratio, + frame_index=frame_index, + device=device, + ) + condition_latents = self._pack_latents( + condition_latents, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + condition_conditioning_mask = torch.full( + condition_latents.shape[:2], strength, device=device, dtype=dtype + ) + + extra_conditioning_latents.append(condition_latents) + extra_conditioning_video_ids.append(condition_video_ids) + extra_conditioning_mask.append(condition_conditioning_mask) + extra_conditioning_num_latents += condition_latents.size(1) + + video_ids = self._prepare_video_ids( + batch_size, + num_latent_frames, + latent_height, + latent_width, + patch_size_t=self.transformer_temporal_patch_size, + patch_size=self.transformer_spatial_patch_size, + device=device, + ) + if len(conditions) > 0: + conditioning_mask = condition_latent_frames_mask.gather(1, video_ids[:, 0]) + else: + conditioning_mask, extra_conditioning_num_latents = None, 0 + video_ids = self._scale_video_ids( + video_ids, + scale_factor=self.vae_spatial_compression_ratio, + scale_factor_t=self.vae_temporal_compression_ratio, + frame_index=0, + device=device, + ) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + + if len(conditions) > 0 and len(extra_conditioning_latents) > 0: + latents = torch.cat([*extra_conditioning_latents, latents], dim=1) + video_ids = torch.cat([*extra_conditioning_video_ids, video_ids], dim=2) + conditioning_mask = torch.cat([*extra_conditioning_mask, conditioning_mask], dim=1) + + return latents, conditioning_mask, video_ids, extra_conditioning_num_latents + + def get_timesteps(self, sigmas, timesteps, num_inference_steps, strength): + num_steps = min(int(num_inference_steps * strength), num_inference_steps) + start_index = max(num_inference_steps - num_steps, 0) + sigmas = sigmas[start_index:] + timesteps = timesteps[start_index:] + return sigmas, timesteps, num_inference_steps - start_index + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + conditions: LTXVideoCondition | list[LTXVideoCondition] = None, + image: PipelineImageInput | list[PipelineImageInput] = None, + video: list[PipelineImageInput] = None, + frame_index: int | list[int] = 0, + strength: float | list[float] = 1.0, + denoise_strength: float = 1.0, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 704, + num_frames: int = 161, + frame_rate: int = 25, + num_inference_steps: int = 50, + timesteps: list[int] = None, + guidance_scale: float = 3, + guidance_rescale: float = 0.0, + image_cond_noise_scale: float = 0.15, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + conditions (`list[LTXVideoCondition], *optional*`): + The list of frame-conditioning items for the video generation.If not provided, conditions will be + created using `image`, `video`, `frame_index` and `strength`. + image (`PipelineImageInput` or `list[PipelineImageInput]`, *optional*): + The image or images to condition the video generation. If not provided, one has to pass `video` or + `conditions`. + video (`list[PipelineImageInput]`, *optional*): + The video to condition the video generation. If not provided, one has to pass `image` or `conditions`. + frame_index (`int` or `list[int]`, *optional*): + The frame index or frame indices at which the image or video will conditionally effect the video + generation. If not provided, one has to pass `conditions`. + strength (`float` or `list[float]`, *optional*): + The strength or strengths of the conditioning effect. If not provided, one has to pass `conditions`. + denoise_strength (`float`, defaults to `1.0`): + The strength of the noise added to the latents for editing. Higher strength leads to more noise added + to the latents, therefore leading to more differences between original video and generated video. This + is useful for video-to-video editing. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + height (`int`, defaults to `512`): + The height in pixels of the generated image. This is set to 480 by default for the best results. + width (`int`, defaults to `704`): + The width in pixels of the generated image. This is set to 848 by default for the best results. + num_frames (`int`, defaults to `161`): + The number of video frames to generate + frame_rate (`int`, defaults to `25`): + Target frame rate of the generated video. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, defaults to `3 `): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + image_cond_noise_scale (`float`, defaults to `0.15`): + Scale of noise added to the conditioning image latents. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + The interpolation factor between random noise and denoised latents at the decode timestep. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `128 `): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + conditions=conditions, + image=image, + video=video, + frame_index=frame_index, + strength=strength, + denoise_strength=denoise_strength, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if conditions is not None: + if not isinstance(conditions, list): + conditions = [conditions] + + strength = [condition.strength for condition in conditions] + frame_index = [condition.frame_index for condition in conditions] + image = [condition.image for condition in conditions] + video = [condition.video for condition in conditions] + elif image is not None or video is not None: + if not isinstance(image, list): + image = [image] + num_conditions = 1 + elif isinstance(image, list): + num_conditions = len(image) + if not isinstance(video, list): + video = [video] + num_conditions = 1 + elif isinstance(video, list): + num_conditions = len(video) + + if not isinstance(frame_index, list): + frame_index = [frame_index] * num_conditions + if not isinstance(strength, list): + strength = [strength] * num_conditions + + device = self._execution_device + vae_dtype = self.vae.dtype + + # 3. Prepare text embeddings & conditioning image/video + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + conditioning_tensors = [] + is_conditioning_image_or_video = image is not None or video is not None + if is_conditioning_image_or_video: + for condition_image, condition_video, condition_frame_index, condition_strength in zip( + image, video, frame_index, strength + ): + if condition_image is not None: + condition_tensor = ( + self.video_processor.preprocess(condition_image, height, width) + .unsqueeze(2) + .to(device, dtype=vae_dtype) + ) + elif condition_video is not None: + condition_tensor = self.video_processor.preprocess_video(condition_video, height, width) + num_frames_input = condition_tensor.size(2) + num_frames_output = self.trim_conditioning_sequence( + condition_frame_index, num_frames_input, num_frames + ) + condition_tensor = condition_tensor[:, :, :num_frames_output] + condition_tensor = condition_tensor.to(device, dtype=vae_dtype) + else: + raise ValueError("Either `image` or `video` must be provided for conditioning.") + + if condition_tensor.size(2) % self.vae_temporal_compression_ratio != 1: + raise ValueError( + f"Number of frames in the video must be of the form (k * {self.vae_temporal_compression_ratio} + 1) " + f"but got {condition_tensor.size(2)} frames." + ) + conditioning_tensors.append(condition_tensor) + + # 4. Prepare timesteps + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + + if timesteps is None: + sigmas = linear_quadratic_schedule(num_inference_steps) + timesteps = sigmas * 1000 + + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + timesteps, + ) + sigmas = self.scheduler.sigmas + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + latent_sigma = None + if denoise_strength < 1: + sigmas, timesteps, num_inference_steps = self.get_timesteps( + sigmas, timesteps, num_inference_steps, denoise_strength + ) + latent_sigma = sigmas[:1].repeat(batch_size * num_videos_per_prompt) + + self._num_timesteps = len(timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents, conditioning_mask, video_coords, extra_conditioning_num_latents = self.prepare_latents( + conditioning_tensors, + strength, + frame_index, + batch_size=batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + num_frames=num_frames, + sigma=latent_sigma, + latents=latents, + generator=generator, + device=device, + dtype=torch.float32, + ) + + video_coords = video_coords.float() + video_coords[:, 0] = video_coords[:, 0] * (1.0 / frame_rate) + + init_latents = latents.clone() if is_conditioning_image_or_video else None + + if self.do_classifier_free_guidance: + video_coords = torch.cat([video_coords, video_coords], dim=0) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + if image_cond_noise_scale > 0 and init_latents is not None: + # Add timestep-dependent noise to the hard-conditioning latents + # This helps with motion continuity, especially when conditioned on a single frame + latents = self.add_noise_to_image_conditioning_latents( + t / 1000.0, + init_latents, + latents, + image_cond_noise_scale, + conditioning_mask, + generator, + ) + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + if is_conditioning_image_or_video: + conditioning_mask_model_input = ( + torch.cat([conditioning_mask, conditioning_mask]) + if self.do_classifier_free_guidance + else conditioning_mask + ) + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]).unsqueeze(-1).float() + if is_conditioning_image_or_video: + timestep = torch.min(timestep, (1 - conditioning_mask_model_input) * 1000.0) + + with self.transformer.cache_context("cond_uncond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + encoder_hidden_states=prompt_embeds, + timestep=timestep, + encoder_attention_mask=prompt_attention_mask, + video_coords=video_coords, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + timestep, _ = timestep.chunk(2) + + if self.guidance_rescale > 0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg( + noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale + ) + + denoised_latents = self.scheduler.step( + -noise_pred, t, latents, per_token_timesteps=timestep, return_dict=False + )[0] + if is_conditioning_image_or_video: + tokens_to_denoise_mask = (t / 1000 - 1e-6 < (1.0 - conditioning_mask)).unsqueeze(-1) + latents = torch.where(tokens_to_denoise_mask, denoised_latents, latents) + else: + latents = denoised_latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if is_conditioning_image_or_video: + latents = latents[:, extra_conditioning_num_latents:] + + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + if output_type == "latent": + video = latents + else: + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = latents.to(prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return LTXPipelineOutput(frames=video) diff --git a/diffusers/pipelines/ltx/pipeline_ltx_i2v_long_multi_prompt.py b/diffusers/pipelines/ltx/pipeline_ltx_i2v_long_multi_prompt.py new file mode 100644 index 0000000000000000000000000000000000000000..838d5afc5c5adc37dc7500217fc35d758868b3fb --- /dev/null +++ b/diffusers/pipelines/ltx/pipeline_ltx_i2v_long_multi_prompt.py @@ -0,0 +1,1410 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import copy +from typing import Any, Callable + +import numpy as np +import PIL +import torch +from transformers import T5EncoderModel, T5TokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTXVideo +from ...models.transformers import LTXVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler, LTXEulerAncestralRFScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import LTXPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTXEulerAncestralRFScheduler, LTXI2VLongMultiPromptPipeline + + >>> pipe = LTXI2VLongMultiPromptPipeline.from_pretrained("LTX-Video-0.9.8-13B-distilled") + >>> # For ComfyUI parity, swap in the RF scheduler (keeps the original config). + >>> pipe.scheduler = LTXEulerAncestralRFScheduler.from_config(pipe.scheduler.config) + >>> pipe = pipe.to("cuda").to(dtype=torch.bfloat16) + >>> # Example A: get decoded frames (PIL) + >>> out = pipe( + ... prompt="a chimpanzee walks | a chimpanzee eats", + ... num_frames=161, + ... height=512, + ... width=704, + ... temporal_tile_size=80, + ... temporal_overlap=24, + ... output_type="pil", + ... return_dict=True, + ... ) + >>> frames = out.frames[0] # list of PIL.Image.Image + >>> # Example B: get latent video and decode later (saves VRAM during sampling) + >>> out_latent = pipe(prompt="a chimpanzee walking", output_type="latent", return_dict=True).frames + >>> frames = pipe.vae_decode_tiled(out_latent, output_type="pil")[0] + ``` +""" + + +def get_latent_coords( + latent_num_frames, latent_height, latent_width, batch_size, device, rope_interpolation_scale, latent_idx +): + """ + Compute latent patch top-left coordinates in (t, y, x) order. + + Args: + latent_num_frames: int. Number of latent frames (T_lat). + latent_height: int. Latent height (H_lat). + latent_width: int. Latent width (W_lat). + batch_size: int. Batch dimension (B). + device: torch.device for the resulting tensor. + rope_interpolation_scale: + tuple[int|float, int|float, int|float]. Scale per (t, y, x) latent step to pixel coords. + latent_idx: int | None. When not None, shifts the time coordinate to align segments: + - <= 0 uses step multiples of rope_interpolation_scale[0] + - > 0 starts at 1 then increments by rope_interpolation_scale[0] + + Returns: + Tensor of shape [B, 3, T_lat * H_lat * W_lat] containing top-left coordinates per latent patch, repeated for each + batch element. + """ + latent_sample_coords = torch.meshgrid( + torch.arange(0, latent_num_frames, 1, device=device), + torch.arange(0, latent_height, 1, device=device), + torch.arange(0, latent_width, 1, device=device), + indexing="ij", + ) + latent_sample_coords = torch.stack(latent_sample_coords, dim=0) + latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_coords = latent_coords.flatten(2) + pixel_coords = latent_coords * torch.tensor(rope_interpolation_scale, device=latent_coords.device)[None, :, None] + if latent_idx is not None: + if latent_idx <= 0: + frame_idx = latent_idx * rope_interpolation_scale[0] + else: + frame_idx = 1 + (latent_idx - 1) * rope_interpolation_scale[0] + if frame_idx == 0: + pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - rope_interpolation_scale[0]).clamp(min=0) + pixel_coords[:, 0] += frame_idx + return pixel_coords + + +# Copied from diffusers.pipelines.ltx.pipeline_ltx.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def adain_normalize_latents( + curr_latents: torch.Tensor, ref_latents: torch.Tensor | None, factor: float +) -> torch.Tensor: + """ + Optional AdaIN normalization: channel-wise mean/variance matching of curr_latents to ref_latents, controlled by + factor. + + Args: + curr_latents: Tensor [B, C, T, H, W]. Current window latents. + ref_latents: + Tensor | None [B, C, T_ref, H, W]. Reference latents (e.g., first window) used to compute target stats. + factor: float in [0, 1]. 0 keeps current stats; 1 matches reference stats. + + Returns: + Tensor with per-channel mean/std blended towards the reference. + """ + if ref_latents is None or factor is None or factor <= 0: + return curr_latents + + eps = torch.tensor(1e-6, device=curr_latents.device, dtype=curr_latents.dtype) + + # Compute per-channel means/stds for current and reference over (T, H, W) + mu_curr = curr_latents.mean(dim=(2, 3, 4), keepdim=True) + sigma_curr = curr_latents.std(dim=(2, 3, 4), keepdim=True) + + mu_ref = ref_latents.mean(dim=(2, 3, 4), keepdim=True).to(device=curr_latents.device, dtype=curr_latents.dtype) + sigma_ref = ref_latents.std(dim=(2, 3, 4), keepdim=True).to(device=curr_latents.device, dtype=curr_latents.dtype) + + # Blend target statistics + mu_blend = (1.0 - float(factor)) * mu_curr + float(factor) * mu_ref + sigma_blend = (1.0 - float(factor)) * sigma_curr + float(factor) * sigma_ref + sigma_blend = torch.clamp(sigma_blend, min=float(eps)) + + # Apply AdaIN + curr_norm = (curr_latents - mu_curr) / (sigma_curr + eps) + return curr_norm * sigma_blend + mu_blend + + +def split_into_temporal_windows( + latent_len: int, temporal_tile_size: int, temporal_overlap: int, compression: int +) -> list[tuple[int, int]]: + """ + Split latent frames into sliding windows. + + Args: + latent_len: int. Number of latent frames (T_lat). + temporal_tile_size: int. Window size in latent frames (> 0). + temporal_overlap: int. Overlap between windows in latent frames (>= 0). + compression: int. VAE temporal compression ratio (unused here; kept for parity). + + Returns: + list[tuple[int, int]]: inclusive-exclusive (start, end) indices per window. + """ + if temporal_tile_size <= 0: + raise ValueError("temporal_tile_size must be > 0") + stride = max(temporal_tile_size - temporal_overlap, 1) + windows = [] + start = 0 + while start < latent_len: + end = min(start + temporal_tile_size, latent_len) + windows.append((start, end)) + if end == latent_len: + break + start = start + stride + return windows + + +def linear_overlap_fuse(prev: torch.Tensor, new: torch.Tensor, overlap: int) -> torch.Tensor: + """ + Temporal linear crossfade between two latent clips over the overlap region. + + Args: + prev: Tensor [B, C, F, H, W]. Previous output segment. + new: Tensor [B, C, F, H, W]. New segment to be appended. + overlap: int. Number of frames to crossfade (overlap <= 1 concatenates without blend). + + Returns: + Tensor [B, C, F_prev + F_new - overlap, H, W] after crossfade at the seam. + """ + if overlap <= 1: + return torch.cat([prev, new], dim=2) + alpha = torch.linspace(1, 0, overlap + 2, device=prev.device, dtype=prev.dtype)[1:-1] + shape = [1] * prev.ndim + shape[2] = alpha.size(0) + alpha = alpha.reshape(shape) + blended = alpha * prev[:, :, -overlap:] + (1 - alpha) * new[:, :, :overlap] + return torch.cat([prev[:, :, :-overlap], blended, new[:, :, overlap:]], dim=2) + + +def inject_prev_tail_latents( + window_latents: torch.Tensor, + prev_tail_latents: torch.Tensor | None, + window_cond_mask_5d: torch.Tensor, + overlap_lat: int, + strength: float | None, + prev_overlap_len: int, +) -> tuple[torch.Tensor, torch.Tensor, int]: + """ + Inject the tail latents from the previous window at the beginning of the current window (first k frames), where k = + min(overlap_lat, T_curr, T_prev_tail). + + Args: + window_latents: Tensor [B, C, T, H, W]. Current window latents. + prev_tail_latents: Tensor | None [B, C, T_prev, H, W]. Tail segment from the previous window. + window_cond_mask_5d: Tensor [B, 1, T, H, W]. Per-token conditioning mask (1 = free, 0 = hard condition). + overlap_lat: int. Number of latent frames to inject from the previous tail. + strength: float | None in [0, 1]. Blend strength; 1.0 replaces, 0.0 keeps original. + prev_overlap_len: int. Accumulated overlap length so far (used for trimming later). + + Returns: + tuple[Tensor, Tensor, int]: (updated_window_latents, updated_cond_mask, updated_prev_overlap_len) + """ + if prev_tail_latents is None or overlap_lat <= 0 or strength is None or strength <= 0: + return window_latents, window_cond_mask_5d, prev_overlap_len + + # Expected shape: [B, C, T, H, W] + T = int(window_latents.shape[2]) + k = min(int(overlap_lat), T, int(prev_tail_latents.shape[2])) + if k <= 0: + return window_latents, window_cond_mask_5d, prev_overlap_len + + tail = prev_tail_latents[:, :, -k:] + mask = torch.full( + (window_cond_mask_5d.shape[0], 1, tail.shape[2], window_cond_mask_5d.shape[3], window_cond_mask_5d.shape[4]), + 1.0 - strength, + dtype=window_cond_mask_5d.dtype, + device=window_cond_mask_5d.device, + ) + + window_latents = torch.cat([window_latents, tail], dim=2) + window_cond_mask_5d = torch.cat([window_cond_mask_5d, mask], dim=2) + return window_latents, window_cond_mask_5d, prev_overlap_len + k + + +def build_video_coords_for_window( + latents: torch.Tensor, + overlap_len: int, + guiding_len: int, + negative_len: int, + rope_interpolation_scale: torch.Tensor, + frame_rate: int, +) -> torch.Tensor: + """ + Build video_coords: [B, 3, S] with order [t, y, x]. + + Args: + latents: Tensor [B, C, T, H, W]. Current window latents (before any trimming). + overlap_len: int. Number of frames from previous tail injected at the head. + guiding_len: int. Number of guidance frames appended at the head. + negative_len: int. Number of negative-index frames appended at the head (typically 1 or 0). + rope_interpolation_scale: tuple[int|float, int|float, int|float]. Scale for (t, y, x). + frame_rate: int. Used to convert time indices into seconds (t /= frame_rate). + + Returns: + Tensor [B, 3, T*H*W] of fractional pixel coordinates per latent patch. + """ + + b, c, f, h, w = latents.shape + pixel_coords = get_latent_coords(f, h, w, b, latents.device, rope_interpolation_scale, 0) + replace_corrds = [] + if overlap_len > 0: + replace_corrds.append(get_latent_coords(overlap_len, h, w, b, latents.device, rope_interpolation_scale, 0)) + if guiding_len > 0: + replace_corrds.append( + get_latent_coords(guiding_len, h, w, b, latents.device, rope_interpolation_scale, overlap_len) + ) + if negative_len > 0: + replace_corrds.append(get_latent_coords(negative_len, h, w, b, latents.device, rope_interpolation_scale, -1)) + if len(replace_corrds) > 0: + replace_corrds = torch.cat(replace_corrds, axis=2) + pixel_coords[:, :, -replace_corrds.shape[2] :] = replace_corrds + fractional_coords = pixel_coords.to(torch.float32) + fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate) + return fractional_coords + + +def parse_prompt_segments(prompt: str | list[str], prompt_segments: list[dict[str, Any]] | None) -> list[str]: + """ + Return a list of positive prompts per window index. + + Args: + prompt: str | list[str]. If str contains '|', parts are split by bars and trimmed. + prompt_segments: + list[dict], optional. Each dict with {"start_window", "end_window", "text"} overrides prompts per window. + + Returns: + list[str] containing the positive prompt for each window index. + """ + if prompt is None: + return [] + if prompt_segments: + max_w = 0 + for seg in prompt_segments: + max_w = max(max_w, int(seg.get("end_window", 0))) + texts = [""] * (max_w + 1) + for seg in prompt_segments: + s = int(seg.get("start_window", 0)) + e = int(seg.get("end_window", s)) + txt = seg.get("text", "") + for w in range(s, e + 1): + texts[w] = txt + # fill empty by last non-empty + last = "" + for i in range(len(texts)): + if texts[i] == "": + texts[i] = last + else: + last = texts[i] + return texts + + # bar-split mode + if isinstance(prompt, str): + parts = [p.strip() for p in prompt.split("|")] + else: + parts = prompt + parts = [p for p in parts if p is not None] + return parts + + +def batch_normalize(latents, reference, factor): + """ + Batch AdaIN-like normalization for latents in dict format (ComfyUI-compatible). + + Args: + latents: dict containing "samples" shaped [B, C, F, H, W] + reference: dict containing "samples" used to compute target stats + factor: float in [0, 1]; 0 = no change, 1 = full match to reference + Returns: + tuple[dict]: a single-element tuple with the updated latents dict. + """ + latents_copy = copy.deepcopy(latents) + t = latents_copy["samples"] # B x C x F x H x W + + for i in range(t.size(0)): # batch + for c in range(t.size(1)): # channel + r_sd, r_mean = torch.std_mean(reference["samples"][i, c], dim=None) # index by original dim order + i_sd, i_mean = torch.std_mean(t[i, c], dim=None) + + t[i, c] = ((t[i, c] - i_mean) / i_sd) * r_sd + r_mean + + latents_copy["samples"] = torch.lerp(latents["samples"], t, factor) + return (latents_copy,) + + +class LTXI2VLongMultiPromptPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin): + r""" + Long-duration I2V (image-to-video) multi-prompt pipeline with ComfyUI parity. + + Key features: + - Temporal sliding-window sampling only (no spatial H/W sharding); autoregressive fusion across windows. + - Multi-prompt segmentation per window with smooth transitions at window heads. + - First-frame hard conditioning via per-token mask for I2V. + - VRAM control via temporal windowing and VAE tiled decoding. + + Reference: https://github.com/Lightricks/LTX-Video + + Args: + transformer ([`LTXVideoTransformer3DModel`]): + Conditional Transformer architecture to denoise the encoded video latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`] or [`LTXEulerAncestralRFScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLLTXVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`T5TokenizerFast`): + Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTXVideo, + text_encoder: T5EncoderModel, + tokenizer: T5TokenizerFast, + transformer: LTXVideoTransformer3DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + if not isinstance(scheduler, LTXEulerAncestralRFScheduler): + logger.warning( + "For ComfyUI parity, `LTXI2VLongMultiPromptPipeline` is typically run with " + "`LTXEulerAncestralRFScheduler`. Got %s.", + scheduler.__class__.__name__, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer", None) is not None else 1 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128 + ) + + self.default_height = 512 + self.default_width = 704 + self.default_frames = 121 + self._current_tile_T = None + + @property + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.guidance_scale + def guidance_scale(self): + return self._guidance_scale + + @property + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.guidance_rescale + def guidance_rescale(self): + return self._guidance_rescale + + @property + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.do_classifier_free_guidance + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.num_timesteps + def num_timesteps(self): + return self._num_timesteps + + @property + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.current_timestep + def current_timestep(self): + return self._current_timestep + + @property + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.attention_kwargs + def attention_kwargs(self): + return self._attention_kwargs + + @property + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.interrupt + def interrupt(self): + return self._interrupt + + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(text_input_ids.to(device))[0] + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + num_frames: int, + device: torch.device, + generator: torch.Generator | None, + dtype: torch.dtype = torch.float32, + latents: torch.Tensor | None = None, + cond_latents: torch.Tensor | None = None, + cond_strength: float = 0.0, + negative_index_latents: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor | None, int, int, int]: + """ + Prepare base latents and optionally inject first-frame conditioning latents. + + Returns: + latents, negative_index_latents, latent_num_frames, latent_height, latent_width + """ + if latents is None: + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + latents = torch.zeros( + (batch_size, num_channels_latents, latent_num_frames, latent_height, latent_width), + device=device, + dtype=dtype, + ) + else: + latent_num_frames = latents.shape[2] + latent_height = latents.shape[3] + latent_width = latents.shape[4] + latents = latents.to(device=device, dtype=dtype) + + if cond_latents is not None and cond_strength > 0: + if negative_index_latents is None: + negative_index_latents = cond_latents + latents[:, :, :1, :, :] = cond_latents + + return latents, negative_index_latents, latent_num_frames, latent_height, latent_width + + # TODO: refactor this out + @torch.no_grad() + def vae_decode_tiled( + self, + latents: torch.Tensor, + decode_timestep: float | None = None, + decode_noise_scale: float | None = None, + horizontal_tiles: int = 4, + vertical_tiles: int = 4, + overlap: int = 3, + last_frame_fix: bool = True, + generator: torch.Generator | None = None, + output_type: str = "pt", + auto_denormalize: bool = True, + compute_dtype: torch.dtype = torch.float32, + enable_vae_tiling: bool = False, + ) -> torch.Tensor | np.ndarray | list[PIL.Image.Image]: + """ + VAE-based spatial tiled decoding (ComfyUI parity) implemented in Diffusers style. + - Linearly feather and blend overlapping tiles to avoid seams. + - Optional last_frame_fix: duplicate the last latent frame before decoding, then drop time_scale_factor frames + at the end. + - Supports timestep_conditioning and decode_noise_scale injection. + - By default, "normalized latents" (the denoising output) are de-normalized internally (auto_denormalize=True). + - Tile fusion is computed in compute_dtype (float32 by default) to reduce blur and color shifts. + + Args: + latents: [B, C_latent, F_latent, H_latent, W_latent] + decode_timestep: Optional decode timestep (effective only if VAE supports timestep_conditioning) + decode_noise_scale: + Optional decode noise interpolation (effective only if VAE supports timestep_conditioning) + horizontal_tiles, vertical_tiles: Number of tiles horizontally/vertically (>= 1) + overlap: Overlap in latent space (in latent pixels, >= 0) + last_frame_fix: Whether to enable the "repeat last frame" fix + generator: Random generator (used for decode_noise_scale noise) + output_type: "latent" | "pt" | "np" | "pil" + - "latent": return latents unchanged (useful for downstream processing) + - "pt": return tensor in VAE output space + - "np"/"pil": post-processed outputs via VideoProcessor.postprocess_video + auto_denormalize: If True, apply LTX de-normalization to `latents` internally (recommended) + compute_dtype: Precision used during tile fusion (float32 default; significantly reduces seam blur) + enable_vae_tiling: If True, delegate tiling to VAE's built-in `tiled_decode` (sets `vae.use_tiling`). + + Returns: + - If output_type="latent": returns input `latents` unchanged + - If output_type="pt": returns [B, C, F, H, W] (values roughly in [-1, 1]) + - If output_type="np"/"pil": returns post-processed outputs via postprocess_video + """ + if output_type == "latent": + return latents + if horizontal_tiles < 1 or vertical_tiles < 1: + raise ValueError("horizontal_tiles and vertical_tiles must be >= 1") + overlap = max(int(overlap), 0) + + # Device and precision + device = self._execution_device + latents = latents.to(device=device, dtype=compute_dtype) + + # De-normalize to VAE space (avoid color artifacts) + if auto_denormalize: + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + # dtype required for VAE forward pass + latents = latents.to(dtype=self.vae.dtype) + + # Temporal/spatial upscaling ratios (parity with ComfyUI's downscale_index_formula) + tsf = int(self.vae_temporal_compression_ratio) + sf = int(self.vae_spatial_compression_ratio) + + # Optional: last_frame_fix (repeat last latent frame) + if last_frame_fix: + latents = torch.cat([latents, latents[:, :, -1:].contiguous()], dim=2) + + b, c_lat, f_lat, h_lat, w_lat = latents.shape + f_out = 1 + (f_lat - 1) * tsf + h_out = h_lat * sf + w_out = w_lat * sf + + # timestep_conditioning + decode-time noise injection (aligned with pipeline) + if getattr(self.vae.config, "timestep_conditioning", False): + dt = float(decode_timestep) if decode_timestep is not None else 0.0 + vt = torch.tensor([dt], device=device, dtype=latents.dtype) + if decode_noise_scale is not None: + dns = torch.tensor([float(decode_noise_scale)], device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + latents = (1 - dns) * latents + dns * noise + else: + vt = None + + if enable_vae_tiling and hasattr(self.vae, "enable_tiling"): + self.vae.enable_tiling() + decoded = self.vae.decode(latents, vt, return_dict=False)[0] + if last_frame_fix: + decoded = decoded[:, :, :-tsf, :, :] + if output_type in ("np", "pil"): + return self.video_processor.postprocess_video(decoded, output_type=output_type) + return decoded + + # Compute base tile sizes (in latent space) + base_tile_h = (h_lat + (vertical_tiles - 1) * overlap) // vertical_tiles + base_tile_w = (w_lat + (horizontal_tiles - 1) * overlap) // horizontal_tiles + + output: torch.Tensor | None = None # [B, C_img, F, H, W], fused using compute_dtype + weights: torch.Tensor | None = None # [B, 1, F, H, W], fused using compute_dtype + + # Iterate tiles in latent space (no temporal tiling) + for v in range(vertical_tiles): + for h in range(horizontal_tiles): + h_start = h * (base_tile_w - overlap) + v_start = v * (base_tile_h - overlap) + + h_end = min(h_start + base_tile_w, w_lat) if h < horizontal_tiles - 1 else w_lat + v_end = min(v_start + base_tile_h, h_lat) if v < vertical_tiles - 1 else h_lat + + # Slice latent tile and decode + tile_latents = latents[:, :, :, v_start:v_end, h_start:h_end] + decoded_tile = self.vae.decode(tile_latents, vt, return_dict=False)[0] # [B, C, F, Ht, Wt] + # Cast to high precision to reduce blending blur + decoded_tile = decoded_tile.to(dtype=compute_dtype) + + # Initialize output buffers (compute_dtype) + if output is None: + output = torch.zeros( + (b, decoded_tile.shape[1], f_out, h_out, w_out), + device=decoded_tile.device, + dtype=compute_dtype, + ) + weights = torch.zeros( + (b, 1, f_out, h_out, w_out), + device=decoded_tile.device, + dtype=compute_dtype, + ) + + # Tile placement in output pixel space + out_h_start = v_start * sf + out_h_end = v_end * sf + out_w_start = h_start * sf + out_w_end = h_end * sf + + tile_out_h = out_h_end - out_h_start + tile_out_w = out_w_end - out_w_start + + # Linear feathering weights [B, 1, F, Ht, Wt] (compute_dtype) + tile_weights = torch.ones( + (b, 1, decoded_tile.shape[2], tile_out_h, tile_out_w), + device=decoded_tile.device, + dtype=compute_dtype, + ) + + overlap_out_h = overlap * sf + overlap_out_w = overlap * sf + + # Horizontal feathering: left/right overlaps + if overlap_out_w > 0: + if h > 0: + h_blend = torch.linspace( + 0, 1, steps=overlap_out_w, device=decoded_tile.device, dtype=compute_dtype + ) + tile_weights[:, :, :, :, :overlap_out_w] *= h_blend.view(1, 1, 1, 1, -1) + if h < horizontal_tiles - 1: + h_blend = torch.linspace( + 1, 0, steps=overlap_out_w, device=decoded_tile.device, dtype=compute_dtype + ) + tile_weights[:, :, :, :, -overlap_out_w:] *= h_blend.view(1, 1, 1, 1, -1) + + # Vertical feathering: top/bottom overlaps + if overlap_out_h > 0: + if v > 0: + v_blend = torch.linspace( + 0, 1, steps=overlap_out_h, device=decoded_tile.device, dtype=compute_dtype + ) + tile_weights[:, :, :, :overlap_out_h, :] *= v_blend.view(1, 1, 1, -1, 1) + if v < vertical_tiles - 1: + v_blend = torch.linspace( + 1, 0, steps=overlap_out_h, device=decoded_tile.device, dtype=compute_dtype + ) + tile_weights[:, :, :, -overlap_out_h:, :] *= v_blend.view(1, 1, 1, -1, 1) + + # Accumulate blended tile + output[:, :, :, out_h_start:out_h_end, out_w_start:out_w_end] += decoded_tile * tile_weights + weights[:, :, :, out_h_start:out_h_end, out_w_start:out_w_end] += tile_weights + + # Normalize, then clamp to [-1, 1] in compute_dtype to avoid color artifacts + output = output / (weights + 1e-8) + output = output.clamp(-1.0, 1.0) + output = output.to(dtype=self.vae.dtype) + + # Optional: drop the last tsf frames after last_frame_fix + if last_frame_fix: + output = output[:, :, :-tsf, :, :] + + if output_type in ("np", "pil"): + return self.video_processor.postprocess_video(output, output_type=output_type) + return output + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + prompt_segments: list[dict[str, Any]] | None = None, + height: int = 512, + width: int = 704, + num_frames: int = 161, + frame_rate: float = 25, + guidance_scale: float = 1.0, + guidance_rescale: float = 0.0, + num_inference_steps: int | None = 8, + sigmas: list[float, torch.Tensor] | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + seed: int | None = 0, + cond_image: "PIL.Image.Image" | torch.Tensor | None = None, + cond_strength: float = 0.5, + latents: torch.Tensor | None = None, + temporal_tile_size: int = 80, + temporal_overlap: int = 24, + temporal_overlap_cond_strength: float = 0.5, + adain_factor: float = 0.25, + guidance_latents: torch.Tensor | None = None, + guiding_strength: float = 1.0, + negative_index_latents: torch.Tensor | None = None, + negative_index_strength: float = 1.0, + skip_steps_sigma_threshold: float | None = 1, + decode_timestep: float | None = 0.05, + decode_noise_scale: float | None = 0.025, + decode_horizontal_tiles: int = 4, + decode_vertical_tiles: int = 4, + decode_overlap: int = 3, + output_type: str | None = "latent", # "latent" | "pt" | "np" | "pil" + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 128, + ): + r""" + Generate an image-to-video sequence via temporal sliding windows and multi-prompt scheduling. + + Args: + prompt (`str` or `list[str]`, *optional*): + Positive text prompt(s) per window. If a single string contains '|', parts are split by bars. + negative_prompt (`str` or `list[str]`, *optional*): + Negative prompt(s) to suppress undesired content. + prompt_segments (`list[dict]`, *optional*): + Segment mapping with {"start_window", "end_window", "text"} to override prompts per window. + height (`int`, defaults to `512`): + Output image height in pixels; must be divisible by 32. + width (`int`, defaults to `704`): + Output image width in pixels; must be divisible by 32. + num_frames (`int`, defaults to `161`): + Number of output frames (in decoded pixel space). + frame_rate (`float`, defaults to `25`): + Frames-per-second; used to normalize temporal coordinates in `video_coords`. + guidance_scale (`float`, defaults to `1.0`): + CFG scale; values > 1 enable classifier-free guidance. + guidance_rescale (`float`, defaults to `0.0`): + Optional rescale to mitigate overexposure under CFG (see `rescale_noise_cfg`). + num_inference_steps (`int`, *optional*, defaults to `8`): + Denoising steps per window. Ignored if `sigmas` is provided. + sigmas (`list[float]` or `torch.Tensor`, *optional*): + Explicit sigma schedule per window; if set, overrides `num_inference_steps`. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + Controls stochasticity; list accepted but first element is used (batch=1). + seed (`int`, *optional*, defaults to `0`): + If provided, seeds the shared generator for global latents and derives a window-local generator with + `seed + w_start` per temporal window. + cond_image (`PIL.Image.Image` or `torch.Tensor`, *optional*): + Conditioning image; fixes frame 0 via per-token mask when `cond_strength > 0`. + cond_strength (`float`, defaults to `0.5`): + Strength of first-frame hard conditioning (smaller cond_mask ⇒ stronger preservation). + latents (`torch.Tensor`, *optional*): + Initial latents [B, C_lat, F_lat, H_lat, W_lat]; if None, sampled with `randn_tensor`. + temporal_tile_size (`int`, defaults to `80`): + Temporal window size (in decoded frames); internally scaled by VAE temporal compression. + temporal_overlap (`int`, defaults to `24`): + Overlap between consecutive windows (in decoded frames); internally scaled by compression. + temporal_overlap_cond_strength (`float`, defaults to `0.5`): + Strength for injecting previous window tail latents at new window head. + adain_factor (`float`, defaults to `0.25`): + AdaIN normalization strength for cross-window consistency (0 disables). + guidance_latents (`torch.Tensor`, *optional*): + Reference latents injected at window head; length trimmed by overlap for subsequent windows. + guiding_strength (`float`, defaults to `1.0`): + Injection strength for `guidance_latents`. + negative_index_latents (`torch.Tensor`, *optional*): + A single-frame latent appended at window head for "negative index" semantics. + negative_index_strength (`float`, defaults to `1.0`): + Injection strength for `negative_index_latents`. + skip_steps_sigma_threshold (`float`, *optional*, defaults to `1`): + Skip steps whose sigma exceeds this threshold. + decode_timestep (`float`, *optional*, defaults to `0.05`): + Decode-time timestep (if VAE supports timestep_conditioning). + decode_noise_scale (`float`, *optional*, defaults to `0.025`): + Decode-time noise mix scale (if VAE supports timestep_conditioning). + decode_horizontal_tiles (`int`, defaults to `4`): + Number of horizontal tiles during VAE decoding. + decode_vertical_tiles (`int`, defaults to `4`): + Number of vertical tiles during VAE decoding. + decode_overlap (`int`, defaults to `3`): + Overlap (in latent pixels) between tiles during VAE decoding. + output_type (`str`, *optional*, defaults to `"latent"`): + The output format of the generated video. Choose between "latent", "pt", "np", or "pil". If "latent", + returns latents without decoding. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + Extra attention parameters forwarded to the transformer. + callback_on_step_end (`PipelineCallback` or `MultiPipelineCallbacks`, *optional*): + Per-step callback hook. + callback_on_step_end_tensor_inputs (`list[str]`, defaults to `["latents"]`): + Keys from locals() to pass into the callback. + max_sequence_length (`int`, defaults to `128`): + Tokenizer max length for prompt encoding. + + Examples: + + Returns: + [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated frames. The output format depends on + `output_type`: + - "latent"/"pt": `torch.Tensor` [B, C, F, H, W]; "latent" is in normalized latent space, "pt" is VAE + output space. + - "np": `np.ndarray` post-processed. + - "pil": `list[PIL.Image.Image]` list of PIL images. + + Shapes: + Latent sizes (when auto-generated): + - F_lat = (num_frames - 1) // vae_temporal_compression_ratio + 1 + - H_lat = height // vae_spatial_compression_ratio + - W_lat = width // vae_spatial_compression_ratio + + Notes: + - Seeding: when `seed` is provided, each temporal window uses a local generator seeded with `seed + + w_start`, while the shared generator is seeded once for global latents if no generator is passed; + otherwise the passed-in generator is reused. + - CFG: unified `noise_pred = uncond + w * (text - uncond)` with optional `guidance_rescale`. + - Memory: denoising performs full-frame predictions (no spatial tiling); decoding can be tiled to avoid + OOM. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Input validation: height/width must be divisible by 32 + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 1. Device & generator + device = self._execution_device + # Normalize generator input: accept list but use the first (batch_size=1) + if isinstance(generator, list): + generator = generator[0] + if seed is not None and generator is None: + generator = torch.Generator(device=device).manual_seed(seed) + + # 2. Optional i2v first frame conditioning: encode cond_image and inject at frame 0 via prepare_latents + cond_latents = None + if cond_image is not None and cond_strength > 0: + img = self.video_processor.preprocess(cond_image, height=height, width=width) + img = img.to(device=device, dtype=self.vae.dtype) + enc = self.vae.encode(img.unsqueeze(2)) # [B, C, 1, h, w] + cond_latents = enc.latent_dist.mode() if hasattr(enc, "latent_dist") else enc.latents + cond_latents = cond_latents.to(torch.float32) + cond_latents = self._normalize_latents( + cond_latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + + # 3. Global initial latents [B,C,F,H,W], optionally seeded/conditioned + latents, negative_index_latents, latent_num_frames, latent_height, latent_width = self.prepare_latents( + batch_size=1, + num_channels_latents=self.transformer.config.in_channels, + height=height, + width=width, + num_frames=num_frames, + device=device, + generator=generator, + dtype=torch.float32, + latents=latents, + cond_latents=cond_latents, + cond_strength=cond_strength, + negative_index_latents=negative_index_latents, + ) + if guidance_latents is not None: + guidance_latents = guidance_latents.to(device=device, dtype=torch.float32) + if latents.shape[2] != guidance_latents.shape[2]: + raise ValueError("The number of frames in `latents` and `guidance_latents` must be the same") + + # 4. Sliding windows in latent frames + tile_size_lat = max(1, temporal_tile_size // self.vae_temporal_compression_ratio) + overlap_lat = max(0, temporal_overlap // self.vae_temporal_compression_ratio) + windows = split_into_temporal_windows( + latent_num_frames, tile_size_lat, overlap_lat, self.vae_temporal_compression_ratio + ) + + # 5. Multi-prompt segments parsing + segment_texts = parse_prompt_segments(prompt, prompt_segments) + + out_latents = None + first_window_latents = None + + # 6. Process each temporal window + for w_idx, (w_start, w_end) in enumerate(windows): + if self.interrupt: + break + + # 6.1 Encode prompt embeddings per window segment + seg_index = min(w_idx, len(segment_texts) - 1) if segment_texts else 0 + pos_text = segment_texts[seg_index] if segment_texts else (prompt if isinstance(prompt, str) else "") + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=[pos_text], + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=1, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + max_sequence_length=max_sequence_length, + device=device, + dtype=None, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 6.2 Window-level timesteps reset: fresh sampling for each temporal window + if sigmas is not None: + s = torch.tensor(sigmas, dtype=torch.float32) if not isinstance(sigmas, torch.Tensor) else sigmas + self.scheduler.set_timesteps(sigmas=s, device=device) + self._num_timesteps = len(sigmas) + else: + self.scheduler.set_timesteps(num_inference_steps=num_inference_steps, device=device) + self._num_timesteps = num_inference_steps + + # 6.3 Extract window latents [B,C,T,H,W] + window_latents = latents[:, :, w_start:w_end] + window_guidance_latents = guidance_latents[:, :, w_start:w_end] if guidance_latents is not None else None + window_T = window_latents.shape[2] + + # 6.4 Build per-window cond mask and inject previous tails / reference + window_cond_mask_5d = torch.ones( + (1, 1, window_T, latent_height, latent_width), device=device, dtype=torch.float32 + ) + self._current_tile_T = window_T + prev_overlap_len = 0 + # Inter-window tail latent injection (Extend) + if w_idx > 0 and overlap_lat > 0 and out_latents is not None: + k = min(overlap_lat, out_latents.shape[2]) + prev_tail = out_latents[:, :, -k:] + window_latents, window_cond_mask_5d, prev_overlap_len = inject_prev_tail_latents( + window_latents, + prev_tail, + window_cond_mask_5d, + overlap_lat, + temporal_overlap_cond_strength, + prev_overlap_len, + ) + # Reference/negative-index latent injection (append 1 frame at window head; controlled by negative_index_strength) + if window_guidance_latents is not None: + guiding_len = ( + window_guidance_latents.shape[2] if w_idx == 0 else window_guidance_latents.shape[2] - overlap_lat + ) + window_latents, window_cond_mask_5d, prev_overlap_len = inject_prev_tail_latents( + window_latents, + window_guidance_latents[:, :, -guiding_len:], + window_cond_mask_5d, + guiding_len, + guiding_strength, + prev_overlap_len, + ) + else: + guiding_len = 0 + window_latents, window_cond_mask_5d, prev_overlap_len = inject_prev_tail_latents( + window_latents, + negative_index_latents, + window_cond_mask_5d, + 1, + negative_index_strength, + prev_overlap_len, + ) + if w_idx == 0 and cond_image is not None and cond_strength > 0: + # First-frame I2V: smaller mask means stronger preservation of the original latent + window_cond_mask_5d[:, :, 0] = 1.0 - cond_strength + + # Update effective window latent sizes (consider injections on T/H/W) + w_B, w_C, w_T_eff, w_H_eff, w_W_eff = window_latents.shape + p = self.transformer_spatial_patch_size + pt = self.transformer_temporal_patch_size + + # 6.5 Pack full-window latents/masks once + # Seeding policy: derive a window-local generator to decouple RNG across windows + if seed is not None: + tile_seed = int(seed) + int(w_start) + local_gen = torch.Generator(device=device).manual_seed(tile_seed) + else: + local_gen = generator + # randn*mask + (1-mask)*latents implements hard-condition initialization + init_rand = randn_tensor(window_latents.shape, generator=local_gen, device=device, dtype=torch.float32) + mixed_latents = init_rand * window_cond_mask_5d + (1 - window_cond_mask_5d) * window_latents + window_latents_packed = self._pack_latents( + window_latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + latents_packed = self._pack_latents( + mixed_latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + cond_mask_tokens = self._pack_latents( + window_cond_mask_5d, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + if self.do_classifier_free_guidance: + cond_mask = torch.cat([cond_mask_tokens, cond_mask_tokens], dim=0) + else: + cond_mask = cond_mask_tokens + + # 6.6 Denoising loop per full window (no spatial tiling) + sigmas_current = self.scheduler.sigmas.to(device=latents_packed.device) + if sigmas_current.shape[0] >= 2: + for i, t in enumerate(self.progress_bar(self.scheduler.timesteps[:-1])): + if self.interrupt: + break + # Skip semantics: if sigma exceeds threshold, skip this step (do not call scheduler.step) + sigma_val = float(sigmas_current[i].item()) + if skip_steps_sigma_threshold is not None and float(skip_steps_sigma_threshold) > 0.0: + if sigma_val > float(skip_steps_sigma_threshold): + continue + + self._current_timestep = t + + # Model input (stack 2 copies under CFG) + latent_model_input = ( + torch.cat([latents_packed] * 2) if self.do_classifier_free_guidance else latents_packed + ) + # Broadcast timesteps, combine with per-token cond mask (I2V at window head) + timestep = t.expand(latent_model_input.shape[0]) + if cond_mask is not None: + # Broadcast timestep to per-token mask under CFG: [B] -> [B, S, 1] + timestep = timestep[:, None, None] * cond_mask + + # Micro-conditions: only provide video_coords (num_frames/height/width set to 1) + rope_interpolation_scale = ( + self.vae_temporal_compression_ratio, + self.vae_spatial_compression_ratio, + self.vae_spatial_compression_ratio, + ) + # Inpainting pre-blend (ComfyUI parity: KSamplerX0Inpaint:400) + if cond_mask_tokens is not None: + latents_packed = latents_packed * cond_mask_tokens + window_latents_packed * ( + 1.0 - cond_mask_tokens + ) + + # Negative-index/overlap lengths (for segmenting time coordinates; RoPE-compatible) + k_negative_count = ( + 1 if (negative_index_latents is not None and float(negative_index_strength) > 0.0) else 0 + ) + k_overlap_count = overlap_lat if (w_idx > 0 and overlap_lat > 0) else 0 + video_coords = build_video_coords_for_window( + latents=window_latents, + overlap_len=int(k_overlap_count), + guiding_len=int(guiding_len), + negative_len=int(k_negative_count), + rope_interpolation_scale=rope_interpolation_scale, + frame_rate=frame_rate, + ) + with self.transformer.cache_context("cond_uncond"): + noise_pred = self.transformer( + hidden_states=latent_model_input.to(dtype=self.transformer.dtype), + encoder_hidden_states=prompt_embeds, + timestep=timestep, + encoder_attention_mask=prompt_attention_mask, + num_frames=1, + height=1, + width=1, + rope_interpolation_scale=rope_interpolation_scale, + video_coords=video_coords, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + # Unified CFG + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + if self.guidance_rescale > 0: + noise_pred = rescale_noise_cfg( + noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale + ) + + # Use global timestep for scheduling, but apply suppressive blending with hard-condition tokens (e.g., first frame) after step to avoid brightness/flicker due to time misalignment + latents_packed = self.scheduler.step( + noise_pred, t, latents_packed, generator=local_gen, return_dict=False + )[0] + # Inpainting post-blend (ComfyUI parity: restore hard-conditioned regions after update) + if cond_mask_tokens is not None: + latents_packed = latents_packed * cond_mask_tokens + window_latents_packed * ( + 1.0 - cond_mask_tokens + ) + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents_packed = callback_outputs.pop("latents", latents_packed) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + if XLA_AVAILABLE: + xm.mark_step() + else: + # Not enough sigmas to perform a valid step; skip this window safely. + pass + + # 6.7 Unpack back to [B,C,T,H,W] once + window_out = self._unpack_latents( + latents_packed, + w_T_eff, + w_H_eff, + w_W_eff, + p, + pt, + ) + if prev_overlap_len > 0: + window_out = window_out[:, :, :-prev_overlap_len] + + # 6.8 Overlap handling and fusion + if out_latents is None: + # First window: keep all latent frames and cache as AdaIN reference + out_latents = window_out + first_window_latents = out_latents + else: + window_out = window_out[:, :, 1:] # Drop the first frame of the new window + if adain_factor > 0 and first_window_latents is not None: + window_out = adain_normalize_latents(window_out, first_window_latents, adain_factor) + overlap_len = max(overlap_lat - 1, 1) + prev_tail_chunk = out_latents[:, :, -window_out.shape[2] :] + fused = linear_overlap_fuse(prev_tail_chunk, window_out, overlap_len) + out_latents = torch.cat([out_latents[:, :, : -window_out.shape[2]], fused], dim=2) + + # 7. Decode or return latent + if output_type == "latent": + video = out_latents + else: + # Decode via tiling to avoid OOM from full-frame decoding; latents are already de-normalized, so keep auto_denormalize disabled + video = self.vae_decode_tiled( + out_latents, + decode_timestep=decode_timestep, + decode_noise_scale=decode_noise_scale, + horizontal_tiles=int(decode_horizontal_tiles), + vertical_tiles=int(decode_vertical_tiles), + overlap=int(decode_overlap), + generator=generator, + output_type=output_type, # Keep type consistent; postprocess is applied afterwards + ) + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return LTXPipelineOutput(frames=video) diff --git a/diffusers/pipelines/ltx/pipeline_ltx_image2video.py b/diffusers/pipelines/ltx/pipeline_ltx_image2video.py new file mode 100644 index 0000000000000000000000000000000000000000..81ecfce50efa79d7273628c27490dc54e80c14d5 --- /dev/null +++ b/diffusers/pipelines/ltx/pipeline_ltx_image2video.py @@ -0,0 +1,954 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import T5EncoderModel, T5TokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTXVideo +from ...models.transformers import LTXVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import LTXPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTXImageToVideoPipeline + >>> from diffusers.utils import export_to_video, load_image + + >>> pipe = LTXImageToVideoPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/a-r-r-o-w/tiny-meme-dataset-captioned/resolve/main/images/8.png" + ... ) + >>> prompt = "A young girl stands calmly in the foreground, looking directly at the camera, as a house fire rages in the background. Flames engulf the structure, with smoke billowing into the air. Firefighters in protective gear rush to the scene, a fire truck labeled '38' visible behind them. The girl's neutral expression contrasts sharply with the chaos of the fire, creating a poignant and emotionally charged scene." + >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted" + + >>> video = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... width=704, + ... height=480, + ... num_frames=161, + ... num_inference_steps=50, + ... ).frames[0] + >>> export_to_video(video, "output.mp4", fps=24) + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin): + r""" + Pipeline for image-to-video generation. + + Reference: https://github.com/Lightricks/LTX-Video + + Args: + transformer ([`LTXVideoTransformer3DModel`]): + Conditional Transformer architecture to denoise the encoded video latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLLTXVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTXVideo, + text_encoder: T5EncoderModel, + tokenizer: T5TokenizerFast, + transformer: LTXVideoTransformer3DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128 + ) + + self.default_height = 512 + self.default_width = 704 + self.default_frames = 121 + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(text_input_ids.to(device))[0] + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128 + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 128, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + def prepare_latents( + self, + image: torch.Tensor | None = None, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 704, + num_frames: int = 161, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + height = height // self.vae_spatial_compression_ratio + width = width // self.vae_spatial_compression_ratio + num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + + shape = (batch_size, num_channels_latents, num_frames, height, width) + mask_shape = (batch_size, 1, num_frames, height, width) + + if latents is not None: + conditioning_mask = latents.new_zeros(mask_shape) + conditioning_mask[:, :, 0] = 1.0 + conditioning_mask = self._pack_latents( + conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ).squeeze(-1) + if latents.ndim != 3 or latents.shape[:2] != conditioning_mask.shape: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is {conditioning_mask.shape + (num_channels_latents,)}." + ) + return latents.to(device=device, dtype=dtype), conditioning_mask + + if isinstance(generator, list): + if len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i].unsqueeze(0).unsqueeze(2)), generator[i]) + for i in range(batch_size) + ] + else: + init_latents = [ + retrieve_latents(self.vae.encode(img.unsqueeze(0).unsqueeze(2)), generator) for img in image + ] + + init_latents = torch.cat(init_latents, dim=0).to(dtype) + init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std) + init_latents = init_latents.repeat(1, 1, num_frames, 1, 1) + conditioning_mask = torch.zeros(mask_shape, device=device, dtype=dtype) + conditioning_mask[:, :, 0] = 1.0 + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = init_latents * conditioning_mask + noise * (1 - conditioning_mask) + + conditioning_mask = self._pack_latents( + conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ).squeeze(-1) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + + return latents, conditioning_mask + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 704, + num_frames: int = 161, + frame_rate: int = 25, + num_inference_steps: int = 50, + timesteps: list[int] = None, + guidance_scale: float = 3, + guidance_rescale: float = 0.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 128, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + height (`int`, defaults to `512`): + The height in pixels of the generated image. This is set to 480 by default for the best results. + width (`int`, defaults to `704`): + The width in pixels of the generated image. This is set to 848 by default for the best results. + num_frames (`int`, defaults to `161`): + The number of video frames to generate + frame_rate (`int`, defaults to `25`): + Target frame rate of the generated video. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, defaults to `3 `): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + The interpolation factor between random noise and denoised latents at the decode timestep. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `128 `): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare latent variables + if latents is None: + image = self.video_processor.preprocess(image, height=height, width=width) + image = image.to(device=device, dtype=prompt_embeds.dtype) + + num_channels_latents = self.transformer.config.in_channels + latents, conditioning_mask = self.prepare_latents( + image, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + ) + + if self.do_classifier_free_guidance: + conditioning_mask = torch.cat([conditioning_mask, conditioning_mask]) + + # 5. Prepare timesteps + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + video_sequence_length = latent_num_frames * latent_height * latent_width + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) + mu = calculate_shift( + video_sequence_length, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Prepare micro-conditions + rope_interpolation_scale = ( + self.vae_temporal_compression_ratio / frame_rate, + self.vae_spatial_compression_ratio, + self.vae_spatial_compression_ratio, + ) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask) + + with self.transformer.cache_context("cond_uncond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + encoder_hidden_states=prompt_embeds, + timestep=timestep, + encoder_attention_mask=prompt_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + rope_interpolation_scale=rope_interpolation_scale, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred.float() + + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + timestep, _ = timestep.chunk(2) + + if self.guidance_rescale > 0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg( + noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale + ) + + # compute the previous noisy sample x_t -> x_t-1 + noise_pred = self._unpack_latents( + noise_pred, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + noise_pred = noise_pred[:, :, 1:] + noise_latents = latents[:, :, 1:] + pred_latents = self.scheduler.step(noise_pred, t, noise_latents, return_dict=False)[0] + + latents = torch.cat([latents[:, :, :1], pred_latents], dim=2) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + video = latents + else: + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = latents.to(prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = torch.randn(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return LTXPipelineOutput(frames=video) diff --git a/diffusers/pipelines/ltx/pipeline_ltx_latent_upsample.py b/diffusers/pipelines/ltx/pipeline_ltx_latent_upsample.py new file mode 100644 index 0000000000000000000000000000000000000000..fbf98ce976d9b0c6edf4918dea40c824c393bd0a --- /dev/null +++ b/diffusers/pipelines/ltx/pipeline_ltx_latent_upsample.py @@ -0,0 +1,314 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch + +from ...image_processor import PipelineImageInput +from ...models import AutoencoderKLLTXVideo +from ...utils import get_logger +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .modeling_latent_upsampler import LTXLatentUpsamplerModel +from .pipeline_output import LTXPipelineOutput + + +logger = get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class LTXLatentUpsamplePipeline(DiffusionPipeline): + model_cpu_offload_seq = "" + + def __init__( + self, + vae: AutoencoderKLLTXVideo, + latent_upsampler: LTXLatentUpsamplerModel, + ) -> None: + super().__init__() + + self.register_modules(vae=vae, latent_upsampler=latent_upsampler) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + + def prepare_latents( + self, + video: torch.Tensor | None = None, + batch_size: int = 1, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + video = video.to(device=device, dtype=self.vae.dtype) + if isinstance(generator, list): + if len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + init_latents = [ + retrieve_latents(self.vae.encode(video[i].unsqueeze(0)), generator[i]) for i in range(batch_size) + ] + else: + init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator) for vid in video] + + init_latents = torch.cat(init_latents, dim=0).to(dtype) + init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std) + return init_latents + + def adain_filter_latent(self, latents: torch.Tensor, reference_latents: torch.Tensor, factor: float = 1.0): + """ + Applies Adaptive Instance Normalization (AdaIN) to a latent tensor based on statistics from a reference latent + tensor. + + Args: + latent (`torch.Tensor`): + Input latents to normalize + reference_latents (`torch.Tensor`): + The reference latents providing style statistics. + factor (`float`): + Blending factor between original and transformed latent. Range: -10.0 to 10.0, Default: 1.0 + + Returns: + torch.Tensor: The transformed latent tensor + """ + result = latents.clone() + + for i in range(latents.size(0)): + for c in range(latents.size(1)): + r_sd, r_mean = torch.std_mean(reference_latents[i, c], dim=None) # index by original dim order + i_sd, i_mean = torch.std_mean(result[i, c], dim=None) + + result[i, c] = ((result[i, c] - i_mean) / i_sd) * r_sd + r_mean + + result = torch.lerp(latents, result, factor) + return result + + def tone_map_latents(self, latents: torch.Tensor, compression: float) -> torch.Tensor: + """ + Applies a non-linear tone-mapping function to latent values to reduce their dynamic range in a perceptually + smooth way using a sigmoid-based compression. + + This is useful for regularizing high-variance latents or for conditioning outputs during generation, especially + when controlling dynamic behavior with a `compression` factor. + + Args: + latents : torch.Tensor + Input latent tensor with arbitrary shape. Expected to be roughly in [-1, 1] or [0, 1] range. + compression : float + Compression strength in the range [0, 1]. + - 0.0: No tone-mapping (identity transform) + - 1.0: Full compression effect + + Returns: + torch.Tensor + The tone-mapped latent tensor of the same shape as input. + """ + # Remap [0-1] to [0-0.75] and apply sigmoid compression in one shot + scale_factor = compression * 0.75 + abs_latents = torch.abs(latents) + + # Sigmoid compression: sigmoid shifts large values toward 0.2, small values stay ~1.0 + # When scale_factor=0, sigmoid term vanishes, when scale_factor=0.75, full effect + sigmoid_term = torch.sigmoid(4.0 * scale_factor * (abs_latents - 1.0)) + scales = 1.0 - 0.8 * scale_factor * sigmoid_term + + filtered = latents * scales + return filtered + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + def check_inputs(self, video, height, width, latents, tone_map_compression_ratio): + if height % self.vae_spatial_compression_ratio != 0 or width % self.vae_spatial_compression_ratio != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if video is not None and latents is not None: + raise ValueError("Only one of `video` or `latents` can be provided.") + if video is None and latents is None: + raise ValueError("One of `video` or `latents` has to be provided.") + + if not (0 <= tone_map_compression_ratio <= 1): + raise ValueError("`tone_map_compression_ratio` must be in the range [0, 1]") + + @torch.no_grad() + def __call__( + self, + video: list[PipelineImageInput] | None = None, + height: int = 512, + width: int = 704, + latents: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + adain_factor: float = 0.0, + tone_map_compression_ratio: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + ): + r""" + Function invoked when calling the pipeline for latent upsampling. + + Args: + video (`list[PipelineImageInput]`, *optional*): + The input video frames to upsample. Mutually exclusive with `latents`. + height (`int`, defaults to `512`): + The height in pixels of the upsampled output. + width (`int`, defaults to `704`): + The width in pixels of the upsampled output. + latents (`torch.Tensor`, *optional*): + Pre-encoded video latents to upsample. Mutually exclusive with `video`. + decode_timestep (`float` or `list[float]`, defaults to `0.0`): + The timestep at which the upsampled latents are decoded. + decode_noise_scale (`float` or `list[float]`, *optional*): + Interpolation factor between random noise and denoised latents at the decode timestep. + adain_factor (`float`, defaults to `0.0`): + Strength of AdaIN statistical matching applied to the upsampled latents. + tone_map_compression_ratio (`float`, defaults to `0.0`): + Compression ratio used for tone mapping the upsampled latents. Must be in the range [0, 1]. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated video. Choose between `PIL.Image`, `np.array`, or `latent`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple. + """ + self.check_inputs( + video=video, + height=height, + width=width, + latents=latents, + tone_map_compression_ratio=tone_map_compression_ratio, + ) + + if video is not None: + # Batched video input is not yet tested/supported. TODO: take a look later + batch_size = 1 + else: + batch_size = latents.shape[0] + device = self._execution_device + + if video is not None: + num_frames = len(video) + if num_frames % self.vae_temporal_compression_ratio != 1: + num_frames = ( + num_frames // self.vae_temporal_compression_ratio * self.vae_temporal_compression_ratio + 1 + ) + video = video[:num_frames] + logger.warning( + f"Video length expected to be of the form `k * {self.vae_temporal_compression_ratio} + 1` but is {len(video)}. Truncating to {num_frames} frames." + ) + video = self.video_processor.preprocess_video(video, height=height, width=width) + video = video.to(device=device, dtype=torch.float32) + + latents = self.prepare_latents( + video=video, + batch_size=batch_size, + dtype=torch.float32, + device=device, + generator=generator, + latents=latents, + ) + + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = latents.to(self.latent_upsampler.dtype) + latents_upsampled = self.latent_upsampler(latents) + + if adain_factor > 0.0: + latents = self.adain_filter_latent(latents_upsampled, latents, adain_factor) + else: + latents = latents_upsampled + + if tone_map_compression_ratio > 0.0: + latents = self.tone_map_latents(latents, tone_map_compression_ratio) + + if output_type == "latent": + latents = self._normalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + video = latents + else: + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return LTXPipelineOutput(frames=video) diff --git a/diffusers/pipelines/ltx/pipeline_output.py b/diffusers/pipelines/ltx/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..f5cb34aa508dc8c6e570f310cbd7615d4cd19154 --- /dev/null +++ b/diffusers/pipelines/ltx/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class LTXPipelineOutput(BaseOutput): + r""" + Output class for LTX pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/diffusers/pipelines/ltx2/__init__.py b/diffusers/pipelines/ltx2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..cc920c1411fa6cbd2279ac062d242f4b016d9fb3 --- /dev/null +++ b/diffusers/pipelines/ltx2/__init__.py @@ -0,0 +1,66 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["connectors"] = ["LTX2TextConnectors"] + _import_structure["image_processor"] = ["LTX2VideoHDRProcessor"] + _import_structure["latent_upsampler"] = ["LTX2LatentUpsamplerModel"] + _import_structure["pipeline_ltx2"] = ["LTX2Pipeline"] + _import_structure["pipeline_ltx2_condition"] = ["LTX2ConditionPipeline", "LTX2VideoCondition"] + _import_structure["pipeline_ltx2_hdr_lora"] = ["LTX2HDRPipeline", "LTX2HDRReferenceCondition"] + _import_structure["pipeline_ltx2_ic_lora"] = ["LTX2InContextPipeline", "LTX2ReferenceCondition"] + _import_structure["pipeline_ltx2_image2video"] = ["LTX2ImageToVideoPipeline"] + _import_structure["pipeline_ltx2_latent_upsample"] = ["LTX2LatentUpsamplePipeline"] + _import_structure["vocoder"] = ["LTX2Vocoder", "LTX2VocoderWithBWE"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .connectors import LTX2TextConnectors + from .image_processor import LTX2VideoHDRProcessor + from .latent_upsampler import LTX2LatentUpsamplerModel + from .pipeline_ltx2 import LTX2Pipeline + from .pipeline_ltx2_condition import LTX2ConditionPipeline, LTX2VideoCondition + from .pipeline_ltx2_hdr_lora import LTX2HDRPipeline, LTX2HDRReferenceCondition + from .pipeline_ltx2_ic_lora import LTX2InContextPipeline, LTX2ReferenceCondition + from .pipeline_ltx2_image2video import LTX2ImageToVideoPipeline + from .pipeline_ltx2_latent_upsample import LTX2LatentUpsamplePipeline + from .vocoder import LTX2Vocoder, LTX2VocoderWithBWE + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/ltx2/connectors.py b/diffusers/pipelines/ltx2/connectors.py new file mode 100644 index 0000000000000000000000000000000000000000..0b0b333209aa9c1cf226f4eab64781fc7715c45c --- /dev/null +++ b/diffusers/pipelines/ltx2/connectors.py @@ -0,0 +1,478 @@ +import math + +import torch +import torch.nn as nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...loaders import PeftAdapterMixin +from ...models.attention import FeedForward +from ...models.modeling_utils import ModelMixin +from ...models.transformers.transformer_ltx2 import LTX2Attention, LTX2AudioVideoAttnProcessor + + +def per_layer_masked_mean_norm( + text_hidden_states: torch.Tensor, + sequence_lengths: torch.Tensor, + device: str | torch.device, + padding_side: str = "left", + scale_factor: int = 8, + eps: float = 1e-6, +): + """ + Performs per-batch per-layer normalization using a masked mean and range on per-layer text encoder hidden_states. + Respects the padding of the hidden states. + + Args: + text_hidden_states (`torch.Tensor` of shape `(batch_size, seq_len, hidden_dim, num_layers)`): + Per-layer hidden_states from a text encoder (e.g. `Gemma3ForConditionalGeneration`). + sequence_lengths (`torch.Tensor of shape `(batch_size,)`): + The number of valid (non-padded) tokens for each batch instance. + device: (`str` or `torch.device`, *optional*): + torch device to place the resulting embeddings on + padding_side: (`str`, *optional*, defaults to `"left"`): + Whether the text tokenizer performs padding on the `"left"` or `"right"`. + scale_factor (`int`, *optional*, defaults to `8`): + Scaling factor to multiply the normalized hidden states by. + eps (`float`, *optional*, defaults to `1e-6`): + A small positive value for numerical stability when performing normalization. + + Returns: + `torch.Tensor` of shape `(batch_size, seq_len, hidden_dim * num_layers)`: + Normed and flattened text encoder hidden states. + """ + batch_size, seq_len, hidden_dim, num_layers = text_hidden_states.shape + original_dtype = text_hidden_states.dtype + + # Create padding mask + token_indices = torch.arange(seq_len, device=device).unsqueeze(0) + if padding_side == "right": + # For right padding, valid tokens are from 0 to sequence_length-1 + mask = token_indices < sequence_lengths[:, None] # [batch_size, seq_len] + elif padding_side == "left": + # For left padding, valid tokens are from (T - sequence_length) to T-1 + start_indices = seq_len - sequence_lengths[:, None] # [batch_size, 1] + mask = token_indices >= start_indices # [B, T] + else: + raise ValueError(f"padding_side must be 'left' or 'right', got {padding_side}") + mask = mask[:, :, None, None] # [batch_size, seq_len] --> [batch_size, seq_len, 1, 1] + + # Compute masked mean over non-padding positions of shape (batch_size, 1, 1, seq_len) + masked_text_hidden_states = text_hidden_states.masked_fill(~mask, 0.0) + num_valid_positions = (sequence_lengths * hidden_dim).view(batch_size, 1, 1, 1) + masked_mean = masked_text_hidden_states.sum(dim=(1, 2), keepdim=True) / (num_valid_positions + eps) + + # Compute min/max over non-padding positions of shape (batch_size, 1, 1 seq_len) + x_min = text_hidden_states.masked_fill(~mask, float("inf")).amin(dim=(1, 2), keepdim=True) + x_max = text_hidden_states.masked_fill(~mask, float("-inf")).amax(dim=(1, 2), keepdim=True) + + # Normalization + normalized_hidden_states = (text_hidden_states - masked_mean) / (x_max - x_min + eps) + normalized_hidden_states = normalized_hidden_states * scale_factor + + # Pack the hidden states to a 3D tensor (batch_size, seq_len, hidden_dim * num_layers) + normalized_hidden_states = normalized_hidden_states.flatten(2) + mask_flat = mask.squeeze(-1).expand(-1, -1, hidden_dim * num_layers) + normalized_hidden_states = normalized_hidden_states.masked_fill(~mask_flat, 0.0) + normalized_hidden_states = normalized_hidden_states.to(dtype=original_dtype) + return normalized_hidden_states + + +def per_token_rms_norm(text_encoder_hidden_states: torch.Tensor, eps: float = 1e-6) -> torch.Tensor: + variance = torch.mean(text_encoder_hidden_states**2, dim=2, keepdim=True) + norm_text_encoder_hidden_states = text_encoder_hidden_states * torch.rsqrt(variance + eps) + return norm_text_encoder_hidden_states + + +class LTX2RotaryPosEmbed1d(nn.Module): + """ + 1D rotary positional embeddings (RoPE) for the LTX 2.0 text encoder connectors. + """ + + def __init__( + self, + dim: int, + base_seq_len: int = 4096, + theta: float = 10000.0, + double_precision: bool = True, + rope_type: str = "interleaved", + num_attention_heads: int = 32, + ): + super().__init__() + if rope_type not in ["interleaved", "split"]: + raise ValueError(f"{rope_type=} not supported. Choose between 'interleaved' and 'split'.") + + self.dim = dim + self.base_seq_len = base_seq_len + self.theta = theta + self.double_precision = double_precision + self.rope_type = rope_type + self.num_attention_heads = num_attention_heads + + def forward( + self, + batch_size: int, + pos: int, + device: str | torch.device, + ) -> tuple[torch.Tensor, torch.Tensor]: + # 1. Get 1D position ids + grid_1d = torch.arange(pos, dtype=torch.float32, device=device) + # Get fractional indices relative to self.base_seq_len + grid_1d = grid_1d / self.base_seq_len + grid = grid_1d.unsqueeze(0).repeat(batch_size, 1) # [batch_size, seq_len] + + # 2. Calculate 1D RoPE frequencies + num_rope_elems = 2 # 1 (because 1D) * 2 (for cos, sin) = 2 + freqs_dtype = torch.float64 if self.double_precision else torch.float32 + pow_indices = torch.pow( + self.theta, + torch.linspace(start=0.0, end=1.0, steps=self.dim // num_rope_elems, dtype=freqs_dtype, device=device), + ) + freqs = (pow_indices * torch.pi / 2.0).to(dtype=torch.float32) + + # 3. Matrix-vector outer product between pos ids of shape (batch_size, seq_len) and freqs vector of shape + # (self.dim // 2,). + freqs = (grid.unsqueeze(-1) * 2 - 1) * freqs # [B, seq_len, self.dim // 2] + + # 4. Get real, interleaved (cos, sin) frequencies, padded to self.dim + if self.rope_type == "interleaved": + cos_freqs = freqs.cos().repeat_interleave(2, dim=-1) + sin_freqs = freqs.sin().repeat_interleave(2, dim=-1) + + if self.dim % num_rope_elems != 0: + cos_padding = torch.ones_like(cos_freqs[:, :, : self.dim % num_rope_elems]) + sin_padding = torch.zeros_like(sin_freqs[:, :, : self.dim % num_rope_elems]) + cos_freqs = torch.cat([cos_padding, cos_freqs], dim=-1) + sin_freqs = torch.cat([sin_padding, sin_freqs], dim=-1) + + elif self.rope_type == "split": + expected_freqs = self.dim // 2 + current_freqs = freqs.shape[-1] + pad_size = expected_freqs - current_freqs + cos_freq = freqs.cos() + sin_freq = freqs.sin() + + if pad_size != 0: + cos_padding = torch.ones_like(cos_freq[:, :, :pad_size]) + sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size]) + + cos_freq = torch.concatenate([cos_padding, cos_freq], axis=-1) + sin_freq = torch.concatenate([sin_padding, sin_freq], axis=-1) + + # Reshape freqs to be compatible with multi-head attention + b = cos_freq.shape[0] + t = cos_freq.shape[1] + + cos_freq = cos_freq.reshape(b, t, self.num_attention_heads, -1) + sin_freq = sin_freq.reshape(b, t, self.num_attention_heads, -1) + + cos_freqs = torch.swapaxes(cos_freq, 1, 2) # (B,H,T,D//2) + sin_freqs = torch.swapaxes(sin_freq, 1, 2) # (B,H,T,D//2) + + return cos_freqs, sin_freqs + + +class LTX2TransformerBlock1d(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + activation_fn: str = "gelu-approximate", + eps: float = 1e-6, + rope_type: str = "interleaved", + apply_gated_attention: bool = False, + ): + super().__init__() + + self.norm1 = torch.nn.RMSNorm(dim, eps=eps, elementwise_affine=False) + self.attn1 = LTX2Attention( + query_dim=dim, + heads=num_attention_heads, + kv_heads=num_attention_heads, + dim_head=attention_head_dim, + rope_type=rope_type, + apply_gated_attention=apply_gated_attention, + processor=LTX2AudioVideoAttnProcessor(), + ) + + self.norm2 = torch.nn.RMSNorm(dim, eps=eps, elementwise_affine=False) + self.ff = FeedForward(dim, activation_fn=activation_fn) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: torch.Tensor | None = None, + rotary_emb: torch.Tensor | None = None, + ) -> torch.Tensor: + norm_hidden_states = self.norm1(hidden_states) + attn_hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask, query_rotary_emb=rotary_emb) + hidden_states = hidden_states + attn_hidden_states + + norm_hidden_states = self.norm2(hidden_states) + ff_hidden_states = self.ff(norm_hidden_states) + hidden_states = hidden_states + ff_hidden_states + + return hidden_states + + +class LTX2ConnectorTransformer1d(nn.Module): + """ + A 1D sequence transformer for modalities such as text. + + In LTX 2.0, this is used to process the text encoder hidden states for each of the video and audio streams. + """ + + _supports_gradient_checkpointing = True + + def __init__( + self, + num_attention_heads: int = 30, + attention_head_dim: int = 128, + num_layers: int = 2, + num_learnable_registers: int | None = 128, + rope_base_seq_len: int = 4096, + rope_theta: float = 10000.0, + rope_double_precision: bool = True, + eps: float = 1e-6, + causal_temporal_positioning: bool = False, + rope_type: str = "interleaved", + gated_attention: bool = False, + ): + super().__init__() + self.num_attention_heads = num_attention_heads + self.inner_dim = num_attention_heads * attention_head_dim + self.causal_temporal_positioning = causal_temporal_positioning + + self.num_learnable_registers = num_learnable_registers + self.learnable_registers = None + if num_learnable_registers is not None: + init_registers = torch.rand(num_learnable_registers, self.inner_dim) * 2.0 - 1.0 + self.learnable_registers = torch.nn.Parameter(init_registers) + + self.rope = LTX2RotaryPosEmbed1d( + self.inner_dim, + base_seq_len=rope_base_seq_len, + theta=rope_theta, + double_precision=rope_double_precision, + rope_type=rope_type, + num_attention_heads=num_attention_heads, + ) + + self.transformer_blocks = torch.nn.ModuleList( + [ + LTX2TransformerBlock1d( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + rope_type=rope_type, + apply_gated_attention=gated_attention, + ) + for _ in range(num_layers) + ] + ) + + self.norm_out = torch.nn.RMSNorm(self.inner_dim, eps=eps, elementwise_affine=False) + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: torch.Tensor | None = None, + attn_mask_binarize_threshold: float = -9000.0, + ) -> tuple[torch.Tensor, torch.Tensor]: + # hidden_states shape: [batch_size, seq_len, hidden_dim] + # attention_mask shape: [batch_size, seq_len] or [batch_size, 1, 1, seq_len] + batch_size, seq_len, _ = hidden_states.shape + + # 1. Replace padding with learned registers, if using + if self.learnable_registers is not None: + if seq_len % self.num_learnable_registers != 0: + raise ValueError( + f"The `hidden_states` sequence length {hidden_states.shape[1]} should be divisible by the number" + f" of learnable registers {self.num_learnable_registers}" + ) + + num_register_repeats = seq_len // self.num_learnable_registers + registers = ( + self.learnable_registers.unsqueeze(0).expand(num_register_repeats, -1, -1).reshape(seq_len, -1) + ) # [seq_len, inner_dim] + + binary_attn_mask = (attention_mask >= attn_mask_binarize_threshold).int() + if binary_attn_mask.ndim == 4: + binary_attn_mask = binary_attn_mask.squeeze(1).squeeze(1) # [B, 1, 1, L] --> [B, L] + + # Move the valid tokens to the front in their original order and fill the tail + # with registers indexed by absolute position, matching the original LTX + # implementation (`_replace_padded_with_learnable_registers`). A stable argsort + # of the inverted mask gathers valid tokens first while preserving their order. + order = torch.argsort(1 - binary_attn_mask, dim=1, stable=True) # [B, L] + front_aligned = torch.gather(hidden_states, 1, order.unsqueeze(-1).expand(-1, -1, hidden_states.shape[-1])) + num_valid = binary_attn_mask.sum(dim=1, keepdim=True) # [B, 1] + positions = torch.arange(seq_len, device=hidden_states.device).unsqueeze(0) # [1, L] + front_mask = (positions < num_valid).unsqueeze(-1) # [B, L, 1] + registers_expanded = registers.unsqueeze(0).expand(batch_size, -1, -1) # [B, L, D] + hidden_states = torch.where(front_mask, front_aligned, registers_expanded.to(hidden_states.dtype)) + + # Overwrite attention_mask with an all-zeros mask if using registers. + attention_mask = torch.zeros_like(attention_mask) + + # 2. Calculate 1D RoPE positional embeddings + rotary_emb = self.rope(batch_size, seq_len, device=hidden_states.device) + + # 3. Run 1D transformer blocks + for block in self.transformer_blocks: + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func(block, hidden_states, attention_mask, rotary_emb) + else: + hidden_states = block(hidden_states, attention_mask=attention_mask, rotary_emb=rotary_emb) + + hidden_states = self.norm_out(hidden_states) + + return hidden_states, attention_mask + + +class LTX2TextConnectors(ModelMixin, PeftAdapterMixin, ConfigMixin): + """ + Text connector stack used by LTX 2.0 to process the packed text encoder hidden states for both the video and audio + streams. + """ + + @register_to_config + def __init__( + self, + caption_channels: int = 3840, # default Gemma-3-12B text encoder hidden_size + text_proj_in_factor: int = 49, # num_layers + 1 for embedding layer = 48 + 1 for Gemma-3-12B + video_connector_num_attention_heads: int = 30, + video_connector_attention_head_dim: int = 128, + video_connector_num_layers: int = 2, + video_connector_num_learnable_registers: int | None = 128, + video_gated_attn: bool = False, + audio_connector_num_attention_heads: int = 30, + audio_connector_attention_head_dim: int = 128, + audio_connector_num_layers: int = 2, + audio_connector_num_learnable_registers: int | None = 128, + audio_gated_attn: bool = False, + connector_rope_base_seq_len: int = 4096, + rope_theta: float = 10000.0, + rope_double_precision: bool = True, + causal_temporal_positioning: bool = False, + rope_type: str = "interleaved", + per_modality_projections: bool = False, + video_hidden_dim: int = 4096, + audio_hidden_dim: int = 2048, + proj_bias: bool = False, + ): + super().__init__() + text_encoder_dim = caption_channels * text_proj_in_factor + if per_modality_projections: + self.video_text_proj_in = nn.Linear(text_encoder_dim, video_hidden_dim, bias=proj_bias) + self.audio_text_proj_in = nn.Linear(text_encoder_dim, audio_hidden_dim, bias=proj_bias) + else: + self.text_proj_in = nn.Linear(text_encoder_dim, caption_channels, bias=proj_bias) + + self.video_connector = LTX2ConnectorTransformer1d( + num_attention_heads=video_connector_num_attention_heads, + attention_head_dim=video_connector_attention_head_dim, + num_layers=video_connector_num_layers, + num_learnable_registers=video_connector_num_learnable_registers, + rope_base_seq_len=connector_rope_base_seq_len, + rope_theta=rope_theta, + rope_double_precision=rope_double_precision, + causal_temporal_positioning=causal_temporal_positioning, + rope_type=rope_type, + gated_attention=video_gated_attn, + ) + self.audio_connector = LTX2ConnectorTransformer1d( + num_attention_heads=audio_connector_num_attention_heads, + attention_head_dim=audio_connector_attention_head_dim, + num_layers=audio_connector_num_layers, + num_learnable_registers=audio_connector_num_learnable_registers, + rope_base_seq_len=connector_rope_base_seq_len, + rope_theta=rope_theta, + rope_double_precision=rope_double_precision, + causal_temporal_positioning=causal_temporal_positioning, + rope_type=rope_type, + gated_attention=audio_gated_attn, + ) + + def forward( + self, + text_encoder_hidden_states: torch.Tensor, + attention_mask: torch.Tensor, + padding_side: str = "left", + scale_factor: int = 8, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Given per-layer text encoder hidden_states, extracts features and runs per-modality connectors to get text + embeddings for the LTX-2.X DiT models. + + Args: + text_encoder_hidden_states (`torch.Tensor`)): + Per-layer text encoder hidden_states. Can either be 4D with shape `(batch_size, seq_len, + caption_channels, text_proj_in_factor) or 3D with the last two dimensions flattened. + attention_mask (`torch.Tensor` of shape `(batch_size, seq_len)`): + Multiplicative binary attention mask where 1s indicate unmasked positions and 0s indicate masked + positions. + padding_side (`str`, *optional*, defaults to `"left"`): + The padding side used by the text encoder's text encoder (either `"left"` or `"right"`). Defaults to + `"left"` as this is what the default Gemma3-12B text encoder uses. Only used if + `per_modality_projections` is `False` (LTX-2.0 models). + scale_factor (`int`, *optional*, defaults to `8`): + Scale factor for masked mean/range normalization. Only used if `per_modality_projections` is `False` + (LTX-2.0 models). + """ + if text_encoder_hidden_states.ndim == 3: + # Ensure shape is [batch_size, seq_len, caption_channels, text_proj_in_factor] + text_encoder_hidden_states = text_encoder_hidden_states.unflatten(2, (self.config.caption_channels, -1)) + + if self.config.per_modality_projections: + # LTX-2.3 + norm_text_encoder_hidden_states = per_token_rms_norm(text_encoder_hidden_states) + + norm_text_encoder_hidden_states = norm_text_encoder_hidden_states.flatten(2, 3) + bool_mask = attention_mask.bool().unsqueeze(-1) + norm_text_encoder_hidden_states = torch.where( + bool_mask, norm_text_encoder_hidden_states, torch.zeros_like(norm_text_encoder_hidden_states) + ) + + # Rescale norms with respect to video and audio dims for feature extractors + video_scale_factor = math.sqrt(self.config.video_hidden_dim / self.config.caption_channels) + video_norm_text_emb = norm_text_encoder_hidden_states * video_scale_factor + audio_scale_factor = math.sqrt(self.config.audio_hidden_dim / self.config.caption_channels) + audio_norm_text_emb = norm_text_encoder_hidden_states * audio_scale_factor + + # Per-Modality Feature extractors + video_text_emb_proj = self.video_text_proj_in(video_norm_text_emb) + audio_text_emb_proj = self.audio_text_proj_in(audio_norm_text_emb) + else: + # LTX-2.0 + sequence_lengths = attention_mask.sum(dim=-1) + norm_text_encoder_hidden_states = per_layer_masked_mean_norm( + text_hidden_states=text_encoder_hidden_states, + sequence_lengths=sequence_lengths, + device=text_encoder_hidden_states.device, + padding_side=padding_side, + scale_factor=scale_factor, + ) + + text_emb_proj = self.text_proj_in(norm_text_encoder_hidden_states) + video_text_emb_proj = text_emb_proj + audio_text_emb_proj = text_emb_proj + + # Convert to additive attention mask for connectors + text_dtype = video_text_emb_proj.dtype + attention_mask = (attention_mask.to(torch.int64) - 1).to(text_dtype) + attention_mask = attention_mask.reshape(attention_mask.shape[0], 1, -1, attention_mask.shape[-1]) + add_attn_mask = attention_mask * torch.finfo(text_dtype).max + + video_text_embedding, video_attn_mask = self.video_connector(video_text_emb_proj, add_attn_mask) + + # Convert video attn mask to binary (multiplicative) mask and mask video text embedding + binary_attn_mask = (video_attn_mask < 1e-6).to(torch.int64) + binary_attn_mask = binary_attn_mask.reshape(video_text_embedding.shape[0], video_text_embedding.shape[1], 1) + video_text_embedding = video_text_embedding * binary_attn_mask + + audio_text_embedding, _ = self.audio_connector(audio_text_emb_proj, add_attn_mask) + + return video_text_embedding, audio_text_embedding, binary_attn_mask.squeeze(-1) diff --git a/diffusers/pipelines/ltx2/export_utils.py b/diffusers/pipelines/ltx2/export_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..40a633b13f9f4938b881cf9c024d48a4087c94f8 --- /dev/null +++ b/diffusers/pipelines/ltx2/export_utils.py @@ -0,0 +1,125 @@ +# Copyright 2025 The Lightricks team and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# `encode_video` moved to `diffusers.utils.export_utils` so other pipelines (Cosmos3) +# can share it. This module remains as a deprecation shim for existing user code that +# does `from diffusers.pipelines.ltx2.export_utils import encode_video`. + +from fractions import Fraction +from pathlib import Path +from typing import Callable + +import numpy as np +import torch + +from ...utils import deprecate, is_av_available +from ...utils.export_utils import encode_video as _encode_video + + +_CAN_USE_AV = is_av_available() +if _CAN_USE_AV: + import av +else: + raise ImportError( + "PyAV is required to use LTX 2.0 video export utilities. You can install it with `pip install av`" + ) + + +def encode_video(*args, **kwargs): + deprecate( + "encode_video", + "0.40.0", + "`encode_video` has moved to `diffusers.utils`. Import it with " + "`from diffusers.utils import encode_video` instead.", + ) + return _encode_video(*args, **kwargs) + + +def encode_hdr_tensor_to_mp4( + frames: torch.Tensor | np.ndarray, + output_mp4: str | Path, + frame_rate: float, + tone_mapping_fn: Callable[[np.ndarray], np.ndarray] | None = None, + tone_map_in_rgb: bool = True, + crf: int = 18, +) -> None: + """ + Converts a linear HDR tensor (for example, as outputted by `LTX2HDRPipeline`) to a SDR `.mp4` file (specifically, a + sRGB-tonemapped H.264 `.mp4`). + + Args: + frames (`torch.Tensor` or `np.ndarray`): + A linear HDR tensors with RGB values in `[0, ∞)` of shape `(F, H, W, 3)`. + output_mp4 (`str` or `pathlib.Path`): + Output MP4 path. + frame_rate (`float`): + Frame rate for the output video. + tone_mapping_fn (`Callable[[np.ndarray], np.ndarray]`, *optional*, defaults to `None`): + An optional tone mapping function which takes a float32 NumPy array of shape `(H, W, 3)` containing linear + HDR values in `[0, ∞)` and returns tone-mapped linear values in `[0, 1]`. The sRGB transfer function (OETF) + is applied afterwards — do **not** pre-apply gamma inside this function. If `None`, defaults to + [`simple_tone_map`], which clips values above `1.0`. The channel ordering of the input array is controlled + by `tone_map_in_rgb`: RGB by default (matching the `LTX2HDRPipeline` output), or BGR when + `tone_map_in_rgb=False`. This is the opposite default to `encode_exr_sequence_to_mp4`. + tone_map_in_rgb (`bool`, *optional*, defaults to `True`): + When `True` (default), frames are passed as RGB to `tone_mapping_fn`, and the output frame is tagged as + `rgb24`. Use this when `tone_mapping_fn` expects RGB input (e.g. operators from `colour-science`). When + `False`, the frames first have their channels flipped to BGR, which is the native format for + `opencv-python` tone mappers (e.g. `cv2.createTonemapReinhard().process`). Note that this is the opposite + default to `encode_exr_sequence_to_mp4`. + crf (`int`, *optional*, defaults to `18`): + libx264 CRF quality factor. Lower values produce higher quality. + """ + if isinstance(frames, torch.Tensor): + frames = frames.cpu().float().numpy() + + container = av.open(str(output_mp4), mode="w") + stream = container.add_stream("libx264", rate=Fraction(frame_rate).limit_denominator(1000)) + stream.pix_fmt = "yuv420p" + stream.options = {"crf": str(crf), "movflags": "+faststart"} + + pix_fmt = "rgb24" if tone_map_in_rgb else "bgr24" + if tone_mapping_fn is None: + # Default to simple tone mapping function which clips values above 1.0 to 1.0. This is what the original + # LTX-2.X code does, but you may want to do some non-trivial tone-mapping to make the sample look better. + def simple_tone_map(x: np.ndarray) -> np.ndarray: + return np.clip(x, 0.0, 1.0) + + tone_mapping_fn = simple_tone_map + + try: + for i, hdr in enumerate(frames): + if not tone_map_in_rgb: + hdr = hdr[..., ::-1] + hdr_mapped = tone_mapping_fn(hdr) + + hdr_mapped = np.clip(hdr_mapped, 0.0, 1.0) # Clamp to [0, 1] in case tone mapper does not + # Apply the sRBG (Rec.709 OETF) transfer function to linear light in [0, 1] + sdr = np.where( + hdr_mapped <= 0.0031308, hdr_mapped * 12.92, 1.055 * np.power(hdr_mapped, 1.0 / 2.4) - 0.055 + ) + out8 = (sdr * 255.0 + 0.5).astype(np.uint8) + + if i == 0: + stream.height, stream.width = out8.shape[:2] + + frame = av.VideoFrame.from_ndarray(out8, format=pix_fmt) + for packet in stream.encode(frame): + container.mux(packet) + + for packet in stream.encode(): + container.mux(packet) + finally: + container.close() diff --git a/diffusers/pipelines/ltx2/image_processor.py b/diffusers/pipelines/ltx2/image_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..a256600739438c670815db233ea905abeee050cb --- /dev/null +++ b/diffusers/pipelines/ltx2/image_processor.py @@ -0,0 +1,175 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import torch +import torch.nn.functional as F + +from ...configuration_utils import register_to_config +from ...utils import logging +from ...video_processor import VideoProcessor + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class LTX2VideoHDRProcessor(VideoProcessor): + r""" + Video processor for the LTX-2 HDR IC-LoRA pipeline. + + Inherits standard video preprocessing from [`VideoProcessor`] and additionally supports: + + - `preprocess_reference_video_hdr`: aspect-ratio-preserving resize followed by reflect-padding to the target size. + For LDR (SDR Rec.709) reference videos, `LogC3.compress_ldr` is an identity clamp, so the numerical output is + equivalent to the standard [-1, 1] normalization used by [`VideoProcessor.preprocess_video`] — only the resize + strategy differs (reflect-pad vs center-crop). + - `postprocess_hdr_video`: applies the LogC3 inverse transform to the VAE's decoded output, mapping `[0, 1]` → + linear HDR `[0, ∞)`. + + Args: + vae_scale_factor (`int`, *optional*, defaults to `32`): + VAE (spatial) scale factor for the LTX-2 video VAE. + resample (`str`, *optional*, defaults to `"bilinear"`): + Resampling filter used by the base [`VaeImageProcessor`] for PIL/tensor resizing. + hdr_transform (`str`, *optional*, defaults to `"logc3"`): + HDR transform identifier. Only `"logc3"` (ARRI EI 800) is currently supported. + """ + + # LogC3 (ARRI EI 800) coefficients, ported from `ltx_core.hdr.LogC3`. + _LOGC3_A = 5.555556 + _LOGC3_B = 0.052272 + _LOGC3_C = 0.247190 + _LOGC3_D = 0.385537 + _LOGC3_E = 5.367655 + _LOGC3_F = 0.092809 + _LOGC3_CUT = 0.010591 + + @register_to_config + def __init__( + self, + vae_scale_factor: int = 32, + resample: str = "bilinear", + hdr_transform: str = "logc3", + ): + super().__init__( + do_resize=True, + vae_scale_factor=vae_scale_factor, + resample=resample, + ) + if hdr_transform != "logc3": + raise ValueError(f"Unsupported HDR transform {hdr_transform!r}. Only 'logc3' is supported.") + + @classmethod + def _logc3_decompress(cls, logc: torch.Tensor) -> torch.Tensor: + r"""Decompress LogC3 `[0, 1]` → linear HDR `[0, ∞)`.""" + logc = torch.clamp(logc, 0.0, 1.0) + cut_log = cls._LOGC3_E * cls._LOGC3_CUT + cls._LOGC3_F + lin_from_log = (torch.pow(10.0, (logc - cls._LOGC3_D) / cls._LOGC3_C) - cls._LOGC3_B) / cls._LOGC3_A + lin_from_lin = (logc - cls._LOGC3_F) / cls._LOGC3_E + return torch.where(logc >= cut_log, lin_from_log, lin_from_lin) + + @staticmethod + def _resize_and_reflect_pad_video(video: torch.Tensor, height: int, width: int) -> torch.Tensor: + r""" + Resize a video tensor preserving aspect ratio, then reflect-pad to the exact target dimensions. + + Args: + video (`torch.Tensor`): Input of shape `(B, C, F, H, W)`. + height (`int`), width (`int`): Target spatial dimensions. + + Returns: + `torch.Tensor`: Resized and padded video of shape `(B, C, F, height, width)`. + """ + b, c, f, src_h, src_w = video.shape + + if height >= src_h and width >= src_w: + new_h, new_w = src_h, src_w + else: + scale = min(height / src_h, width / src_w) + new_h = round(src_h * scale) + new_w = round(src_w * scale) + # (B, C, F, H, W) → (B, F, C, H, W) → (B*F, C, H, W) for 2D per-frame interpolation. + video = video.permute(0, 2, 1, 3, 4).reshape(b * f, c, src_h, src_w) + video = F.interpolate(video, size=(new_h, new_w), mode="bilinear", align_corners=False) + video = video.reshape(b, f, c, new_h, new_w).permute(0, 2, 1, 3, 4) + + pad_bottom = height - new_h + pad_right = width - new_w + if pad_bottom > 0 or pad_right > 0: + # `reflect` pad requires the pad amount to be strictly less than the corresponding input dim. + pad_mode = "reflect" if pad_bottom < new_h and pad_right < new_w else "replicate" + video = video.permute(0, 2, 1, 3, 4).reshape(b * f, c, new_h, new_w) + video = F.pad(video, (0, pad_right, 0, pad_bottom), mode=pad_mode) + video = video.reshape(b, f, c, height, width).permute(0, 2, 1, 3, 4) + + return video + + def preprocess_reference_video_hdr( + self, + video, + height: int, + width: int, + ) -> torch.Tensor: + r""" + Preprocess a reference (SDR) video for HDR IC-LoRA conditioning. + + Runs the input through the standard video preprocessing (normalization to `[-1, 1]`) without resizing, then + applies reflect-pad resize to the target dimensions. For LDR inputs this is numerically equivalent to + `load_video_conditioning_hdr` in the reference implementation (since `LogC3.compress_ldr` is an identity clamp + on `[0, 1]` inputs). + + Args: + video: Input accepted by `VideoProcessor.preprocess_video` (list of PIL images, 4D/5D tensor/array, etc.). + height (`int`), width (`int`): Target spatial dimensions. + + Returns: + `torch.Tensor`: Preprocessed video of shape `(B, C, F, height, width)` with values in `[-1, 1]`. + """ + video = self.preprocess_video(video, height=None, width=None) # (B, C, F, src_h, src_w) in [-1, 1] + video = self._resize_and_reflect_pad_video(video, height, width) + return video + + def postprocess_hdr_video(self, video: torch.Tensor, output_type: str = "np") -> torch.Tensor | np.ndarray: + r""" + Postprocess the VAE's decoded output to linear HDR. + + Args: + video (`torch.Tensor`): + VAE decoded output in VAE range `[-1, 1]`, shape `(B, C, F, H, W)`. + output_type (`str`, *optional*, defaults to `"np"`): + Output type of post-processed video tensor; should be in `["np", "pt"]`. + + Returns: + Returns linear HDR video with values in `[0, ∞)`, depending on `output_type`: + - `output_type="pt"`: `torch.Tensor` with shape `(B, F, H, W, C)` and dtype `float32`. + - `output_type="np"`: `np.ndarray` with shape `(B, F, H, W, C)` and dtype `float32`. + """ + if output_type not in ["np", "pt"]: + logger.warning( + f"output_type {output_type} is not supported for LTX-2.X HDR postprocessing. Supported types are `np`" + f" and `pt`; the output_type will be set to `np`." + ) + output_type = "np" + + video = self.denormalize(video.float()) + # Apply the inverse transform function to get linear HDR light + video = self._logc3_decompress(video) + + # Permute to channels-last: [B, C, F, H, W] --> [B, F, H, W, C] + video = video = video.permute(0, 2, 3, 4, 1).contiguous() + if output_type == "pt": + return video + + video = video.cpu().numpy() + return video diff --git a/diffusers/pipelines/ltx2/latent_upsampler.py b/diffusers/pipelines/ltx2/latent_upsampler.py new file mode 100644 index 0000000000000000000000000000000000000000..329ced36d45b6d804424102a6acdafbca2b798a2 --- /dev/null +++ b/diffusers/pipelines/ltx2/latent_upsampler.py @@ -0,0 +1,285 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math + +import torch +import torch.nn.functional as F + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin + + +RATIONAL_RESAMPLER_SCALE_MAPPING = { + 0.75: (3, 4), + 1.5: (3, 2), + 2.0: (2, 1), + 4.0: (4, 1), +} + + +# Copied from diffusers.pipelines.ltx.modeling_latent_upsampler.ResBlock +class ResBlock(torch.nn.Module): + def __init__(self, channels: int, mid_channels: int | None = None, dims: int = 3): + super().__init__() + if mid_channels is None: + mid_channels = channels + + Conv = torch.nn.Conv2d if dims == 2 else torch.nn.Conv3d + + self.conv1 = Conv(channels, mid_channels, kernel_size=3, padding=1) + self.norm1 = torch.nn.GroupNorm(32, mid_channels) + self.conv2 = Conv(mid_channels, channels, kernel_size=3, padding=1) + self.norm2 = torch.nn.GroupNorm(32, channels) + self.activation = torch.nn.SiLU() + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + residual = hidden_states + hidden_states = self.conv1(hidden_states) + hidden_states = self.norm1(hidden_states) + hidden_states = self.activation(hidden_states) + hidden_states = self.conv2(hidden_states) + hidden_states = self.norm2(hidden_states) + hidden_states = self.activation(hidden_states + residual) + return hidden_states + + +# Copied from diffusers.pipelines.ltx.modeling_latent_upsampler.PixelShuffleND +class PixelShuffleND(torch.nn.Module): + def __init__(self, dims, upscale_factors=(2, 2, 2)): + super().__init__() + + self.dims = dims + self.upscale_factors = upscale_factors + + if dims not in [1, 2, 3]: + raise ValueError("dims must be 1, 2, or 3") + + def forward(self, x): + if self.dims == 3: + # spatiotemporal: b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3) + return ( + x.unflatten(1, (-1, *self.upscale_factors[:3])) + .permute(0, 1, 5, 2, 6, 3, 7, 4) + .flatten(6, 7) + .flatten(4, 5) + .flatten(2, 3) + ) + elif self.dims == 2: + # spatial: b (c p1 p2) h w -> b c (h p1) (w p2) + return ( + x.unflatten(1, (-1, *self.upscale_factors[:2])).permute(0, 1, 4, 2, 5, 3).flatten(4, 5).flatten(2, 3) + ) + elif self.dims == 1: + # temporal: b (c p1) f h w -> b c (f p1) h w + return x.unflatten(1, (-1, *self.upscale_factors[:1])).permute(0, 1, 3, 2, 4, 5).flatten(2, 3) + + +class BlurDownsample(torch.nn.Module): + """ + Anti-aliased spatial downsampling by integer stride using a fixed separable binomial kernel. Applies only on H,W. + Works for dims=2 or dims=3 (per-frame). + """ + + def __init__(self, dims: int, stride: int, kernel_size: int = 5) -> None: + super().__init__() + + if dims not in (2, 3): + raise ValueError(f"`dims` must be either 2 or 3 but is {dims}") + if kernel_size < 3 or kernel_size % 2 != 1: + raise ValueError(f"`kernel_size` must be an odd number >= 3 but is {kernel_size}") + + self.dims = dims + self.stride = stride + self.kernel_size = kernel_size + + # 5x5 separable binomial kernel using binomial coefficients [1, 4, 6, 4, 1] from + # the 4th row of Pascal's triangle. This kernel is used for anti-aliasing and + # provides a smooth approximation of a Gaussian filter (often called a "binomial filter"). + # The 2D kernel is constructed as the outer product and normalized. + k = torch.tensor([math.comb(kernel_size - 1, k) for k in range(kernel_size)]) + k2d = k[:, None] @ k[None, :] + k2d = (k2d / k2d.sum()).float() # shape (kernel_size, kernel_size) + self.register_buffer("kernel", k2d[None, None, :, :]) # (1, 1, kernel_size, kernel_size) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if self.stride == 1: + return x + + if self.dims == 2: + c = x.shape[1] + weight = self.kernel.expand(c, 1, self.kernel_size, self.kernel_size) # depthwise + x = F.conv2d(x, weight=weight, bias=None, stride=self.stride, padding=self.kernel_size // 2, groups=c) + else: + # dims == 3: apply per-frame on H,W + b, c, f, _, _ = x.shape + x = x.transpose(1, 2).flatten(0, 1) # [B, C, F, H, W] --> [B * F, C, H, W] + + weight = self.kernel.expand(c, 1, self.kernel_size, self.kernel_size) # depthwise + x = F.conv2d(x, weight=weight, bias=None, stride=self.stride, padding=self.kernel_size // 2, groups=c) + + h2, w2 = x.shape[-2:] + x = x.unflatten(0, (b, f)).reshape(b, -1, f, h2, w2) # [B * F, C, H, W] --> [B, C, F, H, W] + return x + + +class SpatialRationalResampler(torch.nn.Module): + """ + Scales by the spatial size of the input by a rational number `scale`. For example, `scale = 0.75` will downsample + by a factor of 3 / 4, while `scale = 1.5` will upsample by a factor of 3 / 2. This works by first upsampling the + input by the (integer) numerator of `scale`, and then performing a blur + stride anti-aliased downsample by the + (integer) denominator. + """ + + def __init__(self, mid_channels: int = 1024, scale: float = 2.0): + super().__init__() + self.scale = float(scale) + num_denom = RATIONAL_RESAMPLER_SCALE_MAPPING.get(scale, None) + if num_denom is None: + raise ValueError( + f"The supplied `scale` {scale} is not supported; supported scales are {list(RATIONAL_RESAMPLER_SCALE_MAPPING.keys())}" + ) + self.num, self.den = num_denom + + self.conv = torch.nn.Conv2d(mid_channels, (self.num**2) * mid_channels, kernel_size=3, padding=1) + self.pixel_shuffle = PixelShuffleND(2, upscale_factors=(self.num, self.num)) + self.blur_down = BlurDownsample(dims=2, stride=self.den) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + # Expected x shape: [B * F, C, H, W] + # b, _, f, h, w = x.shape + # x = x.transpose(1, 2).flatten(0, 1) # [B, C, F, H, W] --> [B * F, C, H, W] + x = self.conv(x) + x = self.pixel_shuffle(x) + x = self.blur_down(x) + # x = x.unflatten(0, (b, f)).reshape(b, -1, f, h, w) # [B * F, C, H, W] --> [B, C, F, H, W] + return x + + +class LTX2LatentUpsamplerModel(ModelMixin, ConfigMixin): + """ + Model to spatially upsample VAE latents. + + Args: + in_channels (`int`, defaults to `128`): + Number of channels in the input latent + mid_channels (`int`, defaults to `512`): + Number of channels in the middle layers + num_blocks_per_stage (`int`, defaults to `4`): + Number of ResBlocks to use in each stage (pre/post upsampling) + dims (`int`, defaults to `3`): + Number of dimensions for convolutions (2 or 3) + spatial_upsample (`bool`, defaults to `True`): + Whether to spatially upsample the latent + temporal_upsample (`bool`, defaults to `False`): + Whether to temporally upsample the latent + """ + + @register_to_config + def __init__( + self, + in_channels: int = 128, + mid_channels: int = 1024, + num_blocks_per_stage: int = 4, + dims: int = 3, + spatial_upsample: bool = True, + temporal_upsample: bool = False, + rational_spatial_scale: float = 2.0, + use_rational_resampler: bool = True, + ): + super().__init__() + + self.in_channels = in_channels + self.mid_channels = mid_channels + self.num_blocks_per_stage = num_blocks_per_stage + self.dims = dims + self.spatial_upsample = spatial_upsample + self.temporal_upsample = temporal_upsample + + ConvNd = torch.nn.Conv2d if dims == 2 else torch.nn.Conv3d + + self.initial_conv = ConvNd(in_channels, mid_channels, kernel_size=3, padding=1) + self.initial_norm = torch.nn.GroupNorm(32, mid_channels) + self.initial_activation = torch.nn.SiLU() + + self.res_blocks = torch.nn.ModuleList([ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)]) + + if spatial_upsample and temporal_upsample: + self.upsampler = torch.nn.Sequential( + torch.nn.Conv3d(mid_channels, 8 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(3), + ) + elif spatial_upsample: + if use_rational_resampler: + self.upsampler = SpatialRationalResampler(mid_channels=mid_channels, scale=rational_spatial_scale) + else: + self.upsampler = torch.nn.Sequential( + torch.nn.Conv2d(mid_channels, 4 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(2), + ) + elif temporal_upsample: + self.upsampler = torch.nn.Sequential( + torch.nn.Conv3d(mid_channels, 2 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(1), + ) + else: + raise ValueError("Either spatial_upsample or temporal_upsample must be True") + + self.post_upsample_res_blocks = torch.nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + self.final_conv = ConvNd(mid_channels, in_channels, kernel_size=3, padding=1) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + batch_size, num_channels, num_frames, height, width = hidden_states.shape + + if self.dims == 2: + hidden_states = hidden_states.permute(0, 2, 1, 3, 4).flatten(0, 1) + hidden_states = self.initial_conv(hidden_states) + hidden_states = self.initial_norm(hidden_states) + hidden_states = self.initial_activation(hidden_states) + + for block in self.res_blocks: + hidden_states = block(hidden_states) + + hidden_states = self.upsampler(hidden_states) + + for block in self.post_upsample_res_blocks: + hidden_states = block(hidden_states) + + hidden_states = self.final_conv(hidden_states) + hidden_states = hidden_states.unflatten(0, (batch_size, -1)).permute(0, 2, 1, 3, 4) + else: + hidden_states = self.initial_conv(hidden_states) + hidden_states = self.initial_norm(hidden_states) + hidden_states = self.initial_activation(hidden_states) + + for block in self.res_blocks: + hidden_states = block(hidden_states) + + if self.temporal_upsample: + hidden_states = self.upsampler(hidden_states) + hidden_states = hidden_states[:, :, 1:, :, :] + else: + hidden_states = hidden_states.permute(0, 2, 1, 3, 4).flatten(0, 1) + hidden_states = self.upsampler(hidden_states) + hidden_states = hidden_states.unflatten(0, (batch_size, -1)).permute(0, 2, 1, 3, 4) + + for block in self.post_upsample_res_blocks: + hidden_states = block(hidden_states) + + hidden_states = self.final_conv(hidden_states) + + return hidden_states diff --git a/diffusers/pipelines/ltx2/pipeline_ltx2.py b/diffusers/pipelines/ltx2/pipeline_ltx2.py new file mode 100644 index 0000000000000000000000000000000000000000..493db96e48a794002472973be27f0903b6d13158 --- /dev/null +++ b/diffusers/pipelines/ltx2/pipeline_ltx2.py @@ -0,0 +1,1488 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Gemma3ForConditionalGeneration, Gemma3Processor, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import FromSingleFileMixin, LTX2LoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTX2Audio, AutoencoderKLLTX2Video +from ...models.transformers import LTX2VideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .connectors import LTX2TextConnectors +from .pipeline_output import LTX2PipelineOutput +from .vocoder import LTX2Vocoder, LTX2VocoderWithBWE + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTX2Pipeline + >>> from diffusers.utils import encode_video + + >>> pipe = LTX2Pipeline.from_pretrained("Lightricks/LTX-2", torch_dtype=torch.bfloat16) + >>> pipe.enable_model_cpu_offload() + + >>> prompt = "A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage" + >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted" + + >>> frame_rate = 24.0 + >>> video, audio = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... width=768, + ... height=512, + ... num_frames=121, + ... frame_rate=frame_rate, + ... num_inference_steps=40, + ... guidance_scale=4.0, + ... output_type="np", + ... return_dict=False, + ... ) + + >>> encode_video( + ... video[0], + ... fps=frame_rate, + ... audio=audio[0].float().cpu(), + ... audio_sample_rate=pipe.vocoder.config.output_sampling_rate, # should be 24000 + ... output_path="video.mp4", + ... ) + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTX2Pipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraLoaderMixin): + r""" + Pipeline for text-to-video generation. + + Reference: https://github.com/Lightricks/LTX-Video + + Args: + transformer ([`LTXVideoTransformer3DModel`]): + Conditional Transformer architecture to denoise the encoded video latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLLTXVideo`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + connectors ([`LTX2TextConnectors`]): + Text connector stack used to adapt text encoder hidden states for the video and audio branches. + """ + + model_cpu_offload_seq = "text_encoder->connectors->transformer->vae->audio_vae->vocoder" + _optional_components = ["processor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTX2Video, + audio_vae: AutoencoderKLLTX2Audio, + text_encoder: Gemma3ForConditionalGeneration, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + connectors: LTX2TextConnectors, + transformer: LTX2VideoTransformer3DModel, + vocoder: LTX2Vocoder | LTX2VocoderWithBWE, + processor: Gemma3Processor | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + audio_vae=audio_vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + connectors=connectors, + transformer=transformer, + vocoder=vocoder, + scheduler=scheduler, + processor=processor, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + # TODO: check whether the MEL compression ratio logic here is corrct + self.audio_vae_mel_compression_ratio = ( + self.audio_vae.mel_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.audio_vae_temporal_compression_ratio = ( + self.audio_vae.temporal_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.audio_sampling_rate = ( + self.audio_vae.config.sample_rate if getattr(self, "audio_vae", None) is not None else 16000 + ) + self.audio_hop_length = ( + self.audio_vae.config.mel_hop_length if getattr(self, "audio_vae", None) is not None else 160 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 1024 + ) + + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`str` or `torch.device`): + torch device to place the resulting embeddings on + dtype: (`torch.dtype`): + torch dtype to cast the prompt embeds to + max_sequence_length (`int`, defaults to 1024): Maximum sequence length to use for the prompt. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if getattr(self, "tokenizer", None) is not None: + # Gemma expects left padding for chat-style prompts + self.tokenizer.padding_side = "left" + if self.tokenizer.pad_token is None: + self.tokenizer.pad_token = self.tokenizer.eos_token + + prompt = [p.strip() for p in prompt] + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + text_input_ids = text_input_ids.to(device) + prompt_attention_mask = prompt_attention_mask.to(device) + + text_encoder_outputs = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ) + text_encoder_hidden_states = text_encoder_outputs.hidden_states + text_encoder_hidden_states = torch.stack(text_encoder_hidden_states, dim=-1) + prompt_embeds = text_encoder_hidden_states.flatten(2, 3).to(dtype=dtype) # Pack to 3D + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + @torch.no_grad() + def enhance_prompt( + self, + prompt: str, + system_prompt: str, + max_new_tokens: int = 512, + seed: int = 10, + generator: torch.Generator | None = None, + generation_kwargs: dict[str, Any] | None = None, + device: str | torch.device | None = None, + ): + """ + Enhances the supplied `prompt` by generating a new prompt using the current text encoder (default is a + `transformers.Gemma3ForConditionalGeneration` model) from it and a system prompt. + """ + device = device or self._execution_device + if generation_kwargs is None: + # Set to default generation kwargs + generation_kwargs = {"do_sample": True, "temperature": 0.7} + + messages = [ + {"role": "system", "content": system_prompt}, + {"role": "user", "content": f"user prompt: {prompt}"}, + ] + template = self.processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) + model_inputs = self.processor(text=template, images=None, return_tensors="pt").to(device) + self.text_encoder.to(device) + + # `transformers.GenerationMixin.generate` does not support using a `torch.Generator` to control randomness, + # so manually apply a seed for reproducible generation. + if generator is not None: + # Overwrite seed to generator's initial seed + seed = generator.initial_seed() + torch.manual_seed(seed) + generated_sequences = self.text_encoder.generate( + **model_inputs, + max_new_tokens=max_new_tokens, + **generation_kwargs, + ) # tensor of shape [batch_size, seq_len] + + generated_ids = [seq[len(model_inputs.input_ids[i]) :] for i, seq in enumerate(generated_sequences)] + enhanced_prompt = self.processor.tokenizer.batch_decode(generated_ids, skip_special_tokens=True) + return enhanced_prompt + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + spatio_temporal_guidance_blocks=None, + stg_scale=None, + audio_stg_scale=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + if ((stg_scale > 0.0) or (audio_stg_scale > 0.0)) and not spatio_temporal_guidance_blocks: + raise ValueError( + "Spatio-Temporal Guidance (STG) is specified but no STG blocks are supplied. Please supply a list of" + "block indices at which to apply STG in `spatio_temporal_guidance_blocks`" + ) + + @staticmethod + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_image2video.LTX2ImageToVideoPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + @staticmethod + def _normalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents - latents_mean) / latents_std + + @staticmethod + def _denormalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents * latents_std) + latents_mean + + @staticmethod + def _create_noised_state( + latents: torch.Tensor, noise_scale: float | torch.Tensor, generator: torch.Generator | None = None + ): + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + noised_latents = noise_scale * noise + (1 - noise_scale) * latents + return noised_latents + + @staticmethod + def _pack_audio_latents( + latents: torch.Tensor, patch_size: int | None = None, patch_size_t: int | None = None + ) -> torch.Tensor: + # Audio latents shape: [B, C, L, M], where L is the latent audio length and M is the number of mel bins + if patch_size is not None and patch_size_t is not None: + # Packs the latents into a patch sequence of shape [B, L // p_t * M // p, C * p_t * p] (a ndim=3 tnesor). + # dim=1 is the effective audio sequence length and dim=2 is the effective audio input feature size. + batch_size, num_channels, latent_length, latent_mel_bins = latents.shape + post_patch_latent_length = latent_length / patch_size_t + post_patch_mel_bins = latent_mel_bins / patch_size + latents = latents.reshape( + batch_size, -1, post_patch_latent_length, patch_size_t, post_patch_mel_bins, patch_size + ) + latents = latents.permute(0, 2, 4, 1, 3, 5).flatten(3, 5).flatten(1, 2) + else: + # Packs the latents into a patch sequence of shape [B, L, C * M]. This implicitly assumes a (mel) + # patch_size of M (all mel bins constitutes a single patch) and a patch_size_t of 1. + latents = latents.transpose(1, 2).flatten(2, 3) # [B, C, L, M] --> [B, L, C * M] + return latents + + @staticmethod + def _unpack_audio_latents( + latents: torch.Tensor, + latent_length: int, + num_mel_bins: int, + patch_size: int | None = None, + patch_size_t: int | None = None, + ) -> torch.Tensor: + # Unpacks an audio patch sequence of shape [B, S, D] into a latent spectrogram tensor of shape [B, C, L, M], + # where L is the latent audio length and M is the number of mel bins. + if patch_size is not None and patch_size_t is not None: + batch_size = latents.size(0) + latents = latents.reshape(batch_size, latent_length, num_mel_bins, -1, patch_size_t, patch_size) + latents = latents.permute(0, 3, 1, 4, 2, 5).flatten(4, 5).flatten(2, 3) + else: + # Assume [B, S, D] = [B, L, C * M], which implies that patch_size = M and patch_size_t = 1. + latents = latents.unflatten(2, (-1, num_mel_bins)).transpose(1, 2) + return latents + + def prepare_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 768, + num_frames: int = 121, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + if latents.ndim == 5: + latents = self._normalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + # latents are of shape [B, C, F, H, W], need to be packed + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + if latents.ndim != 3: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is [batch_size, num_seq, num_features]." + ) + latents = self._create_noised_state(latents, noise_scale, generator) + return latents.to(device=device, dtype=dtype) + + height = height // self.vae_spatial_compression_ratio + width = width // self.vae_spatial_compression_ratio + num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + + shape = (batch_size, num_channels_latents, num_frames, height, width) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + return latents + + def prepare_audio_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 8, + audio_latent_length: int = 1, # 1 is just a dummy value + num_mel_bins: int = 64, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + if latents.ndim == 4: + # latents are of shape [B, C, L, M], need to be packed + latents = self._pack_audio_latents(latents) + if latents.ndim != 3: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is [batch_size, num_seq, num_features]." + ) + latents = self._normalize_audio_latents(latents, self.audio_vae.latents_mean, self.audio_vae.latents_std) + latents = self._create_noised_state(latents, noise_scale, generator) + return latents.to(device=device, dtype=dtype) + + latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio + + shape = (batch_size, num_channels_latents, audio_latent_length, latent_mel_bins) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_audio_latents(latents) + return latents + + def convert_velocity_to_x0( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_x0 = sample - denoised_output * scheduler.sigmas[step_idx] + return sample_x0 + + def convert_x0_to_velocity( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_v = (sample - denoised_output) / scheduler.sigmas[step_idx] + return sample_v + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def stg_scale(self): + return self._stg_scale + + @property + def modality_scale(self): + return self._modality_scale + + @property + def audio_guidance_scale(self): + return self._audio_guidance_scale + + @property + def audio_guidance_rescale(self): + return self._audio_guidance_rescale + + @property + def audio_stg_scale(self): + return self._audio_stg_scale + + @property + def audio_modality_scale(self): + return self._audio_modality_scale + + @property + def do_classifier_free_guidance(self): + return (self._guidance_scale > 1.0) or (self._audio_guidance_scale > 1.0) + + @property + def do_spatio_temporal_guidance(self): + return (self._stg_scale > 0.0) or (self._audio_stg_scale > 0.0) + + @property + def do_modality_isolation_guidance(self): + return (self._modality_scale > 1.0) or (self._audio_modality_scale > 1.0) + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + num_inference_steps: int = 40, + sigmas: list[float] | None = None, + timesteps: list[int] = None, + guidance_scale: float = 4.0, + stg_scale: float = 0.0, + modality_scale: float = 1.0, + guidance_rescale: float = 0.0, + audio_guidance_scale: float | None = None, + audio_stg_scale: float | None = None, + audio_modality_scale: float | None = None, + audio_guidance_rescale: float | None = None, + spatio_temporal_guidance_blocks: list[int] | None = None, + noise_scale: float = 0.0, + num_videos_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + audio_latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + use_cross_timestep: bool = False, + system_prompt: str | None = None, + prompt_max_new_tokens: int = 512, + prompt_enhancement_kwargs: dict[str, Any] | None = None, + prompt_enhancement_seed: int = 10, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 1024, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + height (`int`, *optional*, defaults to `512`): + The height in pixels of the generated image. This is set to 480 by default for the best results. + width (`int`, *optional*, defaults to `768`): + The width in pixels of the generated image. This is set to 848 by default for the best results. + num_frames (`int`, *optional*, defaults to `121`): + The number of video frames to generate + frame_rate (`float`, *optional*, defaults to `24.0`): + The frames per second (FPS) of the generated video. + num_inference_steps (`int`, *optional*, defaults to 40): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to `4.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. Used for the video modality (there is + a separate value `audio_guidance_scale` for the audio modality). + stg_scale (`float`, *optional*, defaults to `0.0`): + Video guidance scale for Spatio-Temporal Guidance (STG), proposed in [Spatiotemporal Skip Guidance for + Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664). STG uses a CFG-like estimate + where we move the sample away from a weak sample from a perturbed version of the denoising model. + Enabling STG will result in an additional denoising model forward pass; the default value of `0.0` + means that STG is disabled. + modality_scale (`float`, *optional*, defaults to `1.0`): + Video guidance scale for LTX-2.X modality isolation guidance, where we move the sample away from a + weaker sample generated by the denoising model withy cross-modality (audio-to-video and video-to-audio) + cross attention disabled using a CFG-like estimate. Enabling modality guidance will result in an + additional denoising model forward pass; the default value of `1.0` means that modality guidance is + disabled. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. Used for the video modality. + audio_guidance_scale (`float`, *optional* defaults to `None`): + Audio guidance scale for CFG with respect to the negative prompt. The CFG update rule is the same for + video and audio, but they can use different values for the guidance scale. The LTX-2.X authors suggest + that the `audio_guidance_scale` should be higher relative to the video `guidance_scale` (e.g. for + LTX-2.3 they suggest 3.0 for video and 7.0 for audio). If `None`, defaults to the video value + `guidance_scale`. + audio_stg_scale (`float`, *optional*, defaults to `None`): + Audio guidance scale for STG. As with CFG, the STG update rule is otherwise the same for video and + audio. For LTX-2.3, a value of 1.0 is suggested for both video and audio. If `None`, defaults to the + video value `stg_scale`. + audio_modality_scale (`float`, *optional*, defaults to `None`): + Audio guidance scale for LTX-2.X modality isolation guidance. As with CFG, the modality guidance rule + is otherwise the same for video and audio. For LTX-2.3, a value of 3.0 is suggested for both video and + audio. If `None`, defaults to the video value `modality_scale`. + audio_guidance_rescale (`float`, *optional*, defaults to `None`): + A separate guidance rescale factor for the audio modality. If `None`, defaults to the video value + `guidance_rescale`. + spatio_temporal_guidance_blocks (`list[int]`, *optional*, defaults to `None`): + The zero-indexed transformer block indices at which to apply STG. Must be supplied if STG is used + (`stg_scale` or `audio_stg_scale` is greater than `0`). A value of `[29]` is recommended for LTX-2.0 + and `[28]` is recommended for LTX-2.3. + noise_scale (`float`, *optional*, defaults to `0.0`): + The interpolation factor between random noise and denoised latents at each timestep. Applying noise to + the `latents` and `audio_latents` before continue denoising. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + audio_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for audio + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + The interpolation factor between random noise and denoised latents at the decode timestep. + use_cross_timestep (`bool` *optional*, defaults to `False`): + Whether to use the cross modality (audio is the cross modality of video, and vice versa) sigma when + calculating the cross attention modulation parameters. `True` is the newer (e.g. LTX-2.3) behavior; + `False` is the legacy LTX-2.0 behavior. + system_prompt (`str`, *optional*, defaults to `None`): + Optional system prompt to use for prompt enhancement. The system prompt will be used by the current + text encoder (by default, a `Gemma3ForConditionalGeneration` model) to generate an enhanced prompt from + the original `prompt` to condition generation. If not supplied, prompt enhancement will not be + performed. + prompt_max_new_tokens (`int`, *optional*, defaults to `512`): + The maximum number of new tokens to generate when performing prompt enhancement. + prompt_enhancement_kwargs (`dict[str, Any]`, *optional*, defaults to `None`): + Keyword arguments for `self.text_encoder.generate`. If not supplied, default arguments of + `do_sample=True` and `temperature=0.7` will be used. See + https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationMixin.generate + for more details. + prompt_enhancement_seed (`int`, *optional*, default to `10`): + Random seed for any random operations during prompt enhancement. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTX2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*, defaults to `["latents"]`): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `1024`): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ltx.LTX2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTX2PipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + audio_guidance_scale = audio_guidance_scale or guidance_scale + audio_stg_scale = audio_stg_scale or stg_scale + audio_modality_scale = audio_modality_scale or modality_scale + audio_guidance_rescale = audio_guidance_rescale or guidance_rescale + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + stg_scale=stg_scale, + audio_stg_scale=audio_stg_scale, + ) + + # Per-modality guidance scales (video, audio) + self._guidance_scale = guidance_scale + self._stg_scale = stg_scale + self._modality_scale = modality_scale + self._guidance_rescale = guidance_rescale + self._audio_guidance_scale = audio_guidance_scale + self._audio_stg_scale = audio_stg_scale + self._audio_modality_scale = audio_modality_scale + self._audio_guidance_rescale = audio_guidance_rescale + + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + if system_prompt is not None and prompt is not None: + prompt = self.enhance_prompt( + prompt=prompt, + system_prompt=system_prompt, + max_new_tokens=prompt_max_new_tokens, + seed=prompt_enhancement_seed, + generator=generator, + generation_kwargs=prompt_enhancement_kwargs, + device=device, + ) + + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + tokenizer_padding_side = "left" # Padding side for default Gemma3-12B text encoder + if getattr(self, "tokenizer", None) is not None: + tokenizer_padding_side = getattr(self.tokenizer, "padding_side", "left") + connector_prompt_embeds, connector_audio_prompt_embeds, connector_attention_mask = self.connectors( + prompt_embeds, prompt_attention_mask, padding_side=tokenizer_padding_side + ) + + # 4. Prepare latent variables + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + if latents is not None: + if latents.ndim == 5: + logger.info( + "Got latents of shape [batch_size, latent_dim, latent_frames, latent_height, latent_width], `latent_num_frames`, `latent_height`, `latent_width` will be inferred." + ) + _, _, latent_num_frames, latent_height, latent_width = latents.shape # [B, C, F, H, W] + elif latents.ndim == 3: + logger.warning( + f"You have supplied packed `latents` of shape {latents.shape}, so the latent dims cannot be" + f" inferred. Make sure the supplied `height`, `width`, and `num_frames` are correct." + ) + else: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is either [batch_size, seq_len, num_features] or [batch_size, latent_dim, latent_frames, latent_height, latent_width]." + ) + # video_sequence_length = latent_num_frames * latent_height * latent_width + + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + noise_scale, + torch.float32, + device, + generator, + latents, + ) + + duration_s = num_frames / frame_rate + audio_latents_per_second = ( + self.audio_sampling_rate / self.audio_hop_length / float(self.audio_vae_temporal_compression_ratio) + ) + audio_num_frames = round(duration_s * audio_latents_per_second) + if audio_latents is not None: + if audio_latents.ndim == 4: + logger.info( + "Got audio_latents of shape [batch_size, num_channels, audio_length, mel_bins], `audio_num_frames` will be inferred." + ) + _, _, audio_num_frames, _ = audio_latents.shape # [B, C, L, M] + elif audio_latents.ndim == 3: + logger.warning( + f"You have supplied packed `audio_latents` of shape {audio_latents.shape}, so the latent dims" + f" cannot be inferred. Make sure the supplied `num_frames` and `frame_rate` are correct." + ) + else: + raise ValueError( + f"Provided `audio_latents` tensor has shape {audio_latents.shape}, but the expected shape is either [batch_size, seq_len, num_features] or [batch_size, num_channels, audio_length, mel_bins]." + ) + + num_mel_bins = self.audio_vae.config.mel_bins if getattr(self, "audio_vae", None) is not None else 64 + latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio + num_channels_latents_audio = ( + self.audio_vae.config.latent_channels if getattr(self, "audio_vae", None) is not None else 8 + ) + audio_latents = self.prepare_audio_latents( + batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents_audio, + audio_latent_length=audio_num_frames, + num_mel_bins=num_mel_bins, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=audio_latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + mu = calculate_shift( + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_image_seq_len", 1024), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.95), + self.scheduler.config.get("max_shift", 2.05), + ) + # For now, duplicate the scheduler for use with the audio latents + audio_scheduler = copy.deepcopy(self.scheduler) + _, _ = retrieve_timesteps( + audio_scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # Set begin index to skip nonzero().item() call in scheduler initialization, which triggers GPU sync + self.scheduler.set_begin_index(0) + audio_scheduler.set_begin_index(0) + + # 6. Prepare micro-conditions + # Pre-compute video and audio positional ids as they will be the same at each step of the denoising loop + video_coords = self.transformer.rope.prepare_video_coords( + latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate + ) + audio_coords = self.transformer.audio_rope.prepare_audio_coords( + audio_latents.shape[0], audio_num_frames, audio_latents.device + ) + # Duplicate the positional ids as well if using CFG + if self.do_classifier_free_guidance: + video_coords = video_coords.repeat((2,) + (1,) * (video_coords.ndim - 1)) # Repeat twice in batch dim + audio_coords = audio_coords.repeat((2,) + (1,) * (audio_coords.ndim - 1)) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + audio_latent_model_input = ( + torch.cat([audio_latents] * 2) if self.do_classifier_free_guidance else audio_latents + ) + audio_latent_model_input = audio_latent_model_input.to(prompt_embeds.dtype) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + with self.transformer.cache_context("cond_uncond"): + noise_pred_video, noise_pred_audio = self.transformer( + hidden_states=latent_model_input, + audio_hidden_states=audio_latent_model_input, + encoder_hidden_states=connector_prompt_embeds, + audio_encoder_hidden_states=connector_audio_prompt_embeds, + timestep=timestep, + sigma=timestep, # Used by LTX-2.3 + encoder_attention_mask=connector_attention_mask, + audio_encoder_attention_mask=connector_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_coords, + audio_coords=audio_coords, + isolate_modalities=False, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video = noise_pred_video.float() + noise_pred_audio = noise_pred_audio.float() + + if self.do_classifier_free_guidance: + noise_pred_video_uncond_text, noise_pred_video = noise_pred_video.chunk(2) + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_video_uncond_text = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_text, i, self.scheduler + ) + # Use delta formulation as it works more nicely with multiple guidance terms + video_cfg_delta = (self.guidance_scale - 1) * (noise_pred_video - noise_pred_video_uncond_text) + + noise_pred_audio_uncond_text, noise_pred_audio = noise_pred_audio.chunk(2) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + noise_pred_audio_uncond_text = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_text, i, audio_scheduler + ) + audio_cfg_delta = (self.audio_guidance_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_text + ) + + # Get positive values from merged CFG inputs in case we need to do other DiT forward passes + if self.do_spatio_temporal_guidance or self.do_modality_isolation_guidance: + if i == 0: + # Only split values that remain constant throughout the loop once + video_prompt_embeds = connector_prompt_embeds.chunk(2, dim=0)[1] + audio_prompt_embeds = connector_audio_prompt_embeds.chunk(2, dim=0)[1] + prompt_attn_mask = connector_attention_mask.chunk(2, dim=0)[1] + + video_pos_ids = video_coords.chunk(2, dim=0)[0] + audio_pos_ids = audio_coords.chunk(2, dim=0)[0] + + # Split values that vary each denoising loop iteration + timestep = timestep.chunk(2, dim=0)[0] + else: + video_cfg_delta = audio_cfg_delta = 0 + + video_prompt_embeds = connector_prompt_embeds + audio_prompt_embeds = connector_audio_prompt_embeds + prompt_attn_mask = connector_attention_mask + + video_pos_ids = video_coords + audio_pos_ids = audio_coords + + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + + if self.do_spatio_temporal_guidance: + with self.transformer.cache_context("uncond_stg"): + noise_pred_video_uncond_stg, noise_pred_audio_uncond_stg = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=timestep, + sigma=timestep, # Used by LTX-2.3 + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + isolate_modalities=False, + # Use STG at given blocks to perturb model + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_stg = noise_pred_video_uncond_stg.float() + noise_pred_audio_uncond_stg = noise_pred_audio_uncond_stg.float() + noise_pred_video_uncond_stg = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_stg, i, self.scheduler + ) + noise_pred_audio_uncond_stg = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_stg, i, audio_scheduler + ) + + video_stg_delta = self.stg_scale * (noise_pred_video - noise_pred_video_uncond_stg) + audio_stg_delta = self.audio_stg_scale * (noise_pred_audio - noise_pred_audio_uncond_stg) + else: + video_stg_delta = audio_stg_delta = 0 + + if self.do_modality_isolation_guidance: + with self.transformer.cache_context("uncond_modality"): + noise_pred_video_uncond_modality, noise_pred_audio_uncond_modality = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=timestep, + sigma=timestep, # Used by LTX-2.3 + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + # Turn off A2V and V2A cross attn to isolate video and audio modalities + isolate_modalities=True, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_modality = noise_pred_video_uncond_modality.float() + noise_pred_audio_uncond_modality = noise_pred_audio_uncond_modality.float() + noise_pred_video_uncond_modality = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_modality, i, self.scheduler + ) + noise_pred_audio_uncond_modality = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_modality, i, audio_scheduler + ) + + video_modality_delta = (self.modality_scale - 1) * ( + noise_pred_video - noise_pred_video_uncond_modality + ) + audio_modality_delta = (self.audio_modality_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_modality + ) + else: + video_modality_delta = audio_modality_delta = 0 + + # Now apply all guidance terms + noise_pred_video_g = noise_pred_video + video_cfg_delta + video_stg_delta + video_modality_delta + noise_pred_audio_g = noise_pred_audio + audio_cfg_delta + audio_stg_delta + audio_modality_delta + + # Apply LTX-2.X guidance rescaling + if self.guidance_rescale > 0: + noise_pred_video = rescale_noise_cfg( + noise_pred_video_g, noise_pred_video, guidance_rescale=self.guidance_rescale + ) + else: + noise_pred_video = noise_pred_video_g + + if self.audio_guidance_rescale > 0: + noise_pred_audio = rescale_noise_cfg( + noise_pred_audio_g, noise_pred_audio, guidance_rescale=self.audio_guidance_rescale + ) + else: + noise_pred_audio = noise_pred_audio_g + + # Convert back to velocity for scheduler + noise_pred_video = self.convert_x0_to_velocity(latents, noise_pred_video, i, self.scheduler) + noise_pred_audio = self.convert_x0_to_velocity(audio_latents, noise_pred_audio, i, audio_scheduler) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred_video, t, latents, return_dict=False)[0] + # NOTE: for now duplicate scheduler for audio latents in case self.scheduler sets internal state in + # the step method (such as _step_index) + audio_latents = audio_scheduler.step(noise_pred_audio, t, audio_latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + audio_latents = self._denormalize_audio_latents( + audio_latents, self.audio_vae.latents_mean, self.audio_vae.latents_std + ) + audio_latents = self._unpack_audio_latents(audio_latents, audio_num_frames, num_mel_bins=latent_mel_bins) + + if output_type == "latent": + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + video = latents + audio = audio_latents + else: + latents = latents.to(prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + + latents = latents.to(self.vae.dtype) + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + audio_latents = audio_latents.to(self.audio_vae.dtype) + generated_mel_spectrograms = self.audio_vae.decode(audio_latents, return_dict=False)[0] + audio = self.vocoder(generated_mel_spectrograms) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video, audio) + + return LTX2PipelineOutput(frames=video, audio=audio) diff --git a/diffusers/pipelines/ltx2/pipeline_ltx2_condition.py b/diffusers/pipelines/ltx2/pipeline_ltx2_condition.py new file mode 100644 index 0000000000000000000000000000000000000000..ecc89ae79a30baeb40e0e26a80da875a065d62c0 --- /dev/null +++ b/diffusers/pipelines/ltx2/pipeline_ltx2_condition.py @@ -0,0 +1,1844 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import inspect +import math +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import Gemma3ForConditionalGeneration, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import FromSingleFileMixin, LTX2LoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTX2Audio, AutoencoderKLLTX2Video +from ...models.transformers import LTX2VideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .connectors import LTX2TextConnectors +from .pipeline_output import LTX2PipelineOutput +from .vocoder import LTX2Vocoder, LTX2VocoderWithBWE + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTX2ConditionPipeline + >>> from diffusers.utils import encode_video + >>> from diffusers.pipelines.ltx2.pipeline_ltx2_condition import LTX2VideoCondition + >>> from diffusers.utils import load_image + + >>> pipe = LTX2ConditionPipeline.from_pretrained("Lightricks/LTX-2", torch_dtype=torch.bfloat16) + >>> pipe.enable_model_cpu_offload() + + >>> first_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png" + ... ) + >>> last_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png" + ... ) + >>> first_cond = LTX2VideoCondition(frames=first_image, index=0, strength=1.0) + >>> last_cond = LTX2VideoCondition(frames=last_image, index=-1, strength=1.0) + >>> conditions = [first_cond, last_cond] + >>> prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings." + >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted, static" + + >>> frame_rate = 24.0 + >>> video = pipe( + ... conditions=conditions, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... width=768, + ... height=512, + ... num_frames=121, + ... frame_rate=frame_rate, + ... num_inference_steps=40, + ... guidance_scale=4.0, + ... output_type="np", + ... return_dict=False, + ... ) + >>> video = (video * 255).round().astype("uint8") + >>> video = torch.from_numpy(video) + + >>> encode_video( + ... video[0], + ... fps=frame_rate, + ... audio=audio[0].float().cpu(), + ... audio_sample_rate=pipe.vocoder.config.output_sampling_rate, # should be 24000 + ... output_path="video.mp4", + ... ) + ``` +""" + + +@dataclass +class LTX2VideoCondition: + """ + Defines a single frame-conditioning item for LTX-2 Video - a single frame or a sequence of frames. + + Attributes: + frames (`PIL.Image.Image` or `List[PIL.Image.Image]` or `np.ndarray` or `torch.Tensor`): + The image (or video) to condition the video on. Accepts any type that can be handled by + VideoProcessor.preprocess_video. + index (`int`, defaults to `0`): + The index at which the image or video will conditionally affect the video generation. + strength (`float`, defaults to `1.0`): + The strength of the conditioning effect. A value of `1.0` means the conditioning effect is fully applied. + """ + + frames: PIL.Image.Image | list[PIL.Image.Image] | np.ndarray | torch.Tensor + index: int = 0 + strength: float = 1.0 + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTX2ConditionPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraLoaderMixin): + r""" + Pipeline for video generation which allows image conditions to be inserted at arbitary parts of the video. + + Reference: https://github.com/Lightricks/LTX-Video + + TODO + """ + + model_cpu_offload_seq = "text_encoder->connectors->transformer->vae->audio_vae->vocoder" + _optional_components = ["audio_scheduler"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTX2Video, + audio_vae: AutoencoderKLLTX2Audio, + text_encoder: Gemma3ForConditionalGeneration, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + connectors: LTX2TextConnectors, + transformer: LTX2VideoTransformer3DModel, + vocoder: LTX2Vocoder | LTX2VocoderWithBWE, + audio_scheduler: FlowMatchEulerDiscreteScheduler | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + audio_vae=audio_vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + connectors=connectors, + transformer=transformer, + vocoder=vocoder, + scheduler=scheduler, + audio_scheduler=audio_scheduler, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.audio_vae_mel_compression_ratio = ( + self.audio_vae.mel_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.audio_vae_temporal_compression_ratio = ( + self.audio_vae.temporal_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.audio_sampling_rate = ( + self.audio_vae.config.sample_rate if getattr(self, "audio_vae", None) is not None else 16000 + ) + self.audio_hop_length = ( + self.audio_vae.config.mel_hop_length if getattr(self, "audio_vae", None) is not None else 160 + ) + self.audio_mel_bins = self.audio_vae.config.mel_bins if getattr(self, "audio_vae", None) is not None else 64 + self.audio_latent_channels = ( + self.audio_vae.config.latent_channels if getattr(self, "audio_vae", None) is not None else 8 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio, resample="bilinear") + + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 1024 + ) + tokenizer_padding_side = "left" # Padding side for default Gemma3-12B text encoder + if getattr(self, "tokenizer", None) is not None: + tokenizer_padding_side = getattr(self.tokenizer, "padding_side", "left") + self.tokenizer_padding_side = tokenizer_padding_side + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`str` or `torch.device`): + torch device to place the resulting embeddings on + dtype: (`torch.dtype`): + torch dtype to cast the prompt embeds to + max_sequence_length (`int`, defaults to 1024): Maximum sequence length to use for the prompt. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if getattr(self, "tokenizer", None) is not None: + # Gemma expects left padding for chat-style prompts + self.tokenizer.padding_side = "left" + if self.tokenizer.pad_token is None: + self.tokenizer.pad_token = self.tokenizer.eos_token + + prompt = [p.strip() for p in prompt] + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + text_input_ids = text_input_ids.to(device) + prompt_attention_mask = prompt_attention_mask.to(device) + + text_encoder_outputs = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ) + text_encoder_hidden_states = text_encoder_outputs.hidden_states + text_encoder_hidden_states = torch.stack(text_encoder_hidden_states, dim=-1) + prompt_embeds = text_encoder_hidden_states.flatten(2, 3).to(dtype=dtype) # Pack to 3D + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + latents=None, + audio_latents=None, + spatio_temporal_guidance_blocks=None, + stg_scale=None, + audio_stg_scale=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + if latents is not None and latents.ndim != 5: + raise ValueError( + f"Only unpacked (5D) video latents of shape `[batch_size, latent_channels, latent_frames," + f" latent_height, latent_width] are supported, but got {latents.ndim} dims. If you have packed (3D)" + f" latents, please unpack them (e.g. using the `_unpack_latents` method)." + ) + if audio_latents is not None and audio_latents.ndim != 4: + raise ValueError( + f"Only unpacked (4D) audio latents of shape `[batch_size, num_channels, audio_length, mel_bins] are" + f" supported, but got {latents.ndim} dims. If you have packed (3D) latents, please unpack them (e.g." + f" using the `_unpack_audio_latents` method)." + ) + + if ((stg_scale > 0.0) or (audio_stg_scale > 0.0)) and not spatio_temporal_guidance_blocks: + raise ValueError( + "Spatio-Temporal Guidance (STG) is specified but no STG blocks are supplied. Please supply a list of" + "block indices at which to apply STG in `spatio_temporal_guidance_blocks`" + ) + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_latents + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_image2video.LTX2ImageToVideoPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._create_noised_state + def _create_noised_state( + latents: torch.Tensor, noise_scale: float | torch.Tensor, generator: torch.Generator | None = None + ): + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + noised_latents = noise_scale * noise + (1 - noise_scale) * latents + return noised_latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_audio_latents + def _pack_audio_latents( + latents: torch.Tensor, patch_size: int | None = None, patch_size_t: int | None = None + ) -> torch.Tensor: + # Audio latents shape: [B, C, L, M], where L is the latent audio length and M is the number of mel bins + if patch_size is not None and patch_size_t is not None: + # Packs the latents into a patch sequence of shape [B, L // p_t * M // p, C * p_t * p] (a ndim=3 tnesor). + # dim=1 is the effective audio sequence length and dim=2 is the effective audio input feature size. + batch_size, num_channels, latent_length, latent_mel_bins = latents.shape + post_patch_latent_length = latent_length / patch_size_t + post_patch_mel_bins = latent_mel_bins / patch_size + latents = latents.reshape( + batch_size, -1, post_patch_latent_length, patch_size_t, post_patch_mel_bins, patch_size + ) + latents = latents.permute(0, 2, 4, 1, 3, 5).flatten(3, 5).flatten(1, 2) + else: + # Packs the latents into a patch sequence of shape [B, L, C * M]. This implicitly assumes a (mel) + # patch_size of M (all mel bins constitutes a single patch) and a patch_size_t of 1. + latents = latents.transpose(1, 2).flatten(2, 3) # [B, C, L, M] --> [B, L, C * M] + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_audio_latents + def _unpack_audio_latents( + latents: torch.Tensor, + latent_length: int, + num_mel_bins: int, + patch_size: int | None = None, + patch_size_t: int | None = None, + ) -> torch.Tensor: + # Unpacks an audio patch sequence of shape [B, S, D] into a latent spectrogram tensor of shape [B, C, L, M], + # where L is the latent audio length and M is the number of mel bins. + if patch_size is not None and patch_size_t is not None: + batch_size = latents.size(0) + latents = latents.reshape(batch_size, latent_length, num_mel_bins, -1, patch_size_t, patch_size) + latents = latents.permute(0, 3, 1, 4, 2, 5).flatten(4, 5).flatten(2, 3) + else: + # Assume [B, S, D] = [B, L, C * M], which implies that patch_size = M and patch_size_t = 1. + latents = latents.unflatten(2, (-1, num_mel_bins)).transpose(1, 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._normalize_audio_latents + def _normalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents - latents_mean) / latents_std + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_audio_latents + def _denormalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents * latents_std) + latents_mean + + # Copied from diffusers.pipelines.ltx.pipeline_ltx_condition.LTXConditionPipeline.trim_conditioning_sequence + def trim_conditioning_sequence(self, start_frame: int, sequence_num_frames: int, target_num_frames: int) -> int: + """ + Trim a conditioning sequence to the allowed number of frames. + + Args: + start_frame (int): The target frame number of the first frame in the sequence. + sequence_num_frames (int): The number of frames in the sequence. + target_num_frames (int): The target number of frames in the generated video. + Returns: + int: updated sequence length + """ + scale_factor = self.vae_temporal_compression_ratio + num_frames = min(sequence_num_frames, target_num_frames - start_frame) + # Trim down to a multiple of temporal_scale_factor frames plus 1 + num_frames = (num_frames - 1) // scale_factor * scale_factor + 1 + return num_frames + + def preprocess_conditions( + self, + conditions: LTX2VideoCondition | list[LTX2VideoCondition] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + device: torch.device | None = None, + ) -> tuple[list[torch.Tensor], list[float], list[int], list[int]]: + """ + Preprocesses the condition images/videos to torch tensors. + + Args: + conditions (`LTX2VideoCondition` or `List[LTX2VideoCondition]`, *optional*, defaults to `None`): + A list of image/video condition instances. + height (`int`, *optional*, defaults to `512`): + The desired height in pixels. + width (`int`, *optional*, defaults to `768`): + The desired width in pixels. + num_frames (`int`, *optional*, defaults to `121`): + The desired number of frames in the generated video. + device (`torch.device`, *optional*, defaults to `None`): + The device on which to put the preprocessed image/video tensors. + + Returns: + `Tuple[List[torch.Tensor], List[float], List[int], List[int]]`: + Returns a 4-tuple of lists of length `len(conditions)` as follows: + 1. The first list is a list of preprocessed video tensors of shape [batch_size=1, num_channels, + num_frames, height, width]. + 2. The second list is a list of conditioning strengths. + 3. The third list is a list of latent-space indices for each condition. + 4. The fourth list is a list of (trimmed) pixel-space frame counts per condition. This is needed + for keyframe coord semantics (single-pixel-frame keyframes have a clamped temporal extent). + """ + conditioning_frames, conditioning_strengths, conditioning_indices, conditioning_pixel_frames = [], [], [], [] + + if conditions is None: + conditions = [] + if isinstance(conditions, LTX2VideoCondition): + conditions = [conditions] + + frame_scale_factor = self.vae_temporal_compression_ratio + latent_num_frames = (num_frames - 1) // frame_scale_factor + 1 + for i, condition in enumerate(conditions): + # Create a channels-last video-like array of shape (F, H, W, C) in preparation for resizing. + if isinstance(condition.frames, PIL.Image.Image): + arr = np.array(condition.frames.convert("RGB"))[None] # (1, H, W, 3) + elif isinstance(condition.frames, list) and all(isinstance(f, PIL.Image.Image) for f in condition.frames): + arr = np.stack([np.array(f.convert("RGB")) for f in condition.frames]) # (F, H, W, 3) + elif isinstance(condition.frames, np.ndarray): + arr = condition.frames if condition.frames.ndim == 4 else condition.frames[None] + elif isinstance(condition.frames, torch.Tensor): + t = condition.frames if condition.frames.ndim == 4 else condition.frames.unsqueeze(0) + # Reference layout for video tensors is (F, C, H, W); convert to (F, H, W, C) for the + # resize logic, which expects channels-last. + arr = t.detach().cpu().permute(0, 2, 3, 1).numpy() + else: + raise TypeError(f"Unsupported `frames` type for condition {i}: {type(condition.frames)}") + + src_h, src_w = arr.shape[1], arr.shape[2] + num_cond_frames = arr.shape[0] + # Convert the NumPy array to a channels-first tensor of shape (1, C, F, H, W) + pixels = torch.from_numpy(np.ascontiguousarray(arr)).to(torch.float32) + pixels = pixels.permute(3, 0, 1, 2).unsqueeze(0).to(device) # (1, C, F, H, W) + + # Resize so the longer side fills the target, then center-crop to exact (height, width). + scale = max(height / src_h, width / src_w) + new_h = math.ceil(src_h * scale) + new_w = math.ceil(src_w * scale) + # Flatten (B, C, F, H, W) → (B*F, C, H, W) for the per-frame interpolation + pixels = pixels.permute(0, 2, 1, 3, 4).reshape(num_cond_frames, 3, src_h, src_w) + # NOTE: we avoid using VideoProcessor.preprocess_video here because it uses PIL.Image.resize under the + # hood, which will apply an anti-aliasing pre-filter when downsampling. The original LTX-2.X code simply + # uses F.interpolate, which is reproduced here. + pixels = torch.nn.functional.interpolate(pixels, size=(new_h, new_w), mode="bilinear", align_corners=False) + top = (new_h - height) // 2 + left = (new_w - width) // 2 + pixels = pixels[:, :, top : top + height, left : left + width] + pixels = pixels.reshape(1, num_cond_frames, 3, height, width).permute(0, 2, 1, 3, 4) + + # Map [0, 255] → [-1, 1] (VAE input convention). + condition_pixels = pixels / 127.5 - 1.0 + + # Interpret the index as a latent index, following the original LTX-2 code. + latent_start_idx = condition.index + # Support negative latent indices (e.g. -1 for the last latent index) + if latent_start_idx < 0: + # latent_start_idx will be positive because latent_num_frames is positive + latent_start_idx = latent_start_idx % latent_num_frames + if latent_start_idx >= latent_num_frames: + logger.warning( + f"The starting latent index {latent_start_idx} of condition {i} is too big for the specified number" + f" of latent frames {latent_num_frames}. This condition will be skipped." + ) + continue + + cond_num_frames = condition_pixels.size(2) + start_idx = max((latent_start_idx - 1) * frame_scale_factor + 1, 0) + truncated_cond_frames = self.trim_conditioning_sequence(start_idx, cond_num_frames, num_frames) + condition_pixels = condition_pixels[:, :, :truncated_cond_frames] + + conditioning_frames.append(condition_pixels.to(dtype=self.vae.dtype, device=device)) + conditioning_strengths.append(condition.strength) + conditioning_indices.append(latent_start_idx) + conditioning_pixel_frames.append(truncated_cond_frames) + + return conditioning_frames, conditioning_strengths, conditioning_indices, conditioning_pixel_frames + + def apply_first_frame_conditioning( + self, + latents: torch.Tensor, + conditioning_mask: torch.Tensor, + condition_latents: list[torch.Tensor], + condition_strengths: list[float], + condition_indices: list[int], + latent_height: int, + latent_width: int, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Apply first-frame visual conditioning by overwriting tokens at the first-frame positions. + + Only conditions with `latent_idx == 0` are applied here (matching `VideoConditionByLatentIndex` in the + reference implementation). Conditions at non-zero latent indices are appended as separate keyframe tokens via + `prepare_keyframe_extras` (matching `VideoConditionByKeyframeIndex`) and are skipped here. + + Args: + latents (`torch.Tensor`): + Initial packed (patchified) latents of shape [batch_size, patch_seq_len, hidden_dim]. + conditioning_mask (`torch.Tensor`): + Initial packed (patchified) conditioning mask of shape [batch_size, patch_seq_len, 1] with values in + [0, 1] where 0 means the denoising model output will be fully used and 1 means the condition will be + fully used. + + Returns: + `Tuple[torch.Tensor, torch.Tensor, torch.Tensor]`: + Returns a 3-tuple of tensors where: + 1. The packed video latents with first-frame conditions applied. + 2. The packed conditioning mask with first-frame strengths applied. + 3. The clean conditioning latents at first-frame positions (zeros elsewhere). + """ + clean_latents = torch.zeros_like(latents) + for cond, strength, latent_idx in zip(condition_latents, condition_strengths, condition_indices): + if latent_idx != 0: + # Non-first-frame conditions are handled as keyframe extras (appended tokens) instead. + continue + num_cond_tokens = cond.size(1) + start_token_idx = latent_idx * latent_height * latent_width + end_token_idx = start_token_idx + num_cond_tokens + + latents[:, start_token_idx:end_token_idx] = cond + conditioning_mask[:, start_token_idx:end_token_idx] = strength + clean_latents[:, start_token_idx:end_token_idx] = cond + + return latents, conditioning_mask, clean_latents + + def _prepare_keyframe_coords( + self, + keyframe_latent_num_frames: int, + keyframe_latent_height: int, + keyframe_latent_width: int, + pixel_frame_idx: int, + num_pixel_frames: int, + fps: float, + device: torch.device, + ) -> torch.Tensor: + """ + Compute positional coordinates for a keyframe condition being appended as extra tokens. + + Mirrors `VideoConditionByKeyframeIndex.apply_to` in the reference implementation: + - Latent coords scaled to pixel space *without* the causal fix (since non-zero-index keyframes don't need the + first-frame causal adjustment). + - Temporal axis offset by `pixel_frame_idx` (the pixel-space index at which the keyframe appears). + - For single-pixel-frame keyframes, the per-patch temporal extent is clamped to `[idx, idx + 1)` so the + keyframe occupies a single pixel timestep rather than the VAE-scaled range. + - Temporal coords divided by `fps` to produce seconds. + """ + patch_size = self.transformer_spatial_patch_size + patch_size_t = self.transformer_temporal_patch_size + scale_factors = ( + self.vae_temporal_compression_ratio, + self.vae_spatial_compression_ratio, + self.vae_spatial_compression_ratio, + ) + + grid_f = torch.arange( + start=0, end=keyframe_latent_num_frames, step=patch_size_t, dtype=torch.float32, device=device + ) + grid_h = torch.arange(start=0, end=keyframe_latent_height, step=patch_size, dtype=torch.float32, device=device) + grid_w = torch.arange(start=0, end=keyframe_latent_width, step=patch_size, dtype=torch.float32, device=device) + grid = torch.meshgrid(grid_f, grid_h, grid_w, indexing="ij") + grid = torch.stack(grid, dim=0) + + patch_size_delta = torch.tensor((patch_size_t, patch_size, patch_size), dtype=grid.dtype, device=device) + patch_ends = grid + patch_size_delta.view(3, 1, 1, 1) + + latent_coords = torch.stack([grid, patch_ends], dim=-1) # [3, N_F, N_H, N_W, 2] + latent_coords = latent_coords.flatten(1, 3) # [3, num_patches, 2] + latent_coords = latent_coords.unsqueeze(0) # [1, 3, num_patches, 2] + + scale_tensor = torch.tensor(scale_factors, device=device, dtype=latent_coords.dtype) + broadcast_shape = [1] * latent_coords.ndim + broadcast_shape[1] = -1 + pixel_coords = latent_coords * scale_tensor.view(*broadcast_shape) + + # No causal fix: keyframe coords place the keyframe at `pixel_frame_idx` without the first-frame adjustment. + pixel_coords[:, 0, :, :] = pixel_coords[:, 0, :, :] + pixel_frame_idx + + if num_pixel_frames == 1: + # Single-pixel-frame keyframe: clamp temporal extent to [idx, idx + 1). + pixel_coords[:, 0, :, 1:] = pixel_coords[:, 0, :, :1] + 1 + + pixel_coords[:, 0, :, :] = pixel_coords[:, 0, :, :] / fps + + return pixel_coords + + def prepare_latents( + self, + conditions: LTX2VideoCondition | list[LTX2VideoCondition] | None = None, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + noise_scale: float = 1.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None]: + """ + Prepare noisy video latents, applying frame conditions. + + First-frame conditions (`latent_idx == 0`) are applied by overwriting tokens at the first-frame positions + (`VideoConditionByLatentIndex` semantics). Non-first-frame conditions (`latent_idx > 0`) are concatenated onto + the main latent sequence with per-token `conditioning_mask = strength` (`VideoConditionByKeyframeIndex` + semantics) — the denoising loop's existing timestep formula `t * (1 - conditioning_mask)` and post-process + blend `denoised * (1 - conditioning_mask) + clean * conditioning_mask` then drive them across steps. + + Returns a 4-tuple: + - `latents`: packed noisy latents (base tokens + any keyframe tokens cat'd onto the sequence dim). + - `conditioning_mask`: packed conditioning mask with values in `[0, 1]` — `1` at first-frame positions, + `strength` at keyframe positions, `0` elsewhere. + - `clean_latents`: clean condition values at conditioned positions (zeros elsewhere); same shape as + `latents`. + - `keyframe_coords`: `[B, 3, num_keyframe_patches, 2]` positional coordinates to append to `video_coords`, + or `None` if there are no non-first-frame conditions. + """ + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + + shape = (batch_size, num_channels_latents, latent_num_frames, latent_height, latent_width) + mask_shape = (batch_size, 1, latent_num_frames, latent_height, latent_width) + + if latents is not None: + # Latents are expected to be unpacked (5D) with shape [B, F, C, H, W] + latents = self._normalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + else: + # NOTE: we set the initial latents to zeros rather a sample from the standard Gaussian prior because we + # will sample from the prior later once we have calculated the conditioning mask + latents = torch.zeros(shape, device=device, dtype=dtype) + + conditioning_mask = latents.new_zeros(mask_shape) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + conditioning_mask = self._pack_latents( + conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) # [B, seq_len, 1] + + if latents.ndim != 3 or latents.shape[:2] != conditioning_mask.shape[:2]: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is {conditioning_mask.shape[:2] + (num_channels_latents,)}." + ) + + if isinstance(generator, list): + logger.warning( + f"{self.__class__.__name__} does not support using a list of generators. The first generator in the" + f" list will be used for all (pseudo-)random operations." + ) + + condition_frames, condition_strengths, condition_indices, condition_pixel_frames = self.preprocess_conditions( + conditions, height, width, num_frames, device=device + ) + # Encode each condition through the VAE. We keep both the 5D latent (for coord computation) and the packed + # 3D latent (for first-frame replacement or keyframe append). + condition_latents_5d = [] + condition_latents_packed = [] + for condition_tensor in condition_frames: + condition_latent_5d = retrieve_latents( + self.vae.encode(condition_tensor), + generator=generator[0] if isinstance(generator, list) else generator, + sample_mode="argmax", + ) + condition_latent_5d = self._normalize_latents( + condition_latent_5d, self.vae.latents_mean, self.vae.latents_std + ).to(device=device, dtype=dtype) + condition_latent_packed = self._pack_latents( + condition_latent_5d, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + condition_latents_5d.append(condition_latent_5d) + condition_latents_packed.append(condition_latent_packed) + + # First-frame conditions (latent_idx == 0): replace tokens at the first-frame positions. + # NOTE: following the I2V pipeline, we return a conditioning mask. The original LTX 2 code uses a denoising + # mask, which is the inverse of the conditioning mask (`denoise_mask = 1 - conditioning_mask`). + latents, conditioning_mask, clean_latents = self.apply_first_frame_conditioning( + latents, + conditioning_mask, + condition_latents_packed, + condition_strengths, + condition_indices, + latent_height=latent_height, + latent_width=latent_width, + ) + + # Non-first-frame ("keyframe") conditions (latent_idx > 0): append as extra latent tokens to the noisy latent. + # Each condition gets a all-`strength` conditioning mask and pos ids, which are also appended to those of the + # noisy latent. At each denoising step i, the keyframe conditions get an effective noise level of + # (1 - conditioning_strength) * sigma_i. + frame_scale_factor = self.vae_temporal_compression_ratio + kf_tokens_list, kf_coords_list, kf_mask_list, kf_clean_list = [], [], [], [] + for cond_5d, cond_packed, strength, latent_idx, num_pixel_frames in zip( + condition_latents_5d, + condition_latents_packed, + condition_strengths, + condition_indices, + condition_pixel_frames, + ): + if latent_idx == 0: + continue + + _, _, kf_latent_frames, kf_latent_height, kf_latent_width = cond_5d.shape + pixel_frame_idx = (latent_idx - 1) * frame_scale_factor + 1 + + coords = self._prepare_keyframe_coords( + keyframe_latent_num_frames=kf_latent_frames, + keyframe_latent_height=kf_latent_height, + keyframe_latent_width=kf_latent_width, + pixel_frame_idx=pixel_frame_idx, + num_pixel_frames=num_pixel_frames, + fps=frame_rate, + device=device, + ) + + num_tokens = cond_packed.shape[1] + kf_mask = torch.full( + (cond_packed.shape[0], num_tokens, 1), + float(strength), + device=device, + dtype=conditioning_mask.dtype, + ) + + kf_tokens_list.append(cond_packed) + kf_clean_list.append(cond_packed) + kf_mask_list.append(kf_mask) + kf_coords_list.append(coords) + + if kf_tokens_list: + keyframe_coords = torch.cat(kf_coords_list, dim=2) + latents = torch.cat([latents, torch.cat(kf_tokens_list, dim=1)], dim=1) + conditioning_mask = torch.cat([conditioning_mask, torch.cat(kf_mask_list, dim=1)], dim=1) + clean_latents = torch.cat([clean_latents, torch.cat(kf_clean_list, dim=1)], dim=1) + else: + keyframe_coords = None + + # The conditioning_mask values have the following semantics: + # - mask=0: fully noise tokens (e.g. noisy latents) + # - mask=1: keep fully clean (e.g. I2V first-frame condition, conditions with strength=1) + # - mask in (0, 1): use intermediate noise level mask * sigma_i (noise_scale == sigma_0) + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + scaled_mask = (1.0 - conditioning_mask) * noise_scale # noise to initial noise level `noise_scale` + latents = noise * scaled_mask + latents * (1 - scaled_mask) + + return latents, conditioning_mask, clean_latents, keyframe_coords + + def prepare_audio_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 8, + audio_latent_length: int = 1, # 1 is just a dummy value + num_mel_bins: int = 64, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + # latents expected to be unpacked (4D) with shape [B, C, L, M] + latents = self._pack_audio_latents(latents) + latents = self._normalize_audio_latents(latents, self.audio_vae.latents_mean, self.audio_vae.latents_std) + latents = self._create_noised_state(latents, noise_scale, generator) + return latents.to(device=device, dtype=dtype) + + latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Sample in packed shape (B, L, C * M), following the original LTX-2.X code + packed_shape = (batch_size, audio_latent_length, num_channels_latents * latent_mel_bins) + latents = randn_tensor(packed_shape, generator=generator, device=device, dtype=dtype) + return latents + + def convert_velocity_to_x0( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_x0 = sample - denoised_output * scheduler.sigmas[step_idx] + return sample_x0 + + def convert_x0_to_velocity( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_v = (sample - denoised_output) / scheduler.sigmas[step_idx] + return sample_v + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def stg_scale(self): + return self._stg_scale + + @property + def modality_scale(self): + return self._modality_scale + + @property + def audio_guidance_scale(self): + return self._audio_guidance_scale + + @property + def audio_guidance_rescale(self): + return self._audio_guidance_rescale + + @property + def audio_stg_scale(self): + return self._audio_stg_scale + + @property + def audio_modality_scale(self): + return self._audio_modality_scale + + @property + def do_classifier_free_guidance(self): + return (self._guidance_scale > 1.0) or (self._audio_guidance_scale > 1.0) + + @property + def do_spatio_temporal_guidance(self): + return (self._stg_scale > 0.0) or (self._audio_stg_scale > 0.0) + + @property + def do_modality_isolation_guidance(self): + return (self._modality_scale > 1.0) or (self._audio_modality_scale > 1.0) + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + conditions: LTX2VideoCondition | list[LTX2VideoCondition] | None = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + num_inference_steps: int = 40, + sigmas: list[float] | None = None, + timesteps: list[float] | None = None, + guidance_scale: float = 4.0, + stg_scale: float = 0.0, + modality_scale: float = 1.0, + guidance_rescale: float = 0.0, + audio_guidance_scale: float | None = None, + audio_stg_scale: float | None = None, + audio_modality_scale: float | None = None, + audio_guidance_rescale: float | None = None, + spatio_temporal_guidance_blocks: list[int] | None = None, + noise_scale: float | None = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + audio_latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + use_cross_timestep: bool = False, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 1024, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + conditions (`List[LTXVideoCondition], *optional*`): + The list of frame-conditioning items for the video generation. + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + height (`int`, *optional*, defaults to `512`): + The height in pixels of the generated image. This is set to 480 by default for the best results. + width (`int`, *optional*, defaults to `768`): + The width in pixels of the generated image. This is set to 848 by default for the best results. + num_frames (`int`, *optional*, defaults to `121`): + The number of video frames to generate + frame_rate (`float`, *optional*, defaults to `24.0`): + The frames per second (FPS) of the generated video. + num_inference_steps (`int`, *optional*, defaults to 40): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to `4.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. Used for the video modality (there is + a separate value `audio_guidance_scale` for the audio modality). + stg_scale (`float`, *optional*, defaults to `0.0`): + Video guidance scale for Spatio-Temporal Guidance (STG), proposed in [Spatiotemporal Skip Guidance for + Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664). STG uses a CFG-like estimate + where we move the sample away from a weak sample from a perturbed version of the denoising model. + Enabling STG will result in an additional denoising model forward pass; the default value of `0.0` + means that STG is disabled. + modality_scale (`float`, *optional*, defaults to `1.0`): + Video guidance scale for LTX-2.X modality isolation guidance, where we move the sample away from a + weaker sample generated by the denoising model withy cross-modality (audio-to-video and video-to-audio) + cross attention disabled using a CFG-like estimate. Enabling modality guidance will result in an + additional denoising model forward pass; the default value of `1.0` means that modality guidance is + disabled. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. Used for the video modality. + audio_guidance_scale (`float`, *optional* defaults to `None`): + Audio guidance scale for CFG with respect to the negative prompt. The CFG update rule is the same for + video and audio, but they can use different values for the guidance scale. The LTX-2.X authors suggest + that the `audio_guidance_scale` should be higher relative to the video `guidance_scale` (e.g. for + LTX-2.3 they suggest 3.0 for video and 7.0 for audio). If `None`, defaults to the video value + `guidance_scale`. + audio_stg_scale (`float`, *optional*, defaults to `None`): + Audio guidance scale for STG. As with CFG, the STG update rule is otherwise the same for video and + audio. For LTX-2.3, a value of 1.0 is suggested for both video and audio. If `None`, defaults to the + video value `stg_scale`. + audio_modality_scale (`float`, *optional*, defaults to `None`): + Audio guidance scale for LTX-2.X modality isolation guidance. As with CFG, the modality guidance rule + is otherwise the same for video and audio. For LTX-2.3, a value of 3.0 is suggested for both video and + audio. If `None`, defaults to the video value `modality_scale`. + audio_guidance_rescale (`float`, *optional*, defaults to `None`): + A separate guidance rescale factor for the audio modality. If `None`, defaults to the video value + `guidance_rescale`. + spatio_temporal_guidance_blocks (`list[int]`, *optional*, defaults to `None`): + The zero-indexed transformer block indices at which to apply STG. Must be supplied if STG is used + (`stg_scale` or `audio_stg_scale` is greater than `0`). A value of `[29]` is recommended for LTX-2.0 + and `[28]` is recommended for LTX-2.3. + noise_scale (`float`, *optional*, defaults to `None`): + The interpolation factor between random noise and denoised latents at each timestep. Applying noise to + the `latents` and `audio_latents` before continue denoising. If not set, will be inferred from the + sigma schedule. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + audio_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for audio + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + The interpolation factor between random noise and denoised latents at the decode timestep. + use_cross_timestep (`bool` *optional*, defaults to `False`): + Whether to use the cross modality (audio is the cross modality of video, and vice versa) sigma when + calculating the cross attention modulation parameters. `True` is the newer (e.g. LTX-2.3) behavior; + `False` is the legacy LTX-2.0 behavior. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTX2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `1024`): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ltx.LTX2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTX2PipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + audio_guidance_scale = audio_guidance_scale or guidance_scale + audio_stg_scale = audio_stg_scale or stg_scale + audio_modality_scale = audio_modality_scale or modality_scale + audio_guidance_rescale = audio_guidance_rescale or guidance_rescale + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + latents=latents, + audio_latents=audio_latents, + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + stg_scale=stg_scale, + audio_stg_scale=audio_stg_scale, + ) + + # Per-modality guidance scales (video, audio) + self._guidance_scale = guidance_scale + self._stg_scale = stg_scale + self._modality_scale = modality_scale + self._guidance_rescale = guidance_rescale + self._audio_guidance_scale = audio_guidance_scale + self._audio_stg_scale = audio_stg_scale + self._audio_modality_scale = audio_modality_scale + self._audio_guidance_rescale = audio_guidance_rescale + + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if conditions is not None and not isinstance(conditions, list): + conditions = [conditions] + + # Infer noise scale: first (largest) sigma value if using custom sigmas, else 1.0 + if noise_scale is None: + noise_scale = sigmas[0] if sigmas is not None else 1.0 + + device = self._execution_device + + # 3. Prepare text embeddings + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + connector_prompt_embeds, connector_audio_prompt_embeds, connector_attention_mask = self.connectors( + prompt_embeds, prompt_attention_mask, padding_side=self.tokenizer_padding_side + ) + + # 4. Prepare latent variables + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + if latents is not None: + logger.info( + "Got latents of shape [batch_size, latent_dim, latent_frames, latent_height, latent_width], `latent_num_frames`, `latent_height`, `latent_width` will be inferred." + ) + _, _, latent_num_frames, latent_height, latent_width = latents.shape # [B, C, F, H, W] + # video_sequence_length = latent_num_frames * latent_height * latent_width + + num_channels_latents = self.transformer.config.in_channels + latents, conditioning_mask, clean_latents, keyframe_coords = self.prepare_latents( + conditions=conditions, + batch_size=batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + num_frames=num_frames, + frame_rate=frame_rate, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=latents, + ) + if self.do_classifier_free_guidance: + conditioning_mask = torch.cat([conditioning_mask, conditioning_mask]) + + duration_s = num_frames / frame_rate + audio_latents_per_second = ( + self.audio_sampling_rate / self.audio_hop_length / float(self.audio_vae_temporal_compression_ratio) + ) + audio_num_frames = round(duration_s * audio_latents_per_second) + if audio_latents is not None: + logger.info( + "Got audio_latents of shape [batch_size, num_channels, audio_num_frames, mel_bins], `audio_num_frames` will be inferred." + ) + _, _, audio_num_frames, _ = audio_latents.shape # [B, C, L, M] + + latent_mel_bins = self.audio_mel_bins // self.audio_vae_mel_compression_ratio + audio_latents = self.prepare_audio_latents( + batch_size * num_videos_per_prompt, + num_channels_latents=self.audio_latent_channels, + audio_latent_length=audio_num_frames, + num_mel_bins=self.audio_mel_bins, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=audio_latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + mu = calculate_shift( + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_image_seq_len", 1024), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.95), + self.scheduler.config.get("max_shift", 2.05), + ) + + # For now, duplicate the scheduler for use with the audio latents + if self.audio_scheduler is not None: + audio_scheduler = self.audio_scheduler + else: + audio_scheduler = copy.deepcopy(self.scheduler) + audio_timesteps, _ = retrieve_timesteps( + audio_scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Prepare micro-conditions + # Pre-compute video and audio positional ids as they will be the same at each step of the denoising loop + video_coords = self.transformer.rope.prepare_video_coords( + latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate + ) + audio_coords = self.transformer.audio_rope.prepare_audio_coords( + audio_latents.shape[0], audio_num_frames, audio_latents.device + ) + if keyframe_coords is not None: + video_coords = torch.cat([video_coords, keyframe_coords], dim=2) + # Duplicate the positional ids as well if using CFG + if self.do_classifier_free_guidance: + video_coords = video_coords.repeat((2,) + (1,) * (video_coords.ndim - 1)) # Repeat twice in batch dim + audio_coords = audio_coords.repeat((2,) + (1,) * (audio_coords.ndim - 1)) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + audio_latent_model_input = ( + torch.cat([audio_latents] * 2) if self.do_classifier_free_guidance else audio_latents + ) + audio_latent_model_input = audio_latent_model_input.to(prompt_embeds.dtype) + + timestep = t.expand(latent_model_input.shape[0]) + video_timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask.squeeze(-1)) + + t_audio = audio_timesteps[i] + audio_timestep = t_audio.expand(latent_model_input.shape[0]) + + with self.transformer.cache_context("cond_uncond"): + noise_pred_video, noise_pred_audio = self.transformer( + hidden_states=latent_model_input, + audio_hidden_states=audio_latent_model_input, + encoder_hidden_states=connector_prompt_embeds, + audio_encoder_hidden_states=connector_audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=audio_timestep, + sigma=timestep, # Used by LTX-2.3 + audio_sigma=audio_timestep, + encoder_attention_mask=connector_attention_mask, + audio_encoder_attention_mask=connector_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_coords, + audio_coords=audio_coords, + isolate_modalities=False, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video = noise_pred_video.float() + noise_pred_audio = noise_pred_audio.float() + + if self.do_classifier_free_guidance: + noise_pred_video_uncond_text, noise_pred_video = noise_pred_video.chunk(2) + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_video_uncond_text = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_text, i, self.scheduler + ) + # Use delta formulation as it works more nicely with multiple guidance terms + video_cfg_delta = (self.guidance_scale - 1) * (noise_pred_video - noise_pred_video_uncond_text) + + noise_pred_audio_uncond_text, noise_pred_audio = noise_pred_audio.chunk(2) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + noise_pred_audio_uncond_text = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_text, i, audio_scheduler + ) + audio_cfg_delta = (self.audio_guidance_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_text + ) + + # Get positive values from merged CFG inputs in case we need to do other DiT forward passes + if self.do_spatio_temporal_guidance or self.do_modality_isolation_guidance: + if i == 0: + # Only split values that remain constant throughout the loop once + video_prompt_embeds = connector_prompt_embeds.chunk(2, dim=0)[1] + audio_prompt_embeds = connector_audio_prompt_embeds.chunk(2, dim=0)[1] + prompt_attn_mask = connector_attention_mask.chunk(2, dim=0)[1] + + video_pos_ids = video_coords.chunk(2, dim=0)[0] + audio_pos_ids = audio_coords.chunk(2, dim=0)[0] + + # Split values that vary each denoising loop iteration + timestep = timestep.chunk(2, dim=0)[0] + video_timestep = video_timestep.chunk(2, dim=0)[0] + audio_timestep = audio_timestep.chunk(2, dim=0)[0] + else: + video_cfg_delta = audio_cfg_delta = 0 + + video_prompt_embeds = connector_prompt_embeds + audio_prompt_embeds = connector_audio_prompt_embeds + prompt_attn_mask = connector_attention_mask + + video_pos_ids = video_coords + audio_pos_ids = audio_coords + + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + + if self.do_spatio_temporal_guidance: + with self.transformer.cache_context("uncond_stg"): + noise_pred_video_uncond_stg, noise_pred_audio_uncond_stg = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=audio_timestep, + sigma=timestep, # Used by LTX-2.3 + audio_sigma=audio_timestep, + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + isolate_modalities=False, + # Use STG at given blocks to perturb model + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_stg = noise_pred_video_uncond_stg.float() + noise_pred_audio_uncond_stg = noise_pred_audio_uncond_stg.float() + noise_pred_video_uncond_stg = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_stg, i, self.scheduler + ) + noise_pred_audio_uncond_stg = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_stg, i, audio_scheduler + ) + + video_stg_delta = self.stg_scale * (noise_pred_video - noise_pred_video_uncond_stg) + audio_stg_delta = self.audio_stg_scale * (noise_pred_audio - noise_pred_audio_uncond_stg) + else: + video_stg_delta = audio_stg_delta = 0 + + if self.do_modality_isolation_guidance: + with self.transformer.cache_context("uncond_modality"): + noise_pred_video_uncond_modality, noise_pred_audio_uncond_modality = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=audio_timestep, + sigma=timestep, # Used by LTX-2.3 + audio_sigma=audio_timestep, + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + # Turn off A2V and V2A cross attn to isolate video and audio modalities + isolate_modalities=True, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_modality = noise_pred_video_uncond_modality.float() + noise_pred_audio_uncond_modality = noise_pred_audio_uncond_modality.float() + noise_pred_video_uncond_modality = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_modality, i, self.scheduler + ) + noise_pred_audio_uncond_modality = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_modality, i, audio_scheduler + ) + + video_modality_delta = (self.modality_scale - 1) * ( + noise_pred_video - noise_pred_video_uncond_modality + ) + audio_modality_delta = (self.audio_modality_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_modality + ) + else: + video_modality_delta = audio_modality_delta = 0 + + # Now apply all guidance terms + noise_pred_video_g = noise_pred_video + video_cfg_delta + video_stg_delta + video_modality_delta + noise_pred_audio_g = noise_pred_audio + audio_cfg_delta + audio_stg_delta + audio_modality_delta + + # Apply LTX-2.X guidance rescaling + if self.guidance_rescale > 0: + noise_pred_video = rescale_noise_cfg( + noise_pred_video_g, noise_pred_video, guidance_rescale=self.guidance_rescale + ) + else: + noise_pred_video = noise_pred_video_g + + if self.audio_guidance_rescale > 0: + noise_pred_audio = rescale_noise_cfg( + noise_pred_audio_g, noise_pred_audio, guidance_rescale=self.audio_guidance_rescale + ) + else: + noise_pred_audio = noise_pred_audio_g + + # NOTE: use only the first chunk of conditioning mask in case it is duplicated for CFG + bsz = noise_pred_video.size(0) + # Apply the (packed) conditioning mask to the denoised (x0) sample and clean conditioning. The + # conditioning mask contains conditioning strengths from 0 (always use denoised sample) to 1 (always + # use conditions), with intermediate values specifying how strongly to follow the conditions. + # NOTE: this operation should be applied in sample (x0) space and not velocity space (which is the + # space the denoising model outputs are in) + denoised_sample_cond = ( + noise_pred_video * (1 - conditioning_mask[:bsz]) + clean_latents * conditioning_mask[:bsz] + ).to(noise_pred_video.dtype) + + # Convert the denoised (x0) sample back to a velocity for the scheduler + noise_pred_video = self.convert_x0_to_velocity(latents, denoised_sample_cond, i, self.scheduler) + noise_pred_audio = self.convert_x0_to_velocity(audio_latents, noise_pred_audio, i, audio_scheduler) + + # Compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred_video, t, latents, return_dict=False)[0] + + # NOTE: for now duplicate scheduler for audio latents in case self.scheduler sets internal state in + # the step method (such as _step_index) + audio_latents = audio_scheduler.step(noise_pred_audio, t, audio_latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # Remove any appended keyframe (non-first-frame) condition tokens from the final latent + base_token_count = latent_num_frames * latent_height * latent_width + latents = latents[:, :base_token_count] + + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + audio_latents = self._denormalize_audio_latents( + audio_latents, self.audio_vae.latents_mean, self.audio_vae.latents_std + ) + audio_latents = self._unpack_audio_latents(audio_latents, audio_num_frames, num_mel_bins=latent_mel_bins) + + if output_type == "latent": + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + video = latents + audio = audio_latents + else: + latents = latents.to(prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + + latents = latents.to(self.vae.dtype) + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + audio_latents = audio_latents.to(self.audio_vae.dtype) + generated_mel_spectrograms = self.audio_vae.decode(audio_latents, return_dict=False)[0] + audio = self.vocoder(generated_mel_spectrograms) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video, audio) + + return LTX2PipelineOutput(frames=video, audio=audio) diff --git a/diffusers/pipelines/ltx2/pipeline_ltx2_hdr_lora.py b/diffusers/pipelines/ltx2/pipeline_ltx2_hdr_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..38cd69b66c645d1b43dead607ad3eebeed0b582e --- /dev/null +++ b/diffusers/pipelines/ltx2/pipeline_ltx2_hdr_lora.py @@ -0,0 +1,1611 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import inspect +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import Gemma3ForConditionalGeneration, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import FromSingleFileMixin, LTX2LoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTX2Audio, AutoencoderKLLTX2Video +from ...models.transformers import LTX2VideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .connectors import LTX2TextConnectors +from .image_processor import LTX2VideoHDRProcessor +from .pipeline_output import LTX2PipelineOutput +from .vocoder import LTX2Vocoder, LTX2VocoderWithBWE + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class LTX2HDRReferenceCondition: + r""" + A reference video condition for HDR IC-LoRA conditioning. + + The reference video is encoded into latent tokens and concatenated to the noisy latent sequence during denoising, + allowing the HDR IC-LoRA adapter to condition the generation on the reference video content. + + Matches the `(video_path, strength)` tuples consumed by the reference `HDRICLoraPipeline`'s `video_conditioning` + argument. + + Attributes: + frames (`PIL.Image.Image` or `List[PIL.Image.Image]` or `np.ndarray` or `torch.Tensor`): + The reference video frames. Accepts any type handled by `VideoProcessor.preprocess_video`. + strength (`float`, defaults to `1.0`): + Controls how "clean" the reference tokens appear to the model. A value of `1.0` means fully clean + (per-token timestep=0), `0.0` means fully noisy. + """ + + frames: PIL.Image.Image | list[PIL.Image.Image] | np.ndarray | torch.Tensor + strength: float = 1.0 + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from safetensors import safe_open + >>> from diffusers import LTX2HDRPipeline + >>> from diffusers.pipelines.ltx2.pipeline_ltx2_hdr_lora import LTX2HDRReferenceCondition + >>> from diffusers.pipelines.ltx2.utils import DISTILLED_SIGMA_VALUES + >>> from diffusers.pipelines.ltx2.export_utils import encode_hdr_tensor_to_mp4 + >>> from diffusers.utils import load_video + + >>> pipe = LTX2HDRPipeline.from_pretrained("diffusers/LTX-2.3-Distilled-Diffusers", torch_dtype=torch.bfloat16) + >>> pipe.enable_sequential_cpu_offload(device="cuda") + >>> pipe.load_lora_weights( + ... "Lightricks/LTX-2.3-22b-IC-LoRA-HDR", + ... adapter_name="hdr_lora", + ... weight_name="ltx-2.3-22b-ic-lora-hdr-0.9.safetensors", + ... ) + >>> pipe.set_adapters("hdr_lora", 1.0) + + >>> reference_video = load_video("/path/to/reference.mp4") + >>> ref_cond = LTX2HDRReferenceCondition(frames=reference_video, strength=1.0) + + >>> # Load pre-computed HDR LoRA connector embeddings. + >>> with safe_open("/path/to/connector/embeds.safetensors", framework="pt", device="cuda") as f: + ... connector_video_embeds = f.get_tensor("video_context") + ... connector_audio_embeds = f.get_tensor("audio_context") + + >>> # `hdr_video` is a linear HDR tensor of shape (batch, frames, H, W, C). + >>> hdr_video = pipe( + ... reference_conditions=[ref_cond], + ... connector_video_embeds=connector_video_embeds, + ... connector_audio_embeds=connector_audio_embeds, + ... width=768, + ... height=512, + ... num_frames=121, + ... frame_rate=24.0, + ... num_inference_steps=8, + ... sigmas=DISTILLED_SIGMA_VALUES, + ... guidance_scale=1.0, + ... output_type="pt", + ... return_dict=False, + ... )[0] + + >>> # Convert the HDR video to a SDR sRGB-tonemapped `.mp4` video. + >>> # A custom tone-mapper can be specified via the `tone_mapping_fn` argument. + >>> encode_hdr_tensor_to_mp4(hdr_video[0], "ltx2_hdr_lora_output.mp4", frame_rate=24.0) + ``` +""" + + +# Copied from diffusers.pipelines.ltx2.pipeline_ltx2_ic_lora.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTX2HDRPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraLoaderMixin): + r""" + Pipeline for LTX-2.X HDR video generation with reference video conditioning. + + The pipeline accepts a reference SDR ("normal") video and generates a linear HDR output with values in `[0, ∞)` via + a LogC3 inverse transform which has the same content as the reference video. The motivating use case for this + pipeline is to support LTX-2.X HDR IC-LoRAs, but it should support any LTX-2.X-like model that operates on HDR + inputs as above. + + Compared to [`LTX2InContextPipeline`], the HDR pipeline has the following differences: + + - Video-only (no audio output). The transformer's audio branch is still run since the diffusers transformer API + requires audio inputs, but the decoded audio is discarded and audio-specific guidance scales are fixed to no-op + values to avoid wasted compute. + - No frame-level keyframe conditioning (the reference HDR pipeline does not support this). + + Two-stage inference is supported through separate calls to `__call__`: + + - **Stage 1**: generate video latents at target resolution with HDR IC-LoRA conditioning (`output_type="latent"`). + - **Stage 2**: upsample via [`LTX2LatentUpsamplePipeline`] and refine with this same pipeline (or [`LTX2Pipeline`]) + by passing `latents=upsampled_latents`. The reference HDR stage-2 additionally supports spatial/temporal tiling + of the refinement pass — that optimization is not yet implemented here. + + Reference: https://github.com/Lightricks/LTX-2 Paper: https://huggingface.co/papers/2604.11788 + + Args: + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + Scheduler used in the denoising loop. + vae ([`AutoencoderKLLTX2Video`]): + Video VAE. + audio_vae ([`AutoencoderKLLTX2Audio`]): + Audio VAE. Required for transformer compatibility; its outputs are discarded. + text_encoder ([`transformers.Gemma3ForConditionalGeneration`]): + Text encoder. + tokenizer (`GemmaTokenizer` or `GemmaTokenizerFast`): + Tokenizer for the text encoder. + connectors ([`LTX2TextConnectors`]): + Text connector stack for the transformer. + transformer ([`LTX2VideoTransformer3DModel`]): + Transformer backbone. + vocoder ([`LTX2Vocoder`] or [`LTX2VocoderWithBWE`]): + Vocoder. Required for transformer compatibility; its outputs are discarded. + hdr_transform (`str`, *optional*, defaults to `"logc3"`): + HDR transform identifier applied during postprocessing. Currently only `"logc3"` is supported. + """ + + model_cpu_offload_seq = "text_encoder->connectors->transformer->vae->audio_vae->vocoder" + _optional_components = ["audio_scheduler"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTX2Video, + audio_vae: AutoencoderKLLTX2Audio, + text_encoder: Gemma3ForConditionalGeneration, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + connectors: LTX2TextConnectors, + transformer: LTX2VideoTransformer3DModel, + vocoder: LTX2Vocoder | LTX2VocoderWithBWE, + audio_scheduler: FlowMatchEulerDiscreteScheduler | None = None, + hdr_transform: str = "logc3", + ): + super().__init__() + + self.register_modules( + vae=vae, + audio_vae=audio_vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + connectors=connectors, + transformer=transformer, + vocoder=vocoder, + scheduler=scheduler, + audio_scheduler=audio_scheduler, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.audio_vae_mel_compression_ratio = ( + self.audio_vae.mel_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.audio_vae_temporal_compression_ratio = ( + self.audio_vae.temporal_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.audio_sampling_rate = ( + self.audio_vae.config.sample_rate if getattr(self, "audio_vae", None) is not None else 16000 + ) + self.audio_hop_length = ( + self.audio_vae.config.mel_hop_length if getattr(self, "audio_vae", None) is not None else 160 + ) + self.audio_mel_bins = self.audio_vae.config.mel_bins if getattr(self, "audio_vae", None) is not None else 64 + self.audio_latent_channels = ( + self.audio_vae.config.latent_channels if getattr(self, "audio_vae", None) is not None else 8 + ) + + self.hdr_video_processor = LTX2VideoHDRProcessor( + vae_scale_factor=self.vae_spatial_compression_ratio, + hdr_transform=hdr_transform, + ) + + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 1024 + ) + tokenizer_padding_side = "left" + if getattr(self, "tokenizer", None) is not None: + tokenizer_padding_side = getattr(self.tokenizer, "padding_side", "left") + self.tokenizer_padding_side = tokenizer_padding_side + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`str` or `torch.device`): + torch device to place the resulting embeddings on + dtype: (`torch.dtype`): + torch dtype to cast the prompt embeds to + max_sequence_length (`int`, defaults to 1024): Maximum sequence length to use for the prompt. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if getattr(self, "tokenizer", None) is not None: + # Gemma expects left padding for chat-style prompts + self.tokenizer.padding_side = "left" + if self.tokenizer.pad_token is None: + self.tokenizer.pad_token = self.tokenizer.eos_token + + prompt = [p.strip() for p in prompt] + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + text_input_ids = text_input_ids.to(device) + prompt_attention_mask = prompt_attention_mask.to(device) + + text_encoder_outputs = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ) + text_encoder_hidden_states = text_encoder_outputs.hidden_states + text_encoder_hidden_states = torch.stack(text_encoder_hidden_states, dim=-1) + prompt_embeds = text_encoder_hidden_states.flatten(2, 3).to(dtype=dtype) # Pack to 3D + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + connector_video_embeds=None, + connector_audio_embeds=None, + latents=None, + spatio_temporal_guidance_blocks=None, + stg_scale=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + if connector_video_embeds is None or connector_audio_embeds is None: + raise ValueError( + "Provide a `prompt`, `prompt_embeds` or `connector_video_embeds` and `connector_audio_embeds`" + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if latents is not None and latents.ndim != 5: + raise ValueError( + f"Only unpacked (5D) video latents of shape `[batch_size, latent_channels, latent_frames," + f" latent_height, latent_width] are supported, but got {latents.ndim} dims." + ) + + if (stg_scale is not None and stg_scale > 0.0) and not spatio_temporal_guidance_blocks: + raise ValueError( + "Spatio-Temporal Guidance (STG) is specified but no STG blocks are supplied. Please supply a list of" + " block indices at which to apply STG in `spatio_temporal_guidance_blocks`" + ) + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_latents + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_image2video.LTX2ImageToVideoPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._create_noised_state + def _create_noised_state( + latents: torch.Tensor, noise_scale: float | torch.Tensor, generator: torch.Generator | None = None + ): + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + noised_latents = noise_scale * noise + (1 - noise_scale) * latents + return noised_latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._normalize_audio_latents + def _normalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents - latents_mean) / latents_std + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_audio_latents + def _denormalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents * latents_std) + latents_mean + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_audio_latents + def _pack_audio_latents( + latents: torch.Tensor, patch_size: int | None = None, patch_size_t: int | None = None + ) -> torch.Tensor: + # Audio latents shape: [B, C, L, M], where L is the latent audio length and M is the number of mel bins + if patch_size is not None and patch_size_t is not None: + # Packs the latents into a patch sequence of shape [B, L // p_t * M // p, C * p_t * p] (a ndim=3 tnesor). + # dim=1 is the effective audio sequence length and dim=2 is the effective audio input feature size. + batch_size, num_channels, latent_length, latent_mel_bins = latents.shape + post_patch_latent_length = latent_length / patch_size_t + post_patch_mel_bins = latent_mel_bins / patch_size + latents = latents.reshape( + batch_size, -1, post_patch_latent_length, patch_size_t, post_patch_mel_bins, patch_size + ) + latents = latents.permute(0, 2, 4, 1, 3, 5).flatten(3, 5).flatten(1, 2) + else: + # Packs the latents into a patch sequence of shape [B, L, C * M]. This implicitly assumes a (mel) + # patch_size of M (all mel bins constitutes a single patch) and a patch_size_t of 1. + latents = latents.transpose(1, 2).flatten(2, 3) # [B, C, L, M] --> [B, L, C * M] + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_audio_latents + def _unpack_audio_latents( + latents: torch.Tensor, + latent_length: int, + num_mel_bins: int, + patch_size: int | None = None, + patch_size_t: int | None = None, + ) -> torch.Tensor: + # Unpacks an audio patch sequence of shape [B, S, D] into a latent spectrogram tensor of shape [B, C, L, M], + # where L is the latent audio length and M is the number of mel bins. + if patch_size is not None and patch_size_t is not None: + batch_size = latents.size(0) + latents = latents.reshape(batch_size, latent_length, num_mel_bins, -1, patch_size_t, patch_size) + latents = latents.permute(0, 3, 1, 4, 2, 5).flatten(4, 5).flatten(2, 3) + else: + # Assume [B, S, D] = [B, L, C * M], which implies that patch_size = M and patch_size_t = 1. + latents = latents.unflatten(2, (-1, num_mel_bins)).transpose(1, 2) + return latents + + def prepare_latents( + self, + reference_conditions: list[LTX2HDRReferenceCondition] | None = None, + reference_downscale_factor: int = 1, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor | None, int, torch.Tensor | None]: + r""" + Prepare noisy video latents, applying HDR IC-LoRA reference-video conditioning. + + Builds a packed latent sequence in the order `[base | reference]`: + - Base: either fresh noise (Stage 1, `latents=None`) or pre-existing upsampled latents (Stage 2). + - Reference: HDR-encoded reference-video tokens appended with per-token `conditioning_mask = strength`, + following the same pattern as [`LTX2InContextPipeline.prepare_latents`]. (HDR LoRA does not currently take + per-frame `conditions`, so there is no first-frame / keyframe block in between.) + + Returns a 6-tuple matching [`LTX2InContextPipeline.prepare_latents`]: + - `latents`: packed noisy latents `(B, base + n_ref, C)`. + - `conditioning_mask`: `(B, seq_len, 1)` with `strength` at reference positions, `0` elsewhere. + - `clean_latents`: clean reference values at reference positions (zeros elsewhere); same shape as + `latents`. + - `appended_coords`: `[1, 3, n_ref, 2]` reference coordinates to concat onto `video_coords`, or `None` when + no reference conditions are provided. + - `num_ref_tokens`: count of reference tokens at the END of `latents`. + - `ref_cross_mask`: always `None` for HDR LoRA (no cross-attention masking support). + """ + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + + if isinstance(generator, list): + if len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective" + f" batch size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Build the base noisy latents at the maximum sigma (zeros for Stage 1 fresh noise; normalized provided latents + # for Stage 2). The noise mixing at the bottom converts these into the right partial-denoise state. + if latents is not None: + if latents.ndim == 5: + latents = self._normalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + if latents.ndim != 3: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is [batch_size," + f" num_seq, num_features]." + ) + else: + shape = (batch_size, num_channels_latents, latent_num_frames, latent_height, latent_width) + latents = torch.zeros(shape, device=device, dtype=dtype) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + latents = latents.to(device=device, dtype=dtype) + + # Build conditioning_mask and clean_latents over the base token sequence (zeros — base is unconditioned). + base_seq_len = latents.shape[1] + conditioning_mask = torch.zeros((batch_size, base_seq_len, 1), device=device, dtype=dtype) + clean_latents = torch.zeros_like(latents) + + # Append reference tokens (if any) as a contiguous block at the end of the sequence with per-token + # `conditioning_mask = strength` and `clean_latents = encoded_ref`. + ref_coords: torch.Tensor | None = None + num_ref_tokens = 0 + if reference_conditions is not None and len(reference_conditions) > 0: + ref_latents_packed, ref_coords, _ = self._encode_reference_conditions( + reference_conditions=reference_conditions, + num_frames=num_frames, + height=height, + width=width, + reference_downscale_factor=reference_downscale_factor, + frame_rate=frame_rate, + dtype=dtype, + device=device, + generator=generator[0] if isinstance(generator, list) else generator, + ) + num_ref_tokens = ref_latents_packed.shape[1] + + # All reference videos preprocess to the same shape, so split tokens evenly across conditions. + n_per_ref = num_ref_tokens // len(reference_conditions) + ref_mask_chunks = [ + torch.full( + (batch_size, n_per_ref, 1), + float(ref_cond.strength), + device=device, + dtype=conditioning_mask.dtype, + ) + for ref_cond in reference_conditions + ] + ref_mask_full = torch.cat(ref_mask_chunks, dim=1) + + ref_latents_packed_b = ref_latents_packed.expand(batch_size, -1, -1) + latents = torch.cat([latents, ref_latents_packed_b], dim=1) + conditioning_mask = torch.cat([conditioning_mask, ref_mask_full], dim=1) + clean_latents = torch.cat([clean_latents, ref_latents_packed_b], dim=1) + + # HDR LoRA has no keyframe conditions, so the only appended tokens are reference tokens. + appended_coords = ref_coords + + # The conditioning_mask values have the following semantics: + # - mask=0: fully noise tokens (e.g. noisy latents) + # - mask=1: keep fully clean (e.g. I2V first-frame condition, conditions with strength=1) + # - mask in (0, 1): use intermediate noise level mask * sigma_i (noise_scale == sigma_0) + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + scaled_mask = (1.0 - conditioning_mask) * noise_scale # noise to initial noise level `noise_scale` + latents = noise * scaled_mask + latents * (1 - scaled_mask) + + return latents, conditioning_mask, clean_latents, appended_coords, num_ref_tokens, None + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.prepare_audio_latents + def prepare_audio_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 8, + audio_latent_length: int = 1, # 1 is just a dummy value + num_mel_bins: int = 64, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + # latents expected to be unpacked (4D) with shape [B, C, L, M] + latents = self._pack_audio_latents(latents) + latents = self._normalize_audio_latents(latents, self.audio_vae.latents_mean, self.audio_vae.latents_std) + latents = self._create_noised_state(latents, noise_scale, generator) + return latents.to(device=device, dtype=dtype) + + latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Sample in packed shape (B, L, C * M), following the original LTX-2.X code + packed_shape = (batch_size, audio_latent_length, num_channels_latents * latent_mel_bins) + latents = randn_tensor(packed_shape, generator=generator, device=device, dtype=dtype) + return latents + + def _encode_reference_conditions( + self, + reference_conditions: list[LTX2HDRReferenceCondition], + height: int, + width: int, + num_frames: int, + reference_downscale_factor: int = 1, + frame_rate: float = 24.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]: + """Encode HDR IC-LoRA reference videos into `(reference_latents, reference_coords, reference_cross_mask)`. + + Shared encoding core used by both `prepare_latents` (which folds reference tokens into the main noisy sequence) + and the back-compat shim `prepare_reference_latents`. HDR LoRA does not currently support cross-attention + masking for reference tokens, so the third return is always `None`. + """ + ref_height = height // reference_downscale_factor + ref_width = width // reference_downscale_factor + + if reference_downscale_factor != 1 and ( + height % reference_downscale_factor != 0 or width % reference_downscale_factor != 0 + ): + raise ValueError( + f"Output dimensions ({height}x{width}) must be divisible by reference_downscale_factor " + f"({reference_downscale_factor})." + ) + + all_ref_latents = [] + all_ref_coords = [] + + for ref_cond in reference_conditions: + if isinstance(ref_cond.frames, PIL.Image.Image): + video_like = [ref_cond.frames] + elif isinstance(ref_cond.frames, np.ndarray) and ref_cond.frames.ndim == 3: + video_like = np.expand_dims(ref_cond.frames, axis=0) + elif isinstance(ref_cond.frames, torch.Tensor) and ref_cond.frames.ndim == 3: + video_like = ref_cond.frames.unsqueeze(0) + else: + video_like = ref_cond.frames + + # HDR-specific preprocessing: reflect-pad resize (vs center-crop in the standard IC-LoRA pipeline). + # For LDR reference videos the numerical output of `preprocess_reference_video_hdr` is identical to the + # standard [-1, 1] normalization since LogC3's `compress_ldr` is an identity clamp. + ref_pixels = self.hdr_video_processor.preprocess_reference_video_hdr(video_like, ref_height, ref_width) + ref_pixels = ref_pixels[:, :, :num_frames] + ref_pixels = ref_pixels.to(dtype=self.vae.dtype, device=device) + + ref_latent = retrieve_latents(self.vae.encode(ref_pixels), generator=generator, sample_mode="argmax") + ref_latent = self._normalize_latents(ref_latent, self.vae.latents_mean, self.vae.latents_std).to( + device=device, dtype=dtype + ) + + _, _, ref_latent_frames, ref_latent_height, ref_latent_width = ref_latent.shape + + ref_latent_packed = self._pack_latents( + ref_latent, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + + ref_coords = self.transformer.rope.prepare_video_coords( + batch_size=1, + num_frames=ref_latent_frames, + height=ref_latent_height, + width=ref_latent_width, + device=device, + fps=frame_rate, + ) + if reference_downscale_factor != 1: + ref_coords[:, 1, :, :] = ref_coords[:, 1, :, :] * reference_downscale_factor + ref_coords[:, 2, :, :] = ref_coords[:, 2, :, :] * reference_downscale_factor + + all_ref_latents.append(ref_latent_packed) + all_ref_coords.append(ref_coords) + + reference_latents = torch.cat(all_ref_latents, dim=1) + reference_coords = torch.cat(all_ref_coords, dim=2) + + return reference_latents, reference_coords, None + + def prepare_reference_latents( + self, + reference_conditions: list[LTX2HDRReferenceCondition], + height: int, + width: int, + num_frames: int, + reference_downscale_factor: int = 1, + frame_rate: float = 24.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + r""" + Encode reference videos with HDR preprocessing into packed latent tokens and compute positional coordinates. + + Each reference video is preprocessed via [`LTX2VideoHDRProcessor.preprocess_reference_video_hdr`] (reflect-pad + resize at the reference resolution), VAE-encoded, packed into tokens, and paired with positional coordinates + computed at the reference latent dimensions and scaled by `reference_downscale_factor`. + + Returns a 3-tuple `(reference_latents, reference_coords, reference_denoise_factors)` with the same shapes as + [`LTX2InContextPipeline.prepare_reference_latents`]. + """ + reference_latents, reference_coords, _ = self._encode_reference_conditions( + reference_conditions=reference_conditions, + height=height, + width=width, + num_frames=num_frames, + reference_downscale_factor=reference_downscale_factor, + frame_rate=frame_rate, + dtype=dtype, + device=device, + generator=generator, + ) + + # Materialize per-token denoise factors for callers that still expect the 3-tuple. Each ref video has + # `1 - strength` for all of its tokens; we rebuild this from the per-video token counts. All ref videos + # preprocess to the same shape, so total token count divides equally across them. + n_total = reference_latents.shape[1] + n_per_ref = n_total // max(len(reference_conditions), 1) + denoise_chunks = [ + torch.full((1, n_per_ref), 1.0 - ref_cond.strength, device=reference_latents.device, dtype=torch.float32) + for ref_cond in reference_conditions + ] + reference_denoise_factors = ( + torch.cat(denoise_chunks, dim=1) if denoise_chunks else reference_latents.new_zeros((1, 0)) + ) + return reference_latents, reference_coords, reference_denoise_factors + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.convert_velocity_to_x0 + def convert_velocity_to_x0( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_x0 = sample - denoised_output * scheduler.sigmas[step_idx] + return sample_x0 + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.convert_x0_to_velocity + def convert_x0_to_velocity( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_v = (sample - denoised_output) / scheduler.sigmas[step_idx] + return sample_v + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def stg_scale(self): + return self._stg_scale + + @property + def modality_scale(self): + return self._modality_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def do_spatio_temporal_guidance(self): + return self._stg_scale > 0.0 + + @property + def do_modality_isolation_guidance(self): + return self._modality_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + reference_conditions: LTX2HDRReferenceCondition | list[LTX2HDRReferenceCondition] | None = None, + reference_downscale_factor: int = 1, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + num_inference_steps: int = 8, + sigmas: list[float] | None = None, + timesteps: list[float] | None = None, + guidance_scale: float = 1.0, + stg_scale: float = 0.0, + modality_scale: float = 1.0, + guidance_rescale: float = 0.0, + spatio_temporal_guidance_blocks: list[int] | None = None, + noise_scale: float | None = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + connector_video_embeds: torch.Tensor | None = None, + connector_audio_embeds: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + use_cross_timestep: bool = False, + output_type: str = "pt", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 1024, + ): + r""" + Run HDR IC-LoRA video generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt(s) to guide generation. Either `prompt` or `prompt_embeds` must be supplied. + negative_prompt (`str` or `List[str]`, *optional*): + The negative prompt(s). Ignored when `guidance_scale <= 1.0`. + reference_conditions (`LTX2HDRReferenceCondition` or `List[LTX2HDRReferenceCondition]`, *optional*): + Reference video conditions for HDR IC-LoRA conditioning. + reference_downscale_factor (`int`, *optional*, defaults to `1`): + Ratio between target and reference video resolutions. IC-LoRA models trained with downscaled reference + videos store this factor in their safetensors metadata. + height (`int`, *optional*, defaults to `512`): + Output video height in pixels. Must be divisible by 32. + width (`int`, *optional*, defaults to `768`): + Output video width in pixels. Must be divisible by 32. + num_frames (`int`, *optional*, defaults to `121`): + Number of frames to generate. Must satisfy `(n - 1) % 8 == 0`. + frame_rate (`float`, *optional*, defaults to `24.0`): + Output frame rate (used for temporal positional encoding). + num_inference_steps (`int`, *optional*, defaults to `8`): + Number of denoising steps. Default matches the distilled model schedule. + sigmas (`List[float]`, *optional*): + Custom sigma schedule. Overrides `num_inference_steps` when set. + timesteps (`List[float]`, *optional*): + Custom timesteps schedule. Overrides `num_inference_steps` when set. + guidance_scale (`float`, *optional*, defaults to `1.0`): + Classifier-Free Guidance scale for video. Default `1.0` disables CFG (matches the distilled model). + stg_scale (`float`, *optional*, defaults to `0.0`): + Spatio-Temporal Guidance scale for video. + modality_scale (`float`, *optional*, defaults to `1.0`): + Modality isolation guidance scale for video. + guidance_rescale (`float`, *optional*, defaults to `0.0`): + Video guidance rescale factor. + spatio_temporal_guidance_blocks (`list[int]`, *optional*): + Transformer block indices at which to apply STG. + noise_scale (`float`, *optional*): + Noise scale used when preparing the initial latents. Inferred from the sigma schedule when unset. + num_videos_per_prompt (`int`, *optional*, defaults to `1`): + Number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + Random generator(s) for reproducibility. + latents (`torch.Tensor`, *optional*): + Pre-generated video latents. Pass output from [`LTX2LatentUpsamplePipeline`] here for Stage 2. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Bypasses `prompt`/`tokenizer`/`text_encoder` if supplied. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + connector_video_embeds (`torch.Tensor`, *optional*): + Optional pre-computed connector outputs for the video modality. Used by the HDR LoRA pipeline; if + supplied, will override any `prompt`/`prompt_embeds`. + connector_audio_embeds (`torch.Tensor`, *optional*): + Optional pre-computed connector outputs for the audio modality. Used by the HDR LoRA pipeline; if + supplied, will override any `prompt`/`prompt_embeds`. + decode_timestep (`float` or `list[float]`, defaults to `0.0`): + VAE-decode timestep conditioning (only used by VAE configs with `timestep_conditioning=True`). + decode_noise_scale (`float` or `list[float]`, *optional*): + Interpolation factor between random noise and denoised latents at the decode timestep. + use_cross_timestep (`bool`, *optional*, defaults to `False`): + Whether to use cross-modality sigma for cross-attention modulation. + output_type (`str`, *optional*, defaults to `"pt"`): + One of `"pt"`, `"np"`, or `"latent"`. `"pt"` returns a linear HDR torch tensor in `[0, ∞)` of shape + `(batch_size, num_frames, height, width, channels)`; `"np"` returns the equivalent `float32` NumPy + array; `"latent"` returns the raw denoised latents (skip the HDR decode). + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return an [`LTX2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor`. + callback_on_step_end (`Callable`, *optional*): + A function called at the end of each denoising step, same as [`LTX2InContextPipeline`]. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs passed to `callback_on_step_end`. + max_sequence_length (`int`, *optional*, defaults to `1024`): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`LTX2PipelineOutput`] or `tuple`. When `return_dict=False`, returns `(frames, None)` — the audio slot is + always `None` since this pipeline is video-only. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + connector_video_embeds=connector_video_embeds, + connector_audio_embeds=connector_audio_embeds, + latents=latents, + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + stg_scale=stg_scale, + ) + + # Video-only guidance state. + self._guidance_scale = guidance_scale + self._stg_scale = stg_scale + self._modality_scale = modality_scale + self._guidance_rescale = guidance_rescale + + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + elif prompt_embeds is not None: + batch_size = prompt_embeds.shape[0] + else: + batch_size = connector_video_embeds.shape[0] + + if reference_conditions is not None and not isinstance(reference_conditions, list): + reference_conditions = [reference_conditions] + + if noise_scale is None: + noise_scale = sigmas[0] if sigmas is not None else 1.0 + + device = self._execution_device + + # 3. Prepare text embeddings + if connector_video_embeds is None or connector_audio_embeds is None: + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + connector_prompt_embeds, connector_audio_prompt_embeds, connector_attention_mask = self.connectors( + prompt_embeds, prompt_attention_mask, padding_side=self.tokenizer_padding_side + ) + else: + connector_prompt_embeds = connector_video_embeds.to(device=device, dtype=self.transformer.dtype) + connector_audio_prompt_embeds = connector_audio_embeds.to(device=device, dtype=self.transformer.dtype) + connector_attention_mask = None + + # 4. Prepare video latents + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + if latents is not None: + logger.info( + "Got pre-supplied latents of shape %s; `latent_num_frames`, `latent_height`, and `latent_width` will" + " be inferred.", + tuple(latents.shape), + ) + _, _, latent_num_frames, latent_height, latent_width = latents.shape + + num_channels_latents = self.transformer.config.in_channels + latents, conditioning_mask, clean_latents, appended_coords, num_ref_tokens, _ = self.prepare_latents( + reference_conditions=reference_conditions, + reference_downscale_factor=reference_downscale_factor, + batch_size=batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + num_frames=num_frames, + frame_rate=frame_rate, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=latents, + ) + # Track the base (non-reference) token count so we can trim the appended reference tokens off + # `latents` before unpack/decode at the end. + base_token_count = latents.shape[1] - num_ref_tokens + if self.do_classifier_free_guidance and num_ref_tokens > 0: + conditioning_mask = torch.cat([conditioning_mask, conditioning_mask]) + + # 5. Prepare audio latents. Audio is discarded at the end, but the transformer's audio branch still runs so + # we need well-formed audio inputs. Audio guidance is fixed so no extra audio-only forward passes fire. + duration_s = num_frames / frame_rate + audio_latents_per_second = ( + self.audio_sampling_rate / self.audio_hop_length / float(self.audio_vae_temporal_compression_ratio) + ) + audio_num_frames = round(duration_s * audio_latents_per_second) + + audio_latents = self.prepare_audio_latents( + batch_size * num_videos_per_prompt, + num_channels_latents=self.audio_latent_channels, + audio_latent_length=audio_num_frames, + num_mel_bins=self.audio_mel_bins, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=None, + ) + + # 6. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + mu = calculate_shift( + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_image_seq_len", 1024), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.95), + self.scheduler.config.get("max_shift", 2.05), + ) + if self.audio_scheduler is not None: + audio_scheduler = self.audio_scheduler + else: + audio_scheduler = copy.deepcopy(self.scheduler) + audio_timesteps, _ = retrieve_timesteps( + audio_scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 7. Prepare positional coordinates + video_coords = self.transformer.rope.prepare_video_coords( + latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate + ) + if appended_coords is not None: + # Expand appended_coords to effective batch size (to [B, 3, num_extra_tokens, 2]) + appended_coords = appended_coords.expand(latents.shape[0], -1, -1, -1) + video_coords = torch.cat([video_coords, appended_coords], dim=2) + audio_coords = self.transformer.audio_rope.prepare_audio_coords( + audio_latents.shape[0], audio_num_frames, audio_latents.device + ) + if self.do_classifier_free_guidance: + video_coords = video_coords.repeat((2,) + (1,) * (video_coords.ndim - 1)) + audio_coords = audio_coords.repeat((2,) + (1,) * (audio_coords.ndim - 1)) + + # 8. Denoising loop + video_seq_len = latents.shape[1] + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = latent_model_input.to(connector_prompt_embeds.dtype) + audio_latent_model_input = ( + torch.cat([audio_latents] * 2) if self.do_classifier_free_guidance else audio_latents + ) + audio_latent_model_input = audio_latent_model_input.to(connector_prompt_embeds.dtype) + + timestep_scalar = t.expand(latent_model_input.shape[0]) + if num_ref_tokens > 0: + video_timestep = timestep_scalar.unsqueeze(-1) * (1 - conditioning_mask.squeeze(-1)) + else: + video_timestep = timestep_scalar.unsqueeze(-1).expand(-1, video_seq_len) + + t_audio = audio_timesteps[i] + audio_timestep = t_audio.expand(latent_model_input.shape[0]) + + # --- Main forward pass (cond + uncond for CFG) --- + with self.transformer.cache_context("cond_uncond"): + noise_pred_video, noise_pred_audio = self.transformer( + hidden_states=latent_model_input, + audio_hidden_states=audio_latent_model_input, + encoder_hidden_states=connector_prompt_embeds, + audio_encoder_hidden_states=connector_audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=audio_timestep, + sigma=timestep_scalar, # Used by LTX-2.3 + audio_sigma=audio_timestep, + encoder_attention_mask=connector_attention_mask, + audio_encoder_attention_mask=connector_attention_mask, + video_self_attention_mask=None, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_coords, + audio_coords=audio_coords, + isolate_modalities=False, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video = noise_pred_video.float() + + if self.do_classifier_free_guidance: + noise_pred_video_uncond_text, noise_pred_video = noise_pred_video.chunk(2) + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_video_uncond_text = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_text, i, self.scheduler + ) + video_cfg_delta = (self.guidance_scale - 1) * (noise_pred_video - noise_pred_video_uncond_text) + + if self.do_spatio_temporal_guidance or self.do_modality_isolation_guidance: + if i == 0: + video_prompt_embeds = connector_prompt_embeds.chunk(2, dim=0)[1] + audio_prompt_embeds = connector_audio_prompt_embeds.chunk(2, dim=0)[1] + prompt_attn_mask = connector_attention_mask.chunk(2, dim=0)[1] + video_pos_ids = video_coords.chunk(2, dim=0)[0] + audio_pos_ids = audio_coords.chunk(2, dim=0)[0] + timestep_scalar_single = timestep_scalar.chunk(2, dim=0)[0] + if num_ref_tokens > 0: + video_timestep_single = video_timestep.chunk(2, dim=0)[0] + else: + video_timestep_single = timestep_scalar_single.unsqueeze(-1).expand(-1, video_seq_len) + audio_timestep_single = audio_timestep.chunk(2, dim=0)[0] + else: + video_cfg_delta = 0 + + video_prompt_embeds = connector_prompt_embeds + audio_prompt_embeds = connector_audio_prompt_embeds + prompt_attn_mask = connector_attention_mask + video_pos_ids = video_coords + audio_pos_ids = audio_coords + + timestep_scalar_single = timestep_scalar + if num_ref_tokens > 0: + video_timestep_single = video_timestep + else: + video_timestep_single = timestep_scalar.unsqueeze(-1).expand(-1, video_seq_len) + audio_timestep_single = audio_timestep + + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + + # --- STG forward pass (video only — audio output discarded) --- + if self.do_spatio_temporal_guidance: + with self.transformer.cache_context("uncond_stg"): + noise_pred_video_uncond_stg, noise_pred_audio_uncond_stg = self.transformer( + hidden_states=latents.to(dtype=connector_prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=connector_prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep_single, + audio_timestep=audio_timestep_single, + sigma=timestep_scalar_single, # Used by LTX-2.3 + audio_sigma=audio_timestep_single, + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + video_self_attention_mask=None, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + isolate_modalities=False, + # Use STG at given blocks to perturb model + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_stg = noise_pred_video_uncond_stg.float() + noise_pred_video_uncond_stg = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_stg, i, self.scheduler + ) + video_stg_delta = self.stg_scale * (noise_pred_video - noise_pred_video_uncond_stg) + else: + video_stg_delta = 0 + + # --- Modality isolation guidance forward pass --- + if self.do_modality_isolation_guidance: + with self.transformer.cache_context("uncond_modality"): + noise_pred_video_uncond_mod, noise_pred_audio_uncond_mod = self.transformer( + hidden_states=latents.to(dtype=connector_prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=connector_prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep_single, + audio_timestep=audio_timestep_single, + sigma=timestep_scalar_single, # Used by LTX-2.3 + audio_sigma=audio_timestep_single, + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + video_self_attention_mask=None, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + # Turn off A2V and V2A cross attn to isolate video and audio modalities + isolate_modalities=True, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_mod = noise_pred_video_uncond_mod.float() + noise_pred_video_uncond_mod = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_mod, i, self.scheduler + ) + video_modality_delta = (self.modality_scale - 1) * (noise_pred_video - noise_pred_video_uncond_mod) + else: + video_modality_delta = 0 + + noise_pred_video_g = noise_pred_video + video_cfg_delta + video_stg_delta + video_modality_delta + + if self.guidance_rescale > 0: + noise_pred_video = rescale_noise_cfg( + noise_pred_video_g, noise_pred_video, guidance_rescale=self.guidance_rescale + ) + else: + noise_pred_video = noise_pred_video_g + + # Apply the conditioning mask to apply the reference conditions at the specified strength. + if num_ref_tokens > 0: + bsz = noise_pred_video.size(0) + denoised_sample_cond = ( + noise_pred_video * (1 - conditioning_mask[:bsz]) + + clean_latents.float() * conditioning_mask[:bsz] + ).to(noise_pred_video.dtype) + noise_pred_video = denoised_sample_cond + + noise_pred_video = self.convert_x0_to_velocity(latents, noise_pred_video, i, self.scheduler) + + latents = self.scheduler.step(noise_pred_video, t, latents, return_dict=False)[0] + # Step the audio scheduler so its internal state stays in sync with the video scheduler (audio + # output is discarded at the end, but keeping schedulers aligned avoids surprising behavior if the + # scheduler writes internal indices during `.step()`). + _ = audio_scheduler.step(torch.zeros_like(audio_latents), t, audio_latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Decode + # Trim any appended reference tokens from the latents to recover the generated video only. + latents = latents[:, :base_token_count] + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + if output_type == "latent": + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + video = latents + else: + latents = latents.to(connector_prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = latents.to(self.vae.dtype) + + # VAE decode returns a video tensor in the VAE's native range ([-1, 1]). + decoded = self.vae.decode(latents, timestep, return_dict=False)[0] + # HDR postprocess: LogC3 decompress → linear HDR [0, ∞). Always float32 for HDR fidelity. + video = self.hdr_video_processor.postprocess_hdr_video(decoded, output_type=output_type) + + # Audio is always None for this video-only pipeline. + self.maybe_free_model_hooks() + + if not return_dict: + return (video, None) + + return LTX2PipelineOutput(frames=video, audio=None) diff --git a/diffusers/pipelines/ltx2/pipeline_ltx2_ic_lora.py b/diffusers/pipelines/ltx2/pipeline_ltx2_ic_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..8f2e3504e77788875b8dd7e3677a3515eb21c6d9 --- /dev/null +++ b/diffusers/pipelines/ltx2/pipeline_ltx2_ic_lora.py @@ -0,0 +1,2268 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import inspect +import math +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import Gemma3ForConditionalGeneration, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import FromSingleFileMixin, LTX2LoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTX2Audio, AutoencoderKLLTX2Video +from ...models.transformers import LTX2VideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .connectors import LTX2TextConnectors +from .pipeline_ltx2_condition import LTX2VideoCondition +from .pipeline_output import LTX2PipelineOutput +from .vocoder import LTX2Vocoder, LTX2VocoderWithBWE + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class LTX2ReferenceCondition: + """ + A reference video condition for IC-LoRA (In-Context LoRA) conditioning. + + The reference video is encoded into latent tokens and concatenated to the noisy latent sequence during denoising. + The transformer attends to these extra tokens, allowing the IC-LoRA adapter to condition the generation on the + reference video content (e.g. style, structure, depth, pose). + + Attributes: + frames (`PIL.Image.Image` or `List[PIL.Image.Image]` or `np.ndarray` or `torch.Tensor`): + The reference video frames. Accepts any type handled by `VideoProcessor.preprocess_video`. + strength (`float`, defaults to `1.0`): + Controls how "clean" the reference tokens appear to the model. A value of `1.0` means fully clean + (timestep=0 for reference tokens), `0.0` means fully noisy (same as denoising tokens). + """ + + frames: PIL.Image.Image | list[PIL.Image.Image] | np.ndarray | torch.Tensor + strength: float = 1.0 + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTX2InContextPipeline + >>> from diffusers.pipelines.ltx2.pipeline_ltx2_ic_lora import LTX2ReferenceCondition + >>> from diffusers.pipelines.ltx2.export_utils import encode_video + >>> from diffusers.pipelines.ltx2.utils import DEFAULT_NEGATIVE_PROMPT + >>> from diffusers.utils import load_video + + >>> pipe = LTX2InContextPipeline.from_pretrained("diffusers/LTX-2.3-Diffusers", torch_dtype=torch.bfloat16) + >>> pipe.enable_sequential_cpu_offload(device="cuda") + >>> pipe.load_lora_weights( + ... "Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-In", + ... adapter_name="ic_lora", + ... weight_name="ltx-2-19b-lora-camera-control-dolly-in.safetensors", + ... ) + >>> pipe.set_adapters("ic_lora", 1.0) + + >>> # If the IC LoRA uses reference conditions, you can specify them as follows: + >>> # reference_video = load_video("reference.mp4") + >>> # ref_cond = LTX2ReferenceCondition(frames=reference_video, strength=1.0) + + >>> prompt = "A flowing river in a forest" + >>> frame_rate = 24.0 + >>> video, audio = pipe( + ... prompt=prompt, + ... negative_prompt=DEFAULT_NEGATIVE_PROMPT, + ... # reference_conditions=[ref_cond], + ... width=768, + ... height=512, + ... num_frames=121, + ... frame_rate=frame_rate, + ... num_inference_steps=30, + ... guidance_scale=3.0, + ... output_type="np", + ... return_dict=False, + ... ) + + >>> encode_video( + ... video[0], + ... fps=frame_rate, + ... audio=audio[0].float().cpu(), + ... audio_sample_rate=pipe.vocoder.config.output_sampling_rate, + ... output_path="ic_lora_output.mp4", + ... ) + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTX2InContextPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraLoaderMixin): + r""" + Pipeline for LTX-2.X models with in-context (IC) conditioning. Also supports frame-level image conditions like + `LTX2ConditionPipeline`; both frame and reference conditions can be used together. + + In-context conditioning works by conditioning the generation on a reference video by encoding it into latent tokens + and concatenating them to the noisy latent tokens during denoising. The motivating use case is to support LTX-2.X + IC LoRAs, which may use reference conditions (e.g. a pose video for pose control) to guide generation, but this + pipeline is designed to work with any LTX-2.X-like model trained with in-context reference conditions. + + Two-stage inference is supported through separate calls to `__call__`: + - **Stage 1**: Generate at target resolution with IC-LoRA conditioning (`output_type="latent"`). + - **Stage 2**: Upsample via [`LTX2LatentUpsamplePipeline`], then refine with a distilled LoRA (no IC-LoRA reference + conditioning needed for Stage 2). + + Reference: https://github.com/Lightricks/LTX-Video + + Args: + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLLTX2Video`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + audio_vae ([`AutoencoderKLLTX2Audio`]): + Audio VAE to encode and decode audio spectrograms. + text_encoder ([`Gemma3ForConditionalGeneration`]): + Text encoder model. + tokenizer (`GemmaTokenizer` or `GemmaTokenizerFast`): + Tokenizer for the text encoder. + connectors ([`LTX2TextConnectors`]): + Text connector stack used to adapt text encoder hidden states for the video and audio branches. + transformer ([`LTX2VideoTransformer3DModel`]): + Conditional Transformer architecture to denoise the encoded video latents. + vocoder ([`LTX2Vocoder`] or [`LTX2VocoderWithBWE`]): + Vocoder to convert mel spectrograms to audio waveforms. + """ + + model_cpu_offload_seq = "text_encoder->connectors->transformer->vae->audio_vae->vocoder" + _optional_components = ["audio_scheduler"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTX2Video, + audio_vae: AutoencoderKLLTX2Audio, + text_encoder: Gemma3ForConditionalGeneration, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + connectors: LTX2TextConnectors, + transformer: LTX2VideoTransformer3DModel, + vocoder: LTX2Vocoder | LTX2VocoderWithBWE, + audio_scheduler: FlowMatchEulerDiscreteScheduler | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + audio_vae=audio_vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + connectors=connectors, + transformer=transformer, + vocoder=vocoder, + scheduler=scheduler, + audio_scheduler=audio_scheduler, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.audio_vae_mel_compression_ratio = ( + self.audio_vae.mel_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.audio_vae_temporal_compression_ratio = ( + self.audio_vae.temporal_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.audio_sampling_rate = ( + self.audio_vae.config.sample_rate if getattr(self, "audio_vae", None) is not None else 16000 + ) + self.audio_hop_length = ( + self.audio_vae.config.mel_hop_length if getattr(self, "audio_vae", None) is not None else 160 + ) + self.audio_mel_bins = self.audio_vae.config.mel_bins if getattr(self, "audio_vae", None) is not None else 64 + self.audio_latent_channels = ( + self.audio_vae.config.latent_channels if getattr(self, "audio_vae", None) is not None else 8 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio, resample="bilinear") + + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 1024 + ) + tokenizer_padding_side = "left" + if getattr(self, "tokenizer", None) is not None: + tokenizer_padding_side = getattr(self.tokenizer, "padding_side", "left") + self.tokenizer_padding_side = tokenizer_padding_side + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`str` or `torch.device`): + torch device to place the resulting embeddings on + dtype: (`torch.dtype`): + torch dtype to cast the prompt embeds to + max_sequence_length (`int`, defaults to 1024): Maximum sequence length to use for the prompt. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if getattr(self, "tokenizer", None) is not None: + # Gemma expects left padding for chat-style prompts + self.tokenizer.padding_side = "left" + if self.tokenizer.pad_token is None: + self.tokenizer.pad_token = self.tokenizer.eos_token + + prompt = [p.strip() for p in prompt] + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + text_input_ids = text_input_ids.to(device) + prompt_attention_mask = prompt_attention_mask.to(device) + + text_encoder_outputs = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ) + text_encoder_hidden_states = text_encoder_outputs.hidden_states + text_encoder_hidden_states = torch.stack(text_encoder_hidden_states, dim=-1) + prompt_embeds = text_encoder_hidden_states.flatten(2, 3).to(dtype=dtype) # Pack to 3D + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + latents=None, + audio_latents=None, + spatio_temporal_guidance_blocks=None, + stg_scale=None, + audio_stg_scale=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + if latents is not None and latents.ndim != 5: + raise ValueError( + f"Only unpacked (5D) video latents of shape `[batch_size, latent_channels, latent_frames," + f" latent_height, latent_width] are supported, but got {latents.ndim} dims. If you have packed (3D)" + f" latents, please unpack them (e.g. using the `_unpack_latents` method)." + ) + if audio_latents is not None and audio_latents.ndim != 4: + raise ValueError( + f"Only unpacked (4D) audio latents of shape `[batch_size, num_channels, audio_length, mel_bins] are" + f" supported, but got {audio_latents.ndim} dims. If you have packed (3D) latents, please unpack them" + f" (e.g. using the `_unpack_audio_latents` method)." + ) + + if ((stg_scale > 0.0) or (audio_stg_scale > 0.0)) and not spatio_temporal_guidance_blocks: + raise ValueError( + "Spatio-Temporal Guidance (STG) is specified but no STG blocks are supplied. Please supply a list of" + "block indices at which to apply STG in `spatio_temporal_guidance_blocks`" + ) + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_latents + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_image2video.LTX2ImageToVideoPipeline._normalize_latents + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._create_noised_state + def _create_noised_state( + latents: torch.Tensor, noise_scale: float | torch.Tensor, generator: torch.Generator | None = None + ): + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + noised_latents = noise_scale * noise + (1 - noise_scale) * latents + return noised_latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._normalize_audio_latents + def _normalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents - latents_mean) / latents_std + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_audio_latents + def _denormalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents * latents_std) + latents_mean + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_audio_latents + def _pack_audio_latents( + latents: torch.Tensor, patch_size: int | None = None, patch_size_t: int | None = None + ) -> torch.Tensor: + # Audio latents shape: [B, C, L, M], where L is the latent audio length and M is the number of mel bins + if patch_size is not None and patch_size_t is not None: + # Packs the latents into a patch sequence of shape [B, L // p_t * M // p, C * p_t * p] (a ndim=3 tnesor). + # dim=1 is the effective audio sequence length and dim=2 is the effective audio input feature size. + batch_size, num_channels, latent_length, latent_mel_bins = latents.shape + post_patch_latent_length = latent_length / patch_size_t + post_patch_mel_bins = latent_mel_bins / patch_size + latents = latents.reshape( + batch_size, -1, post_patch_latent_length, patch_size_t, post_patch_mel_bins, patch_size + ) + latents = latents.permute(0, 2, 4, 1, 3, 5).flatten(3, 5).flatten(1, 2) + else: + # Packs the latents into a patch sequence of shape [B, L, C * M]. This implicitly assumes a (mel) + # patch_size of M (all mel bins constitutes a single patch) and a patch_size_t of 1. + latents = latents.transpose(1, 2).flatten(2, 3) # [B, C, L, M] --> [B, L, C * M] + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_audio_latents + def _unpack_audio_latents( + latents: torch.Tensor, + latent_length: int, + num_mel_bins: int, + patch_size: int | None = None, + patch_size_t: int | None = None, + ) -> torch.Tensor: + # Unpacks an audio patch sequence of shape [B, S, D] into a latent spectrogram tensor of shape [B, C, L, M], + # where L is the latent audio length and M is the number of mel bins. + if patch_size is not None and patch_size_t is not None: + batch_size = latents.size(0) + latents = latents.reshape(batch_size, latent_length, num_mel_bins, -1, patch_size_t, patch_size) + latents = latents.permute(0, 3, 1, 4, 2, 5).flatten(4, 5).flatten(2, 3) + else: + # Assume [B, S, D] = [B, L, C * M], which implies that patch_size = M and patch_size_t = 1. + latents = latents.unflatten(2, (-1, num_mel_bins)).transpose(1, 2) + return latents + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.trim_conditioning_sequence + def trim_conditioning_sequence(self, start_frame: int, sequence_num_frames: int, target_num_frames: int) -> int: + """ + Trim a conditioning sequence to the allowed number of frames. + + Args: + start_frame (int): The target frame number of the first frame in the sequence. + sequence_num_frames (int): The number of frames in the sequence. + target_num_frames (int): The target number of frames in the generated video. + Returns: + int: updated sequence length + """ + scale_factor = self.vae_temporal_compression_ratio + num_frames = min(sequence_num_frames, target_num_frames - start_frame) + # Trim down to a multiple of temporal_scale_factor frames plus 1 + num_frames = (num_frames - 1) // scale_factor * scale_factor + 1 + return num_frames + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.preprocess_conditions + def preprocess_conditions( + self, + conditions: LTX2VideoCondition | list[LTX2VideoCondition] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + device: torch.device | None = None, + ) -> tuple[list[torch.Tensor], list[float], list[int], list[int]]: + """ + Preprocesses the condition images/videos to torch tensors. + + Args: + conditions (`LTX2VideoCondition` or `List[LTX2VideoCondition]`, *optional*, defaults to `None`): + A list of image/video condition instances. + height (`int`, *optional*, defaults to `512`): + The desired height in pixels. + width (`int`, *optional*, defaults to `768`): + The desired width in pixels. + num_frames (`int`, *optional*, defaults to `121`): + The desired number of frames in the generated video. + device (`torch.device`, *optional*, defaults to `None`): + The device on which to put the preprocessed image/video tensors. + + Returns: + `Tuple[List[torch.Tensor], List[float], List[int], List[int]]`: + Returns a 4-tuple of lists of length `len(conditions)` as follows: + 1. The first list is a list of preprocessed video tensors of shape [batch_size=1, num_channels, + num_frames, height, width]. + 2. The second list is a list of conditioning strengths. + 3. The third list is a list of latent-space indices for each condition. + 4. The fourth list is a list of (trimmed) pixel-space frame counts per condition. This is needed + for keyframe coord semantics (single-pixel-frame keyframes have a clamped temporal extent). + """ + conditioning_frames, conditioning_strengths, conditioning_indices, conditioning_pixel_frames = [], [], [], [] + + if conditions is None: + conditions = [] + if isinstance(conditions, LTX2VideoCondition): + conditions = [conditions] + + frame_scale_factor = self.vae_temporal_compression_ratio + latent_num_frames = (num_frames - 1) // frame_scale_factor + 1 + for i, condition in enumerate(conditions): + # Create a channels-last video-like array of shape (F, H, W, C) in preparation for resizing. + if isinstance(condition.frames, PIL.Image.Image): + arr = np.array(condition.frames.convert("RGB"))[None] # (1, H, W, 3) + elif isinstance(condition.frames, list) and all(isinstance(f, PIL.Image.Image) for f in condition.frames): + arr = np.stack([np.array(f.convert("RGB")) for f in condition.frames]) # (F, H, W, 3) + elif isinstance(condition.frames, np.ndarray): + arr = condition.frames if condition.frames.ndim == 4 else condition.frames[None] + elif isinstance(condition.frames, torch.Tensor): + t = condition.frames if condition.frames.ndim == 4 else condition.frames.unsqueeze(0) + # Reference layout for video tensors is (F, C, H, W); convert to (F, H, W, C) for the + # resize logic, which expects channels-last. + arr = t.detach().cpu().permute(0, 2, 3, 1).numpy() + else: + raise TypeError(f"Unsupported `frames` type for condition {i}: {type(condition.frames)}") + + src_h, src_w = arr.shape[1], arr.shape[2] + num_cond_frames = arr.shape[0] + # Convert the NumPy array to a channels-first tensor of shape (1, C, F, H, W) + pixels = torch.from_numpy(np.ascontiguousarray(arr)).to(torch.float32) + pixels = pixels.permute(3, 0, 1, 2).unsqueeze(0).to(device) # (1, C, F, H, W) + + # Resize so the longer side fills the target, then center-crop to exact (height, width). + scale = max(height / src_h, width / src_w) + new_h = math.ceil(src_h * scale) + new_w = math.ceil(src_w * scale) + # Flatten (B, C, F, H, W) → (B*F, C, H, W) for the per-frame interpolation + pixels = pixels.permute(0, 2, 1, 3, 4).reshape(num_cond_frames, 3, src_h, src_w) + # NOTE: we avoid using VideoProcessor.preprocess_video here because it uses PIL.Image.resize under the + # hood, which will apply an anti-aliasing pre-filter when downsampling. The original LTX-2.X code simply + # uses F.interpolate, which is reproduced here. + pixels = torch.nn.functional.interpolate(pixels, size=(new_h, new_w), mode="bilinear", align_corners=False) + top = (new_h - height) // 2 + left = (new_w - width) // 2 + pixels = pixels[:, :, top : top + height, left : left + width] + pixels = pixels.reshape(1, num_cond_frames, 3, height, width).permute(0, 2, 1, 3, 4) + + # Map [0, 255] → [-1, 1] (VAE input convention). + condition_pixels = pixels / 127.5 - 1.0 + + # Interpret the index as a latent index, following the original LTX-2 code. + latent_start_idx = condition.index + # Support negative latent indices (e.g. -1 for the last latent index) + if latent_start_idx < 0: + # latent_start_idx will be positive because latent_num_frames is positive + latent_start_idx = latent_start_idx % latent_num_frames + if latent_start_idx >= latent_num_frames: + logger.warning( + f"The starting latent index {latent_start_idx} of condition {i} is too big for the specified number" + f" of latent frames {latent_num_frames}. This condition will be skipped." + ) + continue + + cond_num_frames = condition_pixels.size(2) + start_idx = max((latent_start_idx - 1) * frame_scale_factor + 1, 0) + truncated_cond_frames = self.trim_conditioning_sequence(start_idx, cond_num_frames, num_frames) + condition_pixels = condition_pixels[:, :, :truncated_cond_frames] + + conditioning_frames.append(condition_pixels.to(dtype=self.vae.dtype, device=device)) + conditioning_strengths.append(condition.strength) + conditioning_indices.append(latent_start_idx) + conditioning_pixel_frames.append(truncated_cond_frames) + + return conditioning_frames, conditioning_strengths, conditioning_indices, conditioning_pixel_frames + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.apply_first_frame_conditioning + def apply_first_frame_conditioning( + self, + latents: torch.Tensor, + conditioning_mask: torch.Tensor, + condition_latents: list[torch.Tensor], + condition_strengths: list[float], + condition_indices: list[int], + latent_height: int, + latent_width: int, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Apply first-frame visual conditioning by overwriting tokens at the first-frame positions. + + Only conditions with `latent_idx == 0` are applied here (matching `VideoConditionByLatentIndex` in the + reference implementation). Conditions at non-zero latent indices are appended as separate keyframe tokens via + `prepare_keyframe_extras` (matching `VideoConditionByKeyframeIndex`) and are skipped here. + + Args: + latents (`torch.Tensor`): + Initial packed (patchified) latents of shape [batch_size, patch_seq_len, hidden_dim]. + conditioning_mask (`torch.Tensor`): + Initial packed (patchified) conditioning mask of shape [batch_size, patch_seq_len, 1] with values in + [0, 1] where 0 means the denoising model output will be fully used and 1 means the condition will be + fully used. + + Returns: + `Tuple[torch.Tensor, torch.Tensor, torch.Tensor]`: + Returns a 3-tuple of tensors where: + 1. The packed video latents with first-frame conditions applied. + 2. The packed conditioning mask with first-frame strengths applied. + 3. The clean conditioning latents at first-frame positions (zeros elsewhere). + """ + clean_latents = torch.zeros_like(latents) + for cond, strength, latent_idx in zip(condition_latents, condition_strengths, condition_indices): + if latent_idx != 0: + # Non-first-frame conditions are handled as keyframe extras (appended tokens) instead. + continue + num_cond_tokens = cond.size(1) + start_token_idx = latent_idx * latent_height * latent_width + end_token_idx = start_token_idx + num_cond_tokens + + latents[:, start_token_idx:end_token_idx] = cond + conditioning_mask[:, start_token_idx:end_token_idx] = strength + clean_latents[:, start_token_idx:end_token_idx] = cond + + return latents, conditioning_mask, clean_latents + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline._prepare_keyframe_coords + def _prepare_keyframe_coords( + self, + keyframe_latent_num_frames: int, + keyframe_latent_height: int, + keyframe_latent_width: int, + pixel_frame_idx: int, + num_pixel_frames: int, + fps: float, + device: torch.device, + ) -> torch.Tensor: + """ + Compute positional coordinates for a keyframe condition being appended as extra tokens. + + Mirrors `VideoConditionByKeyframeIndex.apply_to` in the reference implementation: + - Latent coords scaled to pixel space *without* the causal fix (since non-zero-index keyframes don't need the + first-frame causal adjustment). + - Temporal axis offset by `pixel_frame_idx` (the pixel-space index at which the keyframe appears). + - For single-pixel-frame keyframes, the per-patch temporal extent is clamped to `[idx, idx + 1)` so the + keyframe occupies a single pixel timestep rather than the VAE-scaled range. + - Temporal coords divided by `fps` to produce seconds. + """ + patch_size = self.transformer_spatial_patch_size + patch_size_t = self.transformer_temporal_patch_size + scale_factors = ( + self.vae_temporal_compression_ratio, + self.vae_spatial_compression_ratio, + self.vae_spatial_compression_ratio, + ) + + grid_f = torch.arange( + start=0, end=keyframe_latent_num_frames, step=patch_size_t, dtype=torch.float32, device=device + ) + grid_h = torch.arange(start=0, end=keyframe_latent_height, step=patch_size, dtype=torch.float32, device=device) + grid_w = torch.arange(start=0, end=keyframe_latent_width, step=patch_size, dtype=torch.float32, device=device) + grid = torch.meshgrid(grid_f, grid_h, grid_w, indexing="ij") + grid = torch.stack(grid, dim=0) + + patch_size_delta = torch.tensor((patch_size_t, patch_size, patch_size), dtype=grid.dtype, device=device) + patch_ends = grid + patch_size_delta.view(3, 1, 1, 1) + + latent_coords = torch.stack([grid, patch_ends], dim=-1) # [3, N_F, N_H, N_W, 2] + latent_coords = latent_coords.flatten(1, 3) # [3, num_patches, 2] + latent_coords = latent_coords.unsqueeze(0) # [1, 3, num_patches, 2] + + scale_tensor = torch.tensor(scale_factors, device=device, dtype=latent_coords.dtype) + broadcast_shape = [1] * latent_coords.ndim + broadcast_shape[1] = -1 + pixel_coords = latent_coords * scale_tensor.view(*broadcast_shape) + + # No causal fix: keyframe coords place the keyframe at `pixel_frame_idx` without the first-frame adjustment. + pixel_coords[:, 0, :, :] = pixel_coords[:, 0, :, :] + pixel_frame_idx + + if num_pixel_frames == 1: + # Single-pixel-frame keyframe: clamp temporal extent to [idx, idx + 1). + pixel_coords[:, 0, :, 1:] = pixel_coords[:, 0, :, :1] + 1 + + pixel_coords[:, 0, :, :] = pixel_coords[:, 0, :, :] / fps + + return pixel_coords + + def prepare_latents( + self, + conditions: LTX2VideoCondition | list[LTX2VideoCondition] | None = None, + reference_conditions: list[LTX2ReferenceCondition] | None = None, + reference_downscale_factor: int = 1, + conditioning_attention_strength: float = 1.0, + conditioning_attention_mask: torch.Tensor | None = None, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + noise_scale: float = 1.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor | None, int, torch.Tensor | None]: + """ + Prepare noisy video latents, applying frame and reference-video conditioning. + + Conditioning sources are unified into a single packed sequence in the order `[base | keyframe | reference]`: + + - First-frame conditions (`conditions` with `latent_idx == 0`) overwrite tokens at the first-frame positions + (`VideoConditionByLatentIndex` semantics). + - Non-first-frame conditions (`conditions` with `latent_idx > 0`) are concatenated onto the main latent + sequence with per-token `conditioning_mask = strength` (`VideoConditionByKeyframeIndex` semantics). + - IC-LoRA `reference_conditions` (if any) are encoded by the VAE and appended after the keyframes with + per-token `conditioning_mask = strength` (matching the reference repo's `VideoConditionByReferenceLatent` + semantics). + + For all appended tokens the noise mixing below blends them to noise level `(1 - strength) * sigma_max`, and the + existing per-token timestep formula `t * (1 - conditioning_mask)` and the post-process blend `denoised * (1 - + cond_mask) + clean * cond_mask` drive them through the loop. + + Returns a 6-tuple: + - `latents`: packed noisy latents `(B, base + n_keyframe + n_ref, C)`. + - `conditioning_mask`: `(B, seq_len, 1)` with values in `[0, 1]` — `1` at first-frame positions, `strength` + at keyframe / reference positions, `0` elsewhere. + - `clean_latents`: clean condition values at conditioned positions (zeros elsewhere); same shape as + `latents`. + - `appended_coords`: `[1, 3, n_keyframe + n_ref, 2]` positional coordinates to concat onto `video_coords`, + or `None` if no keyframe/reference conditions are provided. + - `num_ref_tokens`: count of reference tokens at the END of `latents` (used by the call site to build the + unified self-attention mask). + - `ref_cross_mask`: `[1, num_ref_tokens]` per-reference-token cross-attention strengths in `[0, 1]`, or + `None` when `conditioning_attention_strength == 1.0` and no pixel-space mask is provided (in which case + attention is uniform). + """ + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + + shape = (batch_size, num_channels_latents, latent_num_frames, latent_height, latent_width) + mask_shape = (batch_size, 1, latent_num_frames, latent_height, latent_width) + + if latents is not None: + # Latents are expected to be unpacked (5D) with shape [B, F, C, H, W] + latents = self._normalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + else: + # NOTE: we set the initial latents to zeros rather a sample from the standard Gaussian prior because we + # will sample from the prior later once we have calculated the conditioning mask + latents = torch.zeros(shape, device=device, dtype=dtype) + + conditioning_mask = latents.new_zeros(mask_shape) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + conditioning_mask = self._pack_latents( + conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) # [B, seq_len, 1] + + if latents.ndim != 3 or latents.shape[:2] != conditioning_mask.shape[:2]: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is {conditioning_mask.shape[:2] + (num_channels_latents,)}." + ) + + if isinstance(generator, list): + logger.warning( + f"{self.__class__.__name__} does not support using a list of generators. The first generator in the" + f" list will be used for all (pseudo-)random operations." + ) + + condition_frames, condition_strengths, condition_indices, condition_pixel_frames = self.preprocess_conditions( + conditions, height, width, num_frames, device=device + ) + # Encode each condition through the VAE. We keep both the 5D latent (for coord computation) and the packed + # 3D latent (for first-frame replacement or keyframe append). + condition_latents_5d = [] + condition_latents_packed = [] + for condition_tensor in condition_frames: + condition_latent_5d = retrieve_latents( + self.vae.encode(condition_tensor), + generator=generator[0] if isinstance(generator, list) else generator, + sample_mode="argmax", + ) + condition_latent_5d = self._normalize_latents( + condition_latent_5d, self.vae.latents_mean, self.vae.latents_std + ).to(device=device, dtype=dtype) + condition_latent_packed = self._pack_latents( + condition_latent_5d, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + condition_latents_5d.append(condition_latent_5d) + condition_latents_packed.append(condition_latent_packed) + + # First-frame conditions (latent_idx == 0): replace tokens at the first-frame positions. + # NOTE: following the I2V pipeline, we return a conditioning mask. The original LTX 2 code uses a denoising + # mask, which is the inverse of the conditioning mask (`denoise_mask = 1 - conditioning_mask`). + latents, conditioning_mask, clean_latents = self.apply_first_frame_conditioning( + latents, + conditioning_mask, + condition_latents_packed, + condition_strengths, + condition_indices, + latent_height=latent_height, + latent_width=latent_width, + ) + + # Non-first-frame ("keyframe") conditions (latent_idx > 0): append as extra latent tokens to the noisy latent. + # Each condition gets a all-`strength` conditioning mask and pos ids, which are also appended to those of the + # noisy latent. At each denoising step i, the keyframe conditions get an effective noise level of + # (1 - conditioning_strength) * sigma_i. + frame_scale_factor = self.vae_temporal_compression_ratio + kf_tokens_list, kf_coords_list, kf_mask_list, kf_clean_list = [], [], [], [] + for cond_5d, cond_packed, strength, latent_idx, num_pixel_frames in zip( + condition_latents_5d, + condition_latents_packed, + condition_strengths, + condition_indices, + condition_pixel_frames, + ): + if latent_idx == 0: + continue + + _, _, kf_latent_frames, kf_latent_height, kf_latent_width = cond_5d.shape + pixel_frame_idx = (latent_idx - 1) * frame_scale_factor + 1 + + coords = self._prepare_keyframe_coords( + keyframe_latent_num_frames=kf_latent_frames, + keyframe_latent_height=kf_latent_height, + keyframe_latent_width=kf_latent_width, + pixel_frame_idx=pixel_frame_idx, + num_pixel_frames=num_pixel_frames, + fps=frame_rate, + device=device, + ) + + num_tokens = cond_packed.shape[1] + kf_mask = torch.full( + (cond_packed.shape[0], num_tokens, 1), + float(strength), + device=device, + dtype=conditioning_mask.dtype, + ) + + kf_tokens_list.append(cond_packed) + kf_clean_list.append(cond_packed) + kf_mask_list.append(kf_mask) + kf_coords_list.append(coords) + + if kf_tokens_list: + keyframe_coords = torch.cat(kf_coords_list, dim=2) + latents = torch.cat([latents, torch.cat(kf_tokens_list, dim=1)], dim=1) + conditioning_mask = torch.cat([conditioning_mask, torch.cat(kf_mask_list, dim=1)], dim=1) + clean_latents = torch.cat([clean_latents, torch.cat(kf_clean_list, dim=1)], dim=1) + else: + keyframe_coords = None + + # IC-LoRA reference-video conditions: encode each reference video, then append it to the main packed + # sequence with per-token `conditioning_mask = strength`. This is the same architectural pattern as + # for non-first-frame conditions above, but we need to keep keyframe and reference conditions separate + # for attention masking. + ref_cross_mask: torch.Tensor | None = None + ref_coords: torch.Tensor | None = None + num_ref_tokens = 0 + if reference_conditions is not None and len(reference_conditions) > 0: + ref_latents_packed, ref_coords, ref_cross_mask = self._encode_reference_conditions( + reference_conditions=reference_conditions, + num_frames=num_frames, + height=height, + width=width, + reference_downscale_factor=reference_downscale_factor, + frame_rate=frame_rate, + conditioning_attention_strength=conditioning_attention_strength, + conditioning_attention_mask=conditioning_attention_mask, + dtype=dtype, + device=device, + generator=generator[0] if isinstance(generator, list) else generator, + ) + num_ref_tokens = ref_latents_packed.shape[1] + + # All reference videos preprocess to the same (ref_height, ref_width, num_frames), so their packed + # token counts are identical. Split `num_ref_tokens` evenly across the conditions and materialize + # the per-token strength mask in `reference_conditions` order, matching the layout the encoder + # emitted. + n_per_ref = num_ref_tokens // len(reference_conditions) + ref_mask_chunks = [ + torch.full( + (batch_size, n_per_ref, 1), + float(ref_cond.strength), + device=device, + dtype=conditioning_mask.dtype, + ) + for ref_cond in reference_conditions + ] + ref_mask_full = torch.cat(ref_mask_chunks, dim=1) + + ref_latents_packed_b = ref_latents_packed.expand(batch_size, -1, -1) + latents = torch.cat([latents, ref_latents_packed_b], dim=1) + conditioning_mask = torch.cat([conditioning_mask, ref_mask_full], dim=1) + clean_latents = torch.cat([clean_latents, ref_latents_packed_b], dim=1) + + # Combine keyframe + reference appended-coords into a single block to concat onto `video_coords` at + # the call site. + if keyframe_coords is not None and ref_coords is not None: + appended_coords = torch.cat([keyframe_coords, ref_coords], dim=2) + elif keyframe_coords is not None: + appended_coords = keyframe_coords + elif ref_coords is not None: + appended_coords = ref_coords + else: + appended_coords = None + + # The conditioning_mask values have the following semantics: + # - mask=0: fully noise tokens (e.g. noisy latents) + # - mask=1: keep fully clean (e.g. I2V first-frame condition, conditions with strength=1) + # - mask in (0, 1): use intermediate noise level mask * sigma_i (noise_scale == sigma_0) + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + scaled_mask = (1.0 - conditioning_mask) * noise_scale # noise to initial noise level `noise_scale` + latents = noise * scaled_mask + latents * (1 - scaled_mask) + + return latents, conditioning_mask, clean_latents, appended_coords, num_ref_tokens, ref_cross_mask + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.prepare_audio_latents + def prepare_audio_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 8, + audio_latent_length: int = 1, # 1 is just a dummy value + num_mel_bins: int = 64, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + # latents expected to be unpacked (4D) with shape [B, C, L, M] + latents = self._pack_audio_latents(latents) + latents = self._normalize_audio_latents(latents, self.audio_vae.latents_mean, self.audio_vae.latents_std) + latents = self._create_noised_state(latents, noise_scale, generator) + return latents.to(device=device, dtype=dtype) + + latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Sample in packed shape (B, L, C * M), following the original LTX-2.X code + packed_shape = (batch_size, audio_latent_length, num_channels_latents * latent_mel_bins) + latents = randn_tensor(packed_shape, generator=generator, device=device, dtype=dtype) + return latents + + def _encode_reference_conditions( + self, + reference_conditions: list[LTX2ReferenceCondition], + height: int, + width: int, + num_frames: int, + reference_downscale_factor: int = 1, + frame_rate: float = 24.0, + conditioning_attention_strength: float = 1.0, + conditioning_attention_mask: torch.Tensor | None = None, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]: + """Encode IC-LoRA reference videos into `(reference_latents, reference_coords, reference_cross_mask)`. + + This is the shared encoding core used by both `prepare_latents` (which folds reference tokens into the main + noisy sequence) and the back-compat shim `prepare_reference_latents` (which exposes the legacy 4-tuple output). + See `prepare_reference_latents` for parameter documentation. + """ + ref_height = height // reference_downscale_factor + ref_width = width // reference_downscale_factor + + mask_needed = conditioning_attention_strength < 1.0 or conditioning_attention_mask is not None + + all_ref_latents = [] + all_ref_coords = [] + all_ref_cross_masks = [] + + for ref_cond in reference_conditions: + # Preprocess reference video frames to the (possibly downscaled) resolution + if isinstance(ref_cond.frames, PIL.Image.Image): + video_like = [ref_cond.frames] + elif isinstance(ref_cond.frames, np.ndarray) and ref_cond.frames.ndim == 3: + video_like = np.expand_dims(ref_cond.frames, axis=0) + elif isinstance(ref_cond.frames, torch.Tensor) and ref_cond.frames.ndim == 3: + video_like = ref_cond.frames.unsqueeze(0) + else: + video_like = ref_cond.frames + + ref_pixels = self.video_processor.preprocess_video(video_like, ref_height, ref_width, resize_mode="crop") + # Trim to num_frames + ref_pixels = ref_pixels[:, :, :num_frames] + ref_pixels = ref_pixels.to(dtype=self.vae.dtype, device=device) + + # Encode through VAE + ref_latent = retrieve_latents(self.vae.encode(ref_pixels), generator=generator, sample_mode="argmax") + ref_latent = self._normalize_latents(ref_latent, self.vae.latents_mean, self.vae.latents_std).to( + device=device, dtype=dtype + ) + + # Get latent dimensions for coordinate computation + _, _, ref_latent_frames, ref_latent_height, ref_latent_width = ref_latent.shape + + # Pack into tokens + ref_latent_packed = self._pack_latents( + ref_latent, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + + # Compute positional coordinates for the reference tokens. We use the transformer's + # prepare_video_coords at the reference video's latent dimensions, then scale spatial coords + # by downscale_factor so they map to the target coordinate space. + ref_coords = self.transformer.rope.prepare_video_coords( + batch_size=1, + num_frames=ref_latent_frames, + height=ref_latent_height, + width=ref_latent_width, + device=device, + fps=frame_rate, + ) + if reference_downscale_factor != 1: + # Scale spatial coordinates (height=axis 1, width=axis 2) to match target space + ref_coords[:, 1, :, :] = ref_coords[:, 1, :, :] * reference_downscale_factor + ref_coords[:, 2, :, :] = ref_coords[:, 2, :, :] * reference_downscale_factor + + num_tokens = ref_latent_packed.shape[1] + + all_ref_latents.append(ref_latent_packed) + all_ref_coords.append(ref_coords) + + if mask_needed: + # Per-reference cross-attention mask. Start from either a downsampled pixel-space mask or a full-1 + # tensor, then scale by conditioning_attention_strength. + if conditioning_attention_mask is not None: + ref_cross = self._downsample_mask_to_latent( + mask=conditioning_attention_mask, + latent_num_frames=ref_latent_frames, + latent_height=ref_latent_height, + latent_width=ref_latent_width, + ).to(device=device, dtype=torch.float32) + else: + ref_cross = torch.ones((1, num_tokens), device=device, dtype=torch.float32) + ref_cross = ref_cross * conditioning_attention_strength + all_ref_cross_masks.append(ref_cross) + + # Concatenate all reference tokens into a single sequence + reference_latents = torch.cat(all_ref_latents, dim=1) # [1, total_ref_tokens, D] + reference_coords = torch.cat(all_ref_coords, dim=2) # [1, 3, total_ref_tokens, 2] + reference_cross_mask = torch.cat(all_ref_cross_masks, dim=1) if mask_needed else None + + return reference_latents, reference_coords, reference_cross_mask + + def prepare_reference_latents( + self, + reference_conditions: list[LTX2ReferenceCondition], + height: int, + width: int, + num_frames: int, + reference_downscale_factor: int = 1, + frame_rate: float = 24.0, + conditioning_attention_strength: float = 1.0, + conditioning_attention_mask: torch.Tensor | None = None, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor | None]: + """ + Encode reference videos into packed latent tokens and compute their positional coordinates. + + Each reference video is independently encoded by the VAE, packed into tokens, and its positional coordinates + are computed with spatial scaling by `reference_downscale_factor` to match the target coordinate space. + + All reference tokens are concatenated into a single sequence. When `conditioning_attention_strength < 1.0` or + `conditioning_attention_mask` is provided, a per-token cross-attention mask is also computed for each reference + video (downsampled to the reference video's latent dimensions) and returned so callers can build a + self-attention mask over the full video sequence. + + Args: + reference_conditions (`list[LTX2ReferenceCondition]`): + The reference video conditions. + height (`int`): + Target video height in pixels (used to determine reference video preprocessing size with + `reference_downscale_factor`). + width (`int`): + Target video width in pixels. + num_frames (`int`): + Number of target video frames. + reference_downscale_factor (`int`, defaults to `1`): + Ratio between target and reference resolutions. A factor of 2 means the reference video is preprocessed + at half the target resolution. Spatial positional coordinates are scaled by this factor to map + reference tokens into the target coordinate space. + frame_rate (`float`, defaults to `24.0`): + Video frame rate (used for temporal coordinate computation). + conditioning_attention_strength (`float`, defaults to `1.0`): + Scalar in `[0, 1]` controlling how strongly reference tokens attend to noisy tokens (and vice versa) in + the self-attention mask. `1.0` means full attention (no masking), `0.0` means reference tokens are + effectively ignored by the noisy tokens. + conditioning_attention_mask (`torch.Tensor`, *optional*): + Optional pixel-space mask of shape `(1, 1, F_pix, H_pix, W_pix)` with values in `[0, 1]` that provides + spatially-varying attention strength. Downsampled to latent space per reference video and multiplied by + `conditioning_attention_strength`. + dtype (`torch.dtype`, *optional*): + Data type for the latents. + device (`torch.device`, *optional*): + Device for the latents. + generator (`torch.Generator`, *optional*): + Random generator for VAE encoding. + + Returns: + A 4-tuple of `(reference_latents, reference_coords, reference_denoise_factors, reference_cross_mask)`: + - `reference_latents`: `[1, total_ref_tokens, hidden_dim]` + - `reference_coords`: `[1, 3, total_ref_tokens, 2]` + - `reference_denoise_factors`: `[1, total_ref_tokens]` — per-token `(1 - strength)` factors + - `reference_cross_mask`: `[1, total_ref_tokens]` per-token noisy↔reference attention strengths in `[0, + 1]`, or `None` when `conditioning_attention_strength == 1.0` and no pixel-space mask is provided (in + which case attention is unmasked). + """ + reference_latents, reference_coords, reference_cross_mask = self._encode_reference_conditions( + reference_conditions=reference_conditions, + height=height, + width=width, + num_frames=num_frames, + reference_downscale_factor=reference_downscale_factor, + frame_rate=frame_rate, + conditioning_attention_strength=conditioning_attention_strength, + conditioning_attention_mask=conditioning_attention_mask, + dtype=dtype, + device=device, + generator=generator, + ) + + # Materialize per-token denoise factors for callers that still expect the 4-tuple. Each ref video has + # `1 - strength` for all of its tokens; we rebuild this from the per-video token counts which we can + # back out from `reference_latents.shape[1]` and the input `reference_conditions` order. + ref_denoise_chunks: list[torch.Tensor] = [] + idx = 0 + # Walk the encoded ref tokens video-by-video. Each ref's token count is fixed by the ref video's latent + # shape, which equals (num_frames -> ref_latent_frames) * ref_latent_h * ref_latent_w. Computing it here + # would duplicate the encoding math; instead we rely on the shape match across all refs being identical + # (same `num_frames`, same downscaled height/width) so we can split equally. + n_total = reference_latents.shape[1] + n_per_ref = n_total // max(len(reference_conditions), 1) + for ref_cond in reference_conditions: + ref_denoise_chunks.append( + torch.full( + (1, n_per_ref), 1.0 - ref_cond.strength, device=reference_latents.device, dtype=torch.float32 + ) + ) + idx += n_per_ref + reference_denoise_factors = ( + torch.cat(ref_denoise_chunks, dim=1) if ref_denoise_chunks else reference_latents.new_zeros((1, 0)) + ) + + return reference_latents, reference_coords, reference_denoise_factors, reference_cross_mask + + @staticmethod + def _downsample_mask_to_latent( + mask: torch.Tensor, + latent_num_frames: int, + latent_height: int, + latent_width: int, + ) -> torch.Tensor: + """ + Downsample a pixel-space attention mask to a flattened per-token latent-space mask. Uses causal temporal + downsampling (the first frame is kept as-is). + + Args: + mask (`torch.Tensor`): + Pixel-space mask of shape `(B, 1, F_pix, H_pix, W_pix)` with values in `[0, 1]`. + latent_num_frames (`int`), latent_height (`int`), latent_width (`int`): + Target latent dimensions. + + Returns: + Flattened latent-space mask of shape `(B, latent_num_frames * latent_height * latent_width)`. + """ + if mask.ndim != 5 or mask.shape[1] != 1: + raise ValueError( + f"Expected `conditioning_attention_mask` of shape (B, 1, F, H, W), got {tuple(mask.shape)}." + ) + b, _, f_pix, _, _ = mask.shape + + # 1. Spatial downsampling (area interpolation per frame). + mask_2d = mask.reshape(b * f_pix, 1, mask.shape[-2], mask.shape[-1]) + spatial_down = torch.nn.functional.interpolate(mask_2d, size=(latent_height, latent_width), mode="area") + spatial_down = spatial_down.reshape(b, 1, f_pix, latent_height, latent_width) + + # 2. Causal temporal downsampling. + first_frame = spatial_down[:, :, :1, :, :] # (B, 1, 1, H_lat, W_lat) + if f_pix > 1 and latent_num_frames > 1: + t = (f_pix - 1) // (latent_num_frames - 1) + if (f_pix - 1) % (latent_num_frames - 1) != 0: + raise ValueError( + f"Pixel frames ({f_pix}) not compatible with latent frames ({latent_num_frames}): " + f"(f_pix - 1) must be divisible by (latent_num_frames - 1)." + ) + rest = spatial_down[:, :, 1:, :, :] + rest = rest.reshape(b, 1, latent_num_frames - 1, t, latent_height, latent_width).mean(dim=3) + latent_mask = torch.cat([first_frame, rest], dim=2) + else: + latent_mask = first_frame + + # 3. Flatten to token order (f, h, w). + return latent_mask.reshape(b, latent_num_frames * latent_height * latent_width) + + @staticmethod + def _build_video_self_attention_mask( + num_noisy_tokens: int, + extras_cross_masks: list[torch.Tensor], + device: torch.device, + dtype: torch.dtype = torch.float32, + ) -> torch.Tensor: + """ + Build the `(1, T_video, T_video)` self-attention mask over `noisy + extras` tokens, where `extras` is a + concatenation of one or more conditioning groups (e.g. keyframes, IC-LoRA references). + + Block structure (mirrors the reference `update_attention_mask` / `ConditioningItemAttentionStrengthWrapper`): + - noisy ↔ noisy: 1.0 (full attention) + - noisy ↔ group_i: `extras_cross_masks[i]` broadcast across the noisy-token axis + - group_i ↔ noisy: `extras_cross_masks[i]` broadcast across the noisy-token axis (symmetric) + - group_i ↔ group_i: 1.0 (tokens in a group fully attend to themselves) + - group_i ↔ group_j (i != j): 0.0 (different conditioning groups don't cross-attend) + + Args: + num_noisy_tokens (`int`): + Number of noisy video tokens. + extras_cross_masks (`list[torch.Tensor]`): + List of per-token cross-attention strengths, one per conditioning group. Each entry has shape `(1, + num_tokens_in_group)` with values in `[0, 1]`. Groups must appear in the same order as their tokens in + the extras block. + device, dtype: + Tensor device and dtype. + + Returns: + Multiplicative self-attention mask of shape `(1, num_noisy_tokens + sum(group_sizes), num_noisy_tokens + + sum(group_sizes))` with values in `[0, 1]`. + """ + total_extras = sum(m.shape[1] for m in extras_cross_masks) + total = num_noisy_tokens + total_extras + + # Initialize to 0 so that between-group blocks remain masked without explicit assignment. + attn_mask = torch.zeros((1, total, total), device=device, dtype=dtype) + attn_mask[:, :num_noisy_tokens, :num_noisy_tokens] = 1.0 # noisy ↔ noisy + + offset = num_noisy_tokens + for cross_mask in extras_cross_masks: + n = cross_mask.shape[1] + cross = cross_mask.to(device=device, dtype=dtype) + # noisy (rows) ↔ this group (cols) + attn_mask[:, :num_noisy_tokens, offset : offset + n] = cross.unsqueeze(1) + # this group (rows) ↔ noisy (cols) + attn_mask[:, offset : offset + n, :num_noisy_tokens] = cross.unsqueeze(2) + # this group ↔ this group (self-attention within the group) + attn_mask[:, offset : offset + n, offset : offset + n] = 1.0 + offset += n + return attn_mask + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.convert_velocity_to_x0 + def convert_velocity_to_x0( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_x0 = sample - denoised_output * scheduler.sigmas[step_idx] + return sample_x0 + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2_condition.LTX2ConditionPipeline.convert_x0_to_velocity + def convert_x0_to_velocity( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_v = (sample - denoised_output) / scheduler.sigmas[step_idx] + return sample_v + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def stg_scale(self): + return self._stg_scale + + @property + def modality_scale(self): + return self._modality_scale + + @property + def audio_guidance_scale(self): + return self._audio_guidance_scale + + @property + def audio_guidance_rescale(self): + return self._audio_guidance_rescale + + @property + def audio_stg_scale(self): + return self._audio_stg_scale + + @property + def audio_modality_scale(self): + return self._audio_modality_scale + + @property + def do_classifier_free_guidance(self): + return (self._guidance_scale > 1.0) or (self._audio_guidance_scale > 1.0) + + @property + def do_spatio_temporal_guidance(self): + return (self._stg_scale > 0.0) or (self._audio_stg_scale > 0.0) + + @property + def do_modality_isolation_guidance(self): + return (self._modality_scale > 1.0) or (self._audio_modality_scale > 1.0) + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + reference_conditions: LTX2ReferenceCondition | list[LTX2ReferenceCondition] | None = None, + conditions: LTX2VideoCondition | list[LTX2VideoCondition] | None = None, + reference_downscale_factor: int = 1, + conditioning_attention_strength: float = 1.0, + conditioning_attention_mask: torch.Tensor | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + num_inference_steps: int = 30, + sigmas: list[float] | None = None, + timesteps: list[float] | None = None, + guidance_scale: float = 3.0, + stg_scale: float = 1.0, + modality_scale: float = 3.0, + guidance_rescale: float = 0.7, + audio_guidance_scale: float | None = 7.0, + audio_stg_scale: float | None = 1.0, + audio_modality_scale: float | None = 3.0, + audio_guidance_rescale: float | None = 0.7, + spatio_temporal_guidance_blocks: list[int] | None = [28], + noise_scale: float | None = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + audio_latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + use_cross_timestep: bool = True, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 1024, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide video generation. Ignored when not using guidance (i.e., ignored if + `guidance_scale` is less than `1`). + reference_conditions (`LTX2ReferenceCondition` or `List[LTX2ReferenceCondition]`, *optional*): + Reference video conditions for IC-LoRA conditioning. Each reference video is encoded into latent tokens + and concatenated to the noisy latent sequence during denoising, allowing the IC-LoRA adapter to + condition the generation on the reference video content. + conditions (`LTX2VideoCondition` or `List[LTX2VideoCondition]`, *optional*): + Frame-level conditioning (same as [`LTX2ConditionPipeline`]). Conditions are inserted at specific + latent positions and blended with the denoised output during each denoising step. + reference_downscale_factor (`int`, *optional*, defaults to `1`): + Ratio between target and reference video resolutions. IC-LoRA models trained with downscaled reference + videos store this factor in their safetensors metadata (`reference_downscale_factor` key). A factor of + `2` means the reference video is preprocessed at half the target resolution and spatial positional + coordinates are scaled accordingly. + conditioning_attention_strength (`float`, *optional*, defaults to `1.0`): + Scalar in `[0, 1]` controlling how strongly noisy tokens and appended reference tokens attend to each + other in the video self-attention. `1.0` = full attention (no masking, same as the base IC-LoRA + behavior). `0.0` = reference tokens are fully masked out of the noisy-token attention (and vice versa). + Only takes effect when `reference_conditions` is provided. + conditioning_attention_mask (`torch.Tensor`, *optional*): + Optional pixel-space spatial attention mask of shape `(1, 1, F_pix, H_pix, W_pix)` with values in `[0, + 1]` that provides per-region attention strength. The mask's spatial-temporal dimensions must match the + reference video's pixel dimensions. Downsampled to latent space using VAE scale factors (with causal + temporal handling for the first frame) and multiplied by `conditioning_attention_strength` to form the + final cross-attention mask between noisy and reference tokens. Only takes effect when + `reference_conditions` is provided. + height (`int`, *optional*, defaults to `512`): + The height in pixels of the generated video. + width (`int`, *optional*, defaults to `768`): + The width in pixels of the generated video. + num_frames (`int`, *optional*, defaults to `121`): + The number of video frames to generate. + frame_rate (`float`, *optional*, defaults to `24.0`): + The frames per second (FPS) of the generated video. + num_inference_steps (`int`, *optional*, defaults to 40): + The number of denoising steps. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process. + guidance_scale (`float`, *optional*, defaults to `4.0`): + Classifier-Free Guidance scale for video. + stg_scale (`float`, *optional*, defaults to `0.0`): + Spatio-Temporal Guidance scale for video. `0.0` disables STG. + modality_scale (`float`, *optional*, defaults to `1.0`): + Modality isolation guidance scale for video. `1.0` disables modality guidance. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor for video. + audio_guidance_scale (`float`, *optional*, defaults to `None`): + CFG scale for audio. If `None`, defaults to `guidance_scale`. + audio_stg_scale (`float`, *optional*, defaults to `None`): + STG scale for audio. If `None`, defaults to `stg_scale`. + audio_modality_scale (`float`, *optional*, defaults to `None`): + Modality guidance scale for audio. If `None`, defaults to `modality_scale`. + audio_guidance_rescale (`float`, *optional*, defaults to `None`): + Guidance rescale for audio. If `None`, defaults to `guidance_rescale`. + spatio_temporal_guidance_blocks (`list[int]`, *optional*): + Transformer block indices at which to apply STG. + noise_scale (`float`, *optional*): + Noise scale for latent initialization. If not set, inferred from the sigma schedule. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + Random generator(s) for reproducibility. + latents (`torch.Tensor`, *optional*): + Pre-generated video latents (5D unpacked). + audio_latents (`torch.Tensor`, *optional*): + Pre-generated audio latents (4D unpacked). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. + prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + Noise scale at decode time. + use_cross_timestep (`bool`, *optional*, defaults to `False`): + Whether to use cross-modality sigma for cross attention modulation. `True` for LTX-2.3+. + output_type (`str`, *optional*, defaults to `"pil"`): + Output format. Choose `"pil"`, `"np"`, or `"latent"`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`LTX2PipelineOutput`] or a plain tuple. + attention_kwargs (`dict`, *optional*): + Additional kwargs passed to the attention processor. + callback_on_step_end (`Callable`, *optional*): + A function called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`List`, *optional*, defaults to `["latents"]`): + Tensor inputs for the callback function. + max_sequence_length (`int`, *optional*, defaults to `1024`): + Maximum sequence length for the text prompt. + + Examples: + + Returns: + [`LTX2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`LTX2PipelineOutput`] is returned, otherwise a `tuple` of `(video, audio)` + is returned. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + audio_guidance_scale = audio_guidance_scale or guidance_scale + audio_stg_scale = audio_stg_scale or stg_scale + audio_modality_scale = audio_modality_scale or modality_scale + audio_guidance_rescale = audio_guidance_rescale or guidance_rescale + + # 1. Check inputs + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + latents=latents, + audio_latents=audio_latents, + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + stg_scale=stg_scale, + audio_stg_scale=audio_stg_scale, + ) + + # Per-modality guidance scales + self._guidance_scale = guidance_scale + self._stg_scale = stg_scale + self._modality_scale = modality_scale + self._guidance_rescale = guidance_rescale + self._audio_guidance_scale = audio_guidance_scale + self._audio_stg_scale = audio_stg_scale + self._audio_modality_scale = audio_modality_scale + self._audio_guidance_rescale = audio_guidance_rescale + + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if conditions is not None and not isinstance(conditions, list): + conditions = [conditions] + if reference_conditions is not None and not isinstance(reference_conditions, list): + reference_conditions = [reference_conditions] + + # Infer noise scale from sigma schedule if not provided + if noise_scale is None: + noise_scale = sigmas[0] if sigmas is not None else 1.0 + + device = self._execution_device + + # 3. Prepare text embeddings + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + connector_prompt_embeds, connector_audio_prompt_embeds, connector_attention_mask = self.connectors( + prompt_embeds, prompt_attention_mask, padding_side=self.tokenizer_padding_side + ) + + # 4. Prepare latent variables + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + if latents is not None: + logger.info( + "Got latents of shape [batch_size, latent_dim, latent_frames, latent_height, latent_width]," + " `latent_num_frames`, `latent_height`, `latent_width` will be inferred." + ) + _, _, latent_num_frames, latent_height, latent_width = latents.shape + + num_channels_latents = self.transformer.config.in_channels + latents, conditioning_mask, clean_latents, appended_coords, num_ref_tokens, ref_cross_mask = ( + self.prepare_latents( + conditions=conditions, + reference_conditions=reference_conditions, + reference_downscale_factor=reference_downscale_factor, + conditioning_attention_strength=conditioning_attention_strength, + conditioning_attention_mask=conditioning_attention_mask, + batch_size=batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + num_frames=num_frames, + frame_rate=frame_rate, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=latents, + ) + ) + # Track the base token count in the generated video, excluding any appended keyframe and reference-video + # condition tokens. + base_token_count = latents.shape[1] - (appended_coords.shape[2] if appended_coords is not None else 0) + + has_conditions = conditions is not None and len(conditions) > 0 + has_appended_tokens = appended_coords is not None + if self.do_classifier_free_guidance and (has_conditions or num_ref_tokens > 0): + conditioning_mask = torch.cat([conditioning_mask, conditioning_mask]) + + # Build a video self-attention mask over three groups: (1) the noisy latents (2) keyframe conditions, if any + # and (3) reference conditions, if any. Tokens are attend to each other across groups as follows: + # - TODO + video_self_attention_mask: torch.Tensor | None = None + if ref_cross_mask is not None: + num_noisy_tokens = latents.shape[1] - num_ref_tokens + video_self_attention_mask = self._build_video_self_attention_mask( + num_noisy_tokens=num_noisy_tokens, + extras_cross_masks=[ref_cross_mask], + device=device, + ) + + # 5. Prepare audio latents + duration_s = num_frames / frame_rate + audio_latents_per_second = ( + self.audio_sampling_rate / self.audio_hop_length / float(self.audio_vae_temporal_compression_ratio) + ) + audio_num_frames = round(duration_s * audio_latents_per_second) + if audio_latents is not None: + logger.info( + "Got audio_latents of shape [batch_size, num_channels, audio_num_frames, mel_bins]," + " `audio_num_frames` will be inferred." + ) + _, _, audio_num_frames, _ = audio_latents.shape + + latent_mel_bins = self.audio_mel_bins // self.audio_vae_mel_compression_ratio + audio_latents = self.prepare_audio_latents( + batch_size * num_videos_per_prompt, + num_channels_latents=self.audio_latent_channels, + audio_latent_length=audio_num_frames, + num_mel_bins=self.audio_mel_bins, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=audio_latents, + ) + + # 6. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + mu = calculate_shift( + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_image_seq_len", 1024), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.95), + self.scheduler.config.get("max_shift", 2.05), + ) + if self.audio_scheduler is not None: + audio_scheduler = self.audio_scheduler + else: + audio_scheduler = copy.deepcopy(self.scheduler) + audio_timesteps, _ = retrieve_timesteps( + audio_scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 7. Prepare positional coordinates + video_coords = self.transformer.rope.prepare_video_coords( + latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate + ) + if appended_coords is not None: + video_coords = torch.cat([video_coords, appended_coords], dim=2) + audio_coords = self.transformer.audio_rope.prepare_audio_coords( + audio_latents.shape[0], audio_num_frames, audio_latents.device + ) + if self.do_classifier_free_guidance: + video_coords = video_coords.repeat((2,) + (1,) * (video_coords.ndim - 1)) + audio_coords = audio_coords.repeat((2,) + (1,) * (audio_coords.ndim - 1)) + + # 8. Denoising loop + video_seq_len = latents.shape[1] + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + audio_latent_model_input = ( + torch.cat([audio_latents] * 2) if self.do_classifier_free_guidance else audio_latents + ) + audio_latent_model_input = audio_latent_model_input.to(prompt_embeds.dtype) + + timestep_scalar = t.expand(latent_model_input.shape[0]) + + if has_conditions or num_ref_tokens > 0: + video_timestep = timestep_scalar.unsqueeze(-1) * (1 - conditioning_mask.squeeze(-1)) + else: + video_timestep = timestep_scalar.unsqueeze(-1).expand(-1, video_seq_len) + + t_audio = audio_timesteps[i] + audio_timestep = t_audio.expand(latent_model_input.shape[0]) + + # --- Main transformer forward pass (conditional + unconditional for CFG) --- + if video_self_attention_mask is not None: + video_self_attention_mask = video_self_attention_mask.expand(latent_model_input.shape[0], -1, -1) + with self.transformer.cache_context("cond_uncond"): + noise_pred_video, noise_pred_audio = self.transformer( + hidden_states=latent_model_input, + audio_hidden_states=audio_latent_model_input, + encoder_hidden_states=connector_prompt_embeds, + audio_encoder_hidden_states=connector_audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=audio_timestep, + sigma=timestep_scalar, # Used by LTX-2.3 + audio_sigma=audio_timestep, + encoder_attention_mask=connector_attention_mask, + audio_encoder_attention_mask=connector_attention_mask, + video_self_attention_mask=video_self_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_coords, + audio_coords=audio_coords, + isolate_modalities=False, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video = noise_pred_video.float() + noise_pred_audio = noise_pred_audio.float() + + if self.do_classifier_free_guidance: + noise_pred_video_uncond_text, noise_pred_video = noise_pred_video.chunk(2) + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_video_uncond_text = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_text, i, self.scheduler + ) + video_cfg_delta = (self.guidance_scale - 1) * (noise_pred_video - noise_pred_video_uncond_text) + + noise_pred_audio_uncond_text, noise_pred_audio = noise_pred_audio.chunk(2) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + noise_pred_audio_uncond_text = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_text, i, audio_scheduler + ) + audio_cfg_delta = (self.audio_guidance_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_text + ) + + if self.do_spatio_temporal_guidance or self.do_modality_isolation_guidance: + if i == 0: + video_prompt_embeds = connector_prompt_embeds.chunk(2, dim=0)[1] + audio_prompt_embeds = connector_audio_prompt_embeds.chunk(2, dim=0)[1] + prompt_attn_mask = connector_attention_mask.chunk(2, dim=0)[1] + + video_pos_ids = video_coords.chunk(2, dim=0)[0] + audio_pos_ids = audio_coords.chunk(2, dim=0)[0] + + timestep_scalar_single = timestep_scalar.chunk(2, dim=0)[0] + if has_conditions or num_ref_tokens > 0: + video_timestep_single = video_timestep.chunk(2, dim=0)[0] + else: + video_timestep_single = timestep_scalar_single.unsqueeze(-1).expand(-1, video_seq_len) + audio_timestep_single = audio_timestep.chunk(2, dim=0)[0] + else: + video_cfg_delta = audio_cfg_delta = 0 + + video_prompt_embeds = connector_prompt_embeds + audio_prompt_embeds = connector_audio_prompt_embeds + prompt_attn_mask = connector_attention_mask + + video_pos_ids = video_coords + audio_pos_ids = audio_coords + + timestep_scalar_single = timestep_scalar + if has_conditions or num_ref_tokens > 0: + video_timestep_single = video_timestep + else: + video_timestep_single = timestep_scalar.unsqueeze(-1).expand(-1, video_seq_len) + audio_timestep_single = audio_timestep + + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + + # --- STG forward pass --- + if self.do_spatio_temporal_guidance: + if video_self_attention_mask is not None: + video_self_attention_mask = video_self_attention_mask.expand(latents.shape[0], -1, -1) + with self.transformer.cache_context("uncond_stg"): + noise_pred_video_uncond_stg, noise_pred_audio_uncond_stg = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep_single, + audio_timestep=audio_timestep_single, + sigma=timestep_scalar_single, # Used by LTX-2.3 + audio_sigma=audio_timestep_single, + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + video_self_attention_mask=video_self_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + isolate_modalities=False, + # Use STG at given blocks to perturb model + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_stg = noise_pred_video_uncond_stg.float() + noise_pred_audio_uncond_stg = noise_pred_audio_uncond_stg.float() + noise_pred_video_uncond_stg = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_stg, i, self.scheduler + ) + noise_pred_audio_uncond_stg = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_stg, i, audio_scheduler + ) + + video_stg_delta = self.stg_scale * (noise_pred_video - noise_pred_video_uncond_stg) + audio_stg_delta = self.audio_stg_scale * (noise_pred_audio - noise_pred_audio_uncond_stg) + else: + video_stg_delta = audio_stg_delta = 0 + + # --- Modality isolation guidance forward pass --- + if self.do_modality_isolation_guidance: + if video_self_attention_mask is not None: + video_self_attention_mask = video_self_attention_mask.expand(latents.shape[0], -1, -1) + with self.transformer.cache_context("uncond_modality"): + noise_pred_video_uncond_mod, noise_pred_audio_uncond_mod = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep_single, + audio_timestep=audio_timestep_single, + sigma=timestep_scalar_single, # Used by LTX-2.3 + audio_sigma=audio_timestep_single, + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + video_self_attention_mask=video_self_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + # Turn off A2V and V2A cross attn to isolate video and audio modalities + isolate_modalities=True, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_mod = noise_pred_video_uncond_mod.float() + noise_pred_audio_uncond_mod = noise_pred_audio_uncond_mod.float() + noise_pred_video_uncond_mod = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_mod, i, self.scheduler + ) + noise_pred_audio_uncond_mod = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_mod, i, audio_scheduler + ) + + video_modality_delta = (self.modality_scale - 1) * (noise_pred_video - noise_pred_video_uncond_mod) + audio_modality_delta = (self.audio_modality_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_mod + ) + else: + video_modality_delta = audio_modality_delta = 0 + + # Apply all guidance terms + noise_pred_video_g = noise_pred_video + video_cfg_delta + video_stg_delta + video_modality_delta + noise_pred_audio_g = noise_pred_audio + audio_cfg_delta + audio_stg_delta + audio_modality_delta + + # Apply guidance rescaling + if self.guidance_rescale > 0: + noise_pred_video = rescale_noise_cfg( + noise_pred_video_g, noise_pred_video, guidance_rescale=self.guidance_rescale + ) + else: + noise_pred_video = noise_pred_video_g + + if self.audio_guidance_rescale > 0: + noise_pred_audio = rescale_noise_cfg( + noise_pred_audio_g, noise_pred_audio, guidance_rescale=self.audio_guidance_rescale + ) + else: + noise_pred_audio = noise_pred_audio_g + + # Apply frame conditioning mask: blend denoised x0 with clean condition latents + if has_conditions: + bsz = noise_pred_video.size(0) + denoised_sample_cond = ( + noise_pred_video * (1 - conditioning_mask[:bsz]) + + clean_latents.float() * conditioning_mask[:bsz] + ).to(noise_pred_video.dtype) + noise_pred_video = denoised_sample_cond + + # Convert back to velocity for scheduler + noise_pred_video = self.convert_x0_to_velocity(latents, noise_pred_video, i, self.scheduler) + noise_pred_audio = self.convert_x0_to_velocity(audio_latents, noise_pred_audio, i, audio_scheduler) + + # Compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred_video, t, latents, return_dict=False)[0] + audio_latents = audio_scheduler.step(noise_pred_audio, t, audio_latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Decode + # Trim any appended keyframe or reference tokens from the latents to recover the generated video only. + latents = latents[:, :base_token_count] + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + audio_latents = self._denormalize_audio_latents( + audio_latents, self.audio_vae.latents_mean, self.audio_vae.latents_std + ) + audio_latents = self._unpack_audio_latents(audio_latents, audio_num_frames, num_mel_bins=latent_mel_bins) + + if output_type == "latent": + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + video = latents + audio = audio_latents + else: + latents = latents.to(prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + + latents = latents.to(self.vae.dtype) + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + audio_latents = audio_latents.to(self.audio_vae.dtype) + generated_mel_spectrograms = self.audio_vae.decode(audio_latents, return_dict=False)[0] + audio = self.vocoder(generated_mel_spectrograms) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video, audio) + + return LTX2PipelineOutput(frames=video, audio=audio) diff --git a/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py b/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py new file mode 100644 index 0000000000000000000000000000000000000000..1f721bf54364d5c1fb1434444dbc6338a2eb4614 --- /dev/null +++ b/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py @@ -0,0 +1,1585 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Gemma3ForConditionalGeneration, Gemma3Processor, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import FromSingleFileMixin, LTX2LoraLoaderMixin +from ...models.autoencoders import AutoencoderKLLTX2Audio, AutoencoderKLLTX2Video +from ...models.transformers import LTX2VideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .connectors import LTX2TextConnectors +from .pipeline_output import LTX2PipelineOutput +from .vocoder import LTX2Vocoder, LTX2VocoderWithBWE + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTX2ImageToVideoPipeline + >>> from diffusers.utils import encode_video + >>> from diffusers.utils import load_image + + >>> pipe = LTX2ImageToVideoPipeline.from_pretrained("Lightricks/LTX-2", torch_dtype=torch.bfloat16) + >>> pipe.enable_model_cpu_offload() + + >>> image = load_image( + ... "https://huggingface.co/datasets/a-r-r-o-w/tiny-meme-dataset-captioned/resolve/main/images/8.png" + ... ) + >>> prompt = "A young girl stands calmly in the foreground, looking directly at the camera, as a house fire rages in the background." + >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted" + + >>> frame_rate = 24.0 + >>> video, audio = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... width=768, + ... height=512, + ... num_frames=121, + ... frame_rate=frame_rate, + ... num_inference_steps=40, + ... guidance_scale=4.0, + ... output_type="np", + ... return_dict=False, + ... ) + + >>> encode_video( + ... video[0], + ... fps=frame_rate, + ... audio=audio[0].float().cpu(), + ... audio_sample_rate=pipe.vocoder.config.output_sampling_rate, # should be 24000 + ... output_path="video.mp4", + ... ) + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class LTX2ImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraLoaderMixin): + r""" + Pipeline for image-to-video generation. + + Reference: https://github.com/Lightricks/LTX-Video + + TODO + """ + + model_cpu_offload_seq = "text_encoder->connectors->transformer->vae->audio_vae->vocoder" + _optional_components = ["processor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLLTX2Video, + audio_vae: AutoencoderKLLTX2Audio, + text_encoder: Gemma3ForConditionalGeneration, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + connectors: LTX2TextConnectors, + transformer: LTX2VideoTransformer3DModel, + vocoder: LTX2Vocoder | LTX2VocoderWithBWE, + processor: Gemma3Processor | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + audio_vae=audio_vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + connectors=connectors, + transformer=transformer, + vocoder=vocoder, + scheduler=scheduler, + processor=processor, + ) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + # TODO: check whether the MEL compression ratio logic here is corrct + self.audio_vae_mel_compression_ratio = ( + self.audio_vae.mel_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.audio_vae_temporal_compression_ratio = ( + self.audio_vae.temporal_compression_ratio if getattr(self, "audio_vae", None) is not None else 4 + ) + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1 + ) + + self.audio_sampling_rate = ( + self.audio_vae.config.sample_rate if getattr(self, "audio_vae", None) is not None else 16000 + ) + self.audio_hop_length = ( + self.audio_vae.config.mel_hop_length if getattr(self, "audio_vae", None) is not None else 160 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio, resample="bilinear") + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 1024 + ) + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + num_videos_per_prompt: int = 1, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`str` or `torch.device`): + torch device to place the resulting embeddings on + dtype: (`torch.dtype`): + torch dtype to cast the prompt embeds to + max_sequence_length (`int`, defaults to 1024): Maximum sequence length to use for the prompt. + """ + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if getattr(self, "tokenizer", None) is not None: + # Gemma expects left padding for chat-style prompts + self.tokenizer.padding_side = "left" + if self.tokenizer.pad_token is None: + self.tokenizer.pad_token = self.tokenizer.eos_token + + prompt = [p.strip() for p in prompt] + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_attention_mask = text_inputs.attention_mask + text_input_ids = text_input_ids.to(device) + prompt_attention_mask = prompt_attention_mask.to(device) + + text_encoder_outputs = self.text_encoder( + input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True + ) + text_encoder_hidden_states = text_encoder_outputs.hidden_states + text_encoder_hidden_states = torch.stack(text_encoder_hidden_states, dim=-1) + prompt_embeds = text_encoder_hidden_states.flatten(2, 3).to(dtype=dtype) # Pack to 3D + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + scale_factor: int = 8, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + scale_factor=scale_factor, + device=device, + dtype=dtype, + ) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + @torch.no_grad() + def enhance_prompt( + self, + image: PipelineImageInput, + prompt: str, + system_prompt: str, + max_new_tokens: int = 512, + seed: int = 10, + generator: torch.Generator | None = None, + generation_kwargs: dict[str, Any] | None = None, + device: str | torch.device | None = None, + ): + """ + Enhances the supplied `prompt` by generating a new prompt using the current text encoder (default is a + `transformers.Gemma3ForConditionalGeneration` model) from it and a system prompt. + """ + device = device or self._execution_device + if generation_kwargs is None: + # Set to default generation kwargs + generation_kwargs = {"do_sample": True, "temperature": 0.7} + + messages = [ + {"role": "system", "content": system_prompt}, + { + "role": "user", + "content": [ + {"type": "image"}, + {"type": "text", "text": f"User Raw Input Prompt: {prompt}."}, + ], + }, + ] + template = self.processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) + model_inputs = self.processor(text=template, images=image, return_tensors="pt").to(device) + self.text_encoder.to(device) + + # `transformers.GenerationMixin.generate` does not support using a `torch.Generator` to control randomness, + # so manually apply a seed for reproducible generation. + if generator is not None: + # Overwrite seed to generator's initial seed + seed = generator.initial_seed() + torch.manual_seed(seed) + generated_sequences = self.text_encoder.generate( + **model_inputs, + max_new_tokens=max_new_tokens, + **generation_kwargs, + ) # tensor of shape [batch_size, seq_len] + + generated_ids = [seq[len(model_inputs.input_ids[i]) :] for i, seq in enumerate(generated_sequences)] + enhanced_prompt = self.processor.tokenizer.batch_decode(generated_ids, skip_special_tokens=True) + return enhanced_prompt + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + spatio_temporal_guidance_blocks=None, + stg_scale=None, + audio_stg_scale=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + if ((stg_scale > 0.0) or (audio_stg_scale > 0.0)) and not spatio_temporal_guidance_blocks: + raise ValueError( + "Spatio-Temporal Guidance (STG) is specified but no STG blocks are supplied. Please supply a list of" + "block indices at which to apply STG in `spatio_temporal_guidance_blocks`" + ) + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_latents + def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor: + # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p]. + # The patch dimensions are then permuted and collapsed into the channel dimension of shape: + # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor). + # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features + batch_size, num_channels, num_frames, height, width = latents.shape + post_patch_num_frames = num_frames // patch_size_t + post_patch_height = height // patch_size + post_patch_width = width // patch_size + latents = latents.reshape( + batch_size, + -1, + post_patch_num_frames, + patch_size_t, + post_patch_height, + patch_size, + post_patch_width, + patch_size, + ) + latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + @staticmethod + def _normalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Normalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = (latents - latents_mean) * scaling_factor / latents_std + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._create_noised_state + def _create_noised_state( + latents: torch.Tensor, noise_scale: float | torch.Tensor, generator: torch.Generator | None = None + ): + noise = randn_tensor(latents.shape, generator=generator, device=latents.device, dtype=latents.dtype) + noised_latents = noise_scale * noise + (1 - noise_scale) * latents + return noised_latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._pack_audio_latents + def _pack_audio_latents( + latents: torch.Tensor, patch_size: int | None = None, patch_size_t: int | None = None + ) -> torch.Tensor: + # Audio latents shape: [B, C, L, M], where L is the latent audio length and M is the number of mel bins + if patch_size is not None and patch_size_t is not None: + # Packs the latents into a patch sequence of shape [B, L // p_t * M // p, C * p_t * p] (a ndim=3 tnesor). + # dim=1 is the effective audio sequence length and dim=2 is the effective audio input feature size. + batch_size, num_channels, latent_length, latent_mel_bins = latents.shape + post_patch_latent_length = latent_length / patch_size_t + post_patch_mel_bins = latent_mel_bins / patch_size + latents = latents.reshape( + batch_size, -1, post_patch_latent_length, patch_size_t, post_patch_mel_bins, patch_size + ) + latents = latents.permute(0, 2, 4, 1, 3, 5).flatten(3, 5).flatten(1, 2) + else: + # Packs the latents into a patch sequence of shape [B, L, C * M]. This implicitly assumes a (mel) + # patch_size of M (all mel bins constitutes a single patch) and a patch_size_t of 1. + latents = latents.transpose(1, 2).flatten(2, 3) # [B, C, L, M] --> [B, L, C * M] + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_audio_latents + def _unpack_audio_latents( + latents: torch.Tensor, + latent_length: int, + num_mel_bins: int, + patch_size: int | None = None, + patch_size_t: int | None = None, + ) -> torch.Tensor: + # Unpacks an audio patch sequence of shape [B, S, D] into a latent spectrogram tensor of shape [B, C, L, M], + # where L is the latent audio length and M is the number of mel bins. + if patch_size is not None and patch_size_t is not None: + batch_size = latents.size(0) + latents = latents.reshape(batch_size, latent_length, num_mel_bins, -1, patch_size_t, patch_size) + latents = latents.permute(0, 3, 1, 4, 2, 5).flatten(4, 5).flatten(2, 3) + else: + # Assume [B, S, D] = [B, L, C * M], which implies that patch_size = M and patch_size_t = 1. + latents = latents.unflatten(2, (-1, num_mel_bins)).transpose(1, 2) + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._normalize_audio_latents + def _normalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents - latents_mean) / latents_std + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_audio_latents + def _denormalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor): + latents_mean = latents_mean.to(latents.device, latents.dtype) + latents_std = latents_std.to(latents.device, latents.dtype) + return (latents * latents_std) + latents_mean + + def prepare_latents( + self, + image: torch.Tensor | None = None, + batch_size: int = 1, + num_channels_latents: int = 128, + height: int = 512, + width: int = 704, + num_frames: int = 161, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + height = height // self.vae_spatial_compression_ratio + width = width // self.vae_spatial_compression_ratio + num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + + shape = (batch_size, num_channels_latents, num_frames, height, width) + mask_shape = (batch_size, 1, num_frames, height, width) + + if latents is not None: + if latents.ndim == 5: + # conditioning_mask needs to the same shape as latents in two stages generation. + batch_size, _, num_frames, height, width = latents.shape + mask_shape = (batch_size, 1, num_frames, height, width) + conditioning_mask = latents.new_zeros(mask_shape) + conditioning_mask[:, :, 0] = 1.0 + + latents = self._normalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = self._create_noised_state(latents, noise_scale * (1 - conditioning_mask), generator) + # latents are of shape [B, C, F, H, W], need to be packed + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + else: + conditioning_mask = latents.new_zeros(mask_shape) + conditioning_mask[:, :, 0] = 1.0 + conditioning_mask = self._pack_latents( + conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ).squeeze(-1) + if latents.ndim != 3 or latents.shape[:2] != conditioning_mask.shape: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is {conditioning_mask.shape + (num_channels_latents,)}." + ) + return latents.to(device=device, dtype=dtype), conditioning_mask + + if isinstance(generator, list): + if len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i].unsqueeze(0).unsqueeze(2)), generator[i], "argmax") + for i in range(batch_size) + ] + else: + init_latents = [ + retrieve_latents(self.vae.encode(img.unsqueeze(0).unsqueeze(2)), generator, "argmax") for img in image + ] + + init_latents = torch.cat(init_latents, dim=0).to(dtype) + init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std) + init_latents = init_latents.repeat(1, 1, num_frames, 1, 1) + + # First condition is image latents and those should be kept clean. + conditioning_mask = torch.zeros(mask_shape, device=device, dtype=dtype) + conditioning_mask[:, :, 0] = 1.0 + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # Interpolation. + latents = init_latents * conditioning_mask + noise * (1 - conditioning_mask) + + conditioning_mask = self._pack_latents( + conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ).squeeze(-1) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + + return latents, conditioning_mask + + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline.prepare_audio_latents + def prepare_audio_latents( + self, + batch_size: int = 1, + num_channels_latents: int = 8, + audio_latent_length: int = 1, # 1 is just a dummy value + num_mel_bins: int = 64, + noise_scale: float = 0.0, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + if latents.ndim == 4: + # latents are of shape [B, C, L, M], need to be packed + latents = self._pack_audio_latents(latents) + if latents.ndim != 3: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is [batch_size, num_seq, num_features]." + ) + latents = self._normalize_audio_latents(latents, self.audio_vae.latents_mean, self.audio_vae.latents_std) + latents = self._create_noised_state(latents, noise_scale, generator) + return latents.to(device=device, dtype=dtype) + + latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio + + shape = (batch_size, num_channels_latents, audio_latent_length, latent_mel_bins) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_audio_latents(latents) + return latents + + def convert_velocity_to_x0( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_x0 = sample - denoised_output * scheduler.sigmas[step_idx] + return sample_x0 + + def convert_x0_to_velocity( + self, sample: torch.Tensor, denoised_output: torch.Tensor, step_idx: int, scheduler: Any | None = None + ) -> torch.Tensor: + if scheduler is None: + scheduler = self.scheduler + + sample_v = (sample - denoised_output) / scheduler.sigmas[step_idx] + return sample_v + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def stg_scale(self): + return self._stg_scale + + @property + def modality_scale(self): + return self._modality_scale + + @property + def audio_guidance_scale(self): + return self._audio_guidance_scale + + @property + def audio_guidance_rescale(self): + return self._audio_guidance_rescale + + @property + def audio_stg_scale(self): + return self._audio_stg_scale + + @property + def audio_modality_scale(self): + return self._audio_modality_scale + + @property + def do_classifier_free_guidance(self): + return (self._guidance_scale > 1.0) or (self._audio_guidance_scale > 1.0) + + @property + def do_spatio_temporal_guidance(self): + return (self._stg_scale > 0.0) or (self._audio_stg_scale > 0.0) + + @property + def do_modality_isolation_guidance(self): + return (self._modality_scale > 1.0) or (self._audio_modality_scale > 1.0) + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + frame_rate: float = 24.0, + num_inference_steps: int = 40, + sigmas: list[float] | None = None, + timesteps: list[int] | None = None, + guidance_scale: float = 4.0, + stg_scale: float = 0.0, + modality_scale: float = 1.0, + guidance_rescale: float = 0.0, + audio_guidance_scale: float | None = None, + audio_stg_scale: float | None = None, + audio_modality_scale: float | None = None, + audio_guidance_rescale: float | None = None, + spatio_temporal_guidance_blocks: list[int] | None = None, + noise_scale: float = 0.0, + num_videos_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + audio_latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + use_cross_timestep: bool = False, + system_prompt: str | None = None, + prompt_max_new_tokens: int = 512, + prompt_enhancement_kwargs: dict[str, Any] | None = None, + prompt_enhancement_seed: int = 10, + output_type: str = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 1024, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + height (`int`, *optional*, defaults to `512`): + The height in pixels of the generated image. This is set to 480 by default for the best results. + width (`int`, *optional*, defaults to `768`): + The width in pixels of the generated image. This is set to 848 by default for the best results. + num_frames (`int`, *optional*, defaults to `121`): + The number of video frames to generate + frame_rate (`float`, *optional*, defaults to `24.0`): + The frames per second (FPS) of the generated video. + num_inference_steps (`int`, *optional*, defaults to 40): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to `4.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. Used for the video modality (there is + a separate value `audio_guidance_scale` for the audio modality). + stg_scale (`float`, *optional*, defaults to `0.0`): + Video guidance scale for Spatio-Temporal Guidance (STG), proposed in [Spatiotemporal Skip Guidance for + Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664). STG uses a CFG-like estimate + where we move the sample away from a weak sample from a perturbed version of the denoising model. + Enabling STG will result in an additional denoising model forward pass; the default value of `0.0` + means that STG is disabled. + modality_scale (`float`, *optional*, defaults to `1.0`): + Video guidance scale for LTX-2.X modality isolation guidance, where we move the sample away from a + weaker sample generated by the denoising model withy cross-modality (audio-to-video and video-to-audio) + cross attention disabled using a CFG-like estimate. Enabling modality guidance will result in an + additional denoising model forward pass; the default value of `1.0` means that modality guidance is + disabled. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. Used for the video modality. + audio_guidance_scale (`float`, *optional* defaults to `None`): + Audio guidance scale for CFG with respect to the negative prompt. The CFG update rule is the same for + video and audio, but they can use different values for the guidance scale. The LTX-2.X authors suggest + that the `audio_guidance_scale` should be higher relative to the video `guidance_scale` (e.g. for + LTX-2.3 they suggest 3.0 for video and 7.0 for audio). If `None`, defaults to the video value + `guidance_scale`. + audio_stg_scale (`float`, *optional*, defaults to `None`): + Audio guidance scale for STG. As with CFG, the STG update rule is otherwise the same for video and + audio. For LTX-2.3, a value of 1.0 is suggested for both video and audio. If `None`, defaults to the + video value `stg_scale`. + audio_modality_scale (`float`, *optional*, defaults to `None`): + Audio guidance scale for LTX-2.X modality isolation guidance. As with CFG, the modality guidance rule + is otherwise the same for video and audio. For LTX-2.3, a value of 3.0 is suggested for both video and + audio. If `None`, defaults to the video value `modality_scale`. + audio_guidance_rescale (`float`, *optional*, defaults to `None`): + A separate guidance rescale factor for the audio modality. If `None`, defaults to the video value + `guidance_rescale`. + spatio_temporal_guidance_blocks (`list[int]`, *optional*, defaults to `None`): + The zero-indexed transformer block indices at which to apply STG. Must be supplied if STG is used + (`stg_scale` or `audio_stg_scale` is greater than `0`). A value of `[29]` is recommended for LTX-2.0 + and `[28]` is recommended for LTX-2.3. + noise_scale (`float`, *optional*, defaults to `0.0`): + The interpolation factor between random noise and denoised latents at each timestep. Applying noise to + the `latents` and `audio_latents` before continue denoising. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + audio_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for audio + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + The interpolation factor between random noise and denoised latents at the decode timestep. + use_cross_timestep (`bool` *optional*, defaults to `False`): + Whether to use the cross modality (audio is the cross modality of video, and vice versa) sigma when + calculating the cross attention modulation parameters. `True` is the newer (e.g. LTX-2.3) behavior; + `False` is the legacy LTX-2.0 behavior. + system_prompt (`str`, *optional*, defaults to `None`): + Optional system prompt to use for prompt enhancement. The system prompt will be used by the current + text encoder (by default, a `Gemma3ForConditionalGeneration` model) to generate an enhanced prompt from + the original `prompt` to condition generation. If not supplied, prompt enhancement will not be + performed. + prompt_max_new_tokens (`int`, *optional*, defaults to `512`): + The maximum number of new tokens to generate when performing prompt enhancement. + prompt_enhancement_kwargs (`dict[str, Any]`, *optional*, defaults to `None`): + Keyword arguments for `self.text_encoder.generate`. If not supplied, default arguments of + `do_sample=True` and `temperature=0.7` will be used. See + https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationMixin.generate + for more details. + prompt_enhancement_seed (`int`, *optional*, default to `10`): + Random seed for any random operations during prompt enhancement. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTX2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `1024`): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ltx.LTX2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTX2PipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + audio_guidance_scale = audio_guidance_scale or guidance_scale + audio_stg_scale = audio_stg_scale or stg_scale + audio_modality_scale = audio_modality_scale or modality_scale + audio_guidance_rescale = audio_guidance_rescale or guidance_rescale + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + stg_scale=stg_scale, + audio_stg_scale=audio_stg_scale, + ) + + # Per-modality guidance scales (video, audio) + self._guidance_scale = guidance_scale + self._stg_scale = stg_scale + self._modality_scale = modality_scale + self._guidance_rescale = guidance_rescale + self._audio_guidance_scale = audio_guidance_scale + self._audio_stg_scale = audio_stg_scale + self._audio_modality_scale = audio_modality_scale + self._audio_guidance_rescale = audio_guidance_rescale + + self._attention_kwargs = attention_kwargs + self._interrupt = False + self._current_timestep = None + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + if system_prompt is not None and prompt is not None: + prompt = self.enhance_prompt( + image=image, + prompt=prompt, + system_prompt=system_prompt, + max_new_tokens=prompt_max_new_tokens, + seed=prompt_enhancement_seed, + generator=generator, + generation_kwargs=prompt_enhancement_kwargs, + device=device, + ) + + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=max_sequence_length, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + tokenizer_padding_side = "left" # Padding side for default Gemma3-12B text encoder + if getattr(self, "tokenizer", None) is not None: + tokenizer_padding_side = getattr(self.tokenizer, "padding_side", "left") + connector_prompt_embeds, connector_audio_prompt_embeds, connector_attention_mask = self.connectors( + prompt_embeds, prompt_attention_mask, padding_side=tokenizer_padding_side + ) + + # 4. Prepare latent variables + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + if latents is not None: + if latents.ndim == 5: + logger.info( + "Got latents of shape [batch_size, latent_dim, latent_frames, latent_height, latent_width], `latent_num_frames`, `latent_height`, `latent_width` will be inferred." + ) + _, _, latent_num_frames, latent_height, latent_width = latents.shape # [B, C, F, H, W] + elif latents.ndim == 3: + logger.warning( + f"You have supplied packed `latents` of shape {latents.shape}, so the latent dims cannot be" + f" inferred. Make sure the supplied `height`, `width`, and `num_frames` are correct." + ) + else: + raise ValueError( + f"Provided `latents` tensor has shape {latents.shape}, but the expected shape is either [batch_size, seq_len, num_features] or [batch_size, latent_dim, latent_frames, latent_height, latent_width]." + ) + # video_sequence_length = latent_num_frames * latent_height * latent_width + + if latents is None: + image = self.video_processor.preprocess(image, height=height, width=width) + image = image.to(device=device, dtype=prompt_embeds.dtype) + + num_channels_latents = self.transformer.config.in_channels + latents, conditioning_mask = self.prepare_latents( + image, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + noise_scale, + torch.float32, + device, + generator, + latents, + ) + if self.do_classifier_free_guidance: + conditioning_mask = torch.cat([conditioning_mask, conditioning_mask]) + + duration_s = num_frames / frame_rate + audio_latents_per_second = ( + self.audio_sampling_rate / self.audio_hop_length / float(self.audio_vae_temporal_compression_ratio) + ) + audio_num_frames = round(duration_s * audio_latents_per_second) + if audio_latents is not None: + if audio_latents.ndim == 4: + logger.info( + "Got audio_latents of shape [batch_size, num_channels, audio_length, mel_bins], `audio_num_frames` will be inferred." + ) + _, _, audio_num_frames, _ = audio_latents.shape # [B, C, L, M] + elif audio_latents.ndim == 3: + logger.warning( + f"You have supplied packed `audio_latents` of shape {audio_latents.shape}, so the latent dims" + f" cannot be inferred. Make sure the supplied `num_frames` and `frame_rate` are correct." + ) + else: + raise ValueError( + f"Provided `audio_latents` tensor has shape {audio_latents.shape}, but the expected shape is either [batch_size, seq_len, num_features] or [batch_size, num_channels, audio_length, mel_bins]." + ) + + num_mel_bins = self.audio_vae.config.mel_bins if getattr(self, "audio_vae", None) is not None else 64 + latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio + num_channels_latents_audio = ( + self.audio_vae.config.latent_channels if getattr(self, "audio_vae", None) is not None else 8 + ) + audio_latents = self.prepare_audio_latents( + batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents_audio, + audio_latent_length=audio_num_frames, + num_mel_bins=num_mel_bins, + noise_scale=noise_scale, + dtype=torch.float32, + device=device, + generator=generator, + latents=audio_latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + mu = calculate_shift( + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_image_seq_len", 1024), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.95), + self.scheduler.config.get("max_shift", 2.05), + ) + + # For now, duplicate the scheduler for use with the audio latents + audio_scheduler = copy.deepcopy(self.scheduler) + _, _ = retrieve_timesteps( + audio_scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Prepare micro-conditions + # Pre-compute video and audio positional ids as they will be the same at each step of the denoising loop + video_coords = self.transformer.rope.prepare_video_coords( + latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate + ) + audio_coords = self.transformer.audio_rope.prepare_audio_coords( + audio_latents.shape[0], audio_num_frames, audio_latents.device + ) + # Duplicate the positional ids as well if using CFG + if self.do_classifier_free_guidance: + video_coords = video_coords.repeat((2,) + (1,) * (video_coords.ndim - 1)) # Repeat twice in batch dim + audio_coords = audio_coords.repeat((2,) + (1,) * (audio_coords.ndim - 1)) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + audio_latent_model_input = ( + torch.cat([audio_latents] * 2) if self.do_classifier_free_guidance else audio_latents + ) + audio_latent_model_input = audio_latent_model_input.to(prompt_embeds.dtype) + + timestep = t.expand(latent_model_input.shape[0]) + video_timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask) + + with self.transformer.cache_context("cond_uncond"): + noise_pred_video, noise_pred_audio = self.transformer( + hidden_states=latent_model_input, + audio_hidden_states=audio_latent_model_input, + encoder_hidden_states=connector_prompt_embeds, + audio_encoder_hidden_states=connector_audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=timestep, + sigma=timestep, # Used by LTX-2.3 + encoder_attention_mask=connector_attention_mask, + audio_encoder_attention_mask=connector_attention_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_coords, + audio_coords=audio_coords, + isolate_modalities=False, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video = noise_pred_video.float() + noise_pred_audio = noise_pred_audio.float() + + if self.do_classifier_free_guidance: + noise_pred_video_uncond_text, noise_pred_video = noise_pred_video.chunk(2) + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_video_uncond_text = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_text, i, self.scheduler + ) + # Use delta formulation as it works more nicely with multiple guidance terms + video_cfg_delta = (self.guidance_scale - 1) * (noise_pred_video - noise_pred_video_uncond_text) + + noise_pred_audio_uncond_text, noise_pred_audio = noise_pred_audio.chunk(2) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + noise_pred_audio_uncond_text = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_text, i, audio_scheduler + ) + audio_cfg_delta = (self.audio_guidance_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_text + ) + + # Get positive values from merged CFG inputs in case we need to do other DiT forward passes + if self.do_spatio_temporal_guidance or self.do_modality_isolation_guidance: + if i == 0: + # Only split values that remain constant throughout the loop once + video_prompt_embeds = connector_prompt_embeds.chunk(2, dim=0)[1] + audio_prompt_embeds = connector_audio_prompt_embeds.chunk(2, dim=0)[1] + prompt_attn_mask = connector_attention_mask.chunk(2, dim=0)[1] + + video_pos_ids = video_coords.chunk(2, dim=0)[0] + audio_pos_ids = audio_coords.chunk(2, dim=0)[0] + + # Split values that vary each denoising loop iteration + timestep = timestep.chunk(2, dim=0)[0] + video_timestep = video_timestep.chunk(2, dim=0)[0] + else: + video_cfg_delta = audio_cfg_delta = 0 + + video_prompt_embeds = connector_prompt_embeds + audio_prompt_embeds = connector_audio_prompt_embeds + prompt_attn_mask = connector_attention_mask + + video_pos_ids = video_coords + audio_pos_ids = audio_coords + + noise_pred_video = self.convert_velocity_to_x0(latents, noise_pred_video, i, self.scheduler) + noise_pred_audio = self.convert_velocity_to_x0(audio_latents, noise_pred_audio, i, audio_scheduler) + + if self.do_spatio_temporal_guidance: + with self.transformer.cache_context("uncond_stg"): + noise_pred_video_uncond_stg, noise_pred_audio_uncond_stg = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=timestep, + sigma=timestep, # Used by LTX-2.3 + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + isolate_modalities=False, + # Use STG at given blocks to perturb model + spatio_temporal_guidance_blocks=spatio_temporal_guidance_blocks, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_stg = noise_pred_video_uncond_stg.float() + noise_pred_audio_uncond_stg = noise_pred_audio_uncond_stg.float() + noise_pred_video_uncond_stg = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_stg, i, self.scheduler + ) + noise_pred_audio_uncond_stg = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_stg, i, audio_scheduler + ) + + video_stg_delta = self.stg_scale * (noise_pred_video - noise_pred_video_uncond_stg) + audio_stg_delta = self.audio_stg_scale * (noise_pred_audio - noise_pred_audio_uncond_stg) + else: + video_stg_delta = audio_stg_delta = 0 + + if self.do_modality_isolation_guidance: + with self.transformer.cache_context("uncond_modality"): + noise_pred_video_uncond_modality, noise_pred_audio_uncond_modality = self.transformer( + hidden_states=latents.to(dtype=prompt_embeds.dtype), + audio_hidden_states=audio_latents.to(dtype=prompt_embeds.dtype), + encoder_hidden_states=video_prompt_embeds, + audio_encoder_hidden_states=audio_prompt_embeds, + timestep=video_timestep, + audio_timestep=timestep, + sigma=timestep, # Used by LTX-2.3 + encoder_attention_mask=prompt_attn_mask, + audio_encoder_attention_mask=prompt_attn_mask, + num_frames=latent_num_frames, + height=latent_height, + width=latent_width, + fps=frame_rate, + audio_num_frames=audio_num_frames, + video_coords=video_pos_ids, + audio_coords=audio_pos_ids, + # Turn off A2V and V2A cross attn to isolate video and audio modalities + isolate_modalities=True, + spatio_temporal_guidance_blocks=None, + perturbation_mask=None, + use_cross_timestep=use_cross_timestep, + attention_kwargs=attention_kwargs, + return_dict=False, + ) + noise_pred_video_uncond_modality = noise_pred_video_uncond_modality.float() + noise_pred_audio_uncond_modality = noise_pred_audio_uncond_modality.float() + noise_pred_video_uncond_modality = self.convert_velocity_to_x0( + latents, noise_pred_video_uncond_modality, i, self.scheduler + ) + noise_pred_audio_uncond_modality = self.convert_velocity_to_x0( + audio_latents, noise_pred_audio_uncond_modality, i, audio_scheduler + ) + + video_modality_delta = (self.modality_scale - 1) * ( + noise_pred_video - noise_pred_video_uncond_modality + ) + audio_modality_delta = (self.audio_modality_scale - 1) * ( + noise_pred_audio - noise_pred_audio_uncond_modality + ) + else: + video_modality_delta = audio_modality_delta = 0 + + # Now apply all guidance terms + noise_pred_video_g = noise_pred_video + video_cfg_delta + video_stg_delta + video_modality_delta + noise_pred_audio_g = noise_pred_audio + audio_cfg_delta + audio_stg_delta + audio_modality_delta + + # Apply LTX-2.X guidance rescaling + if self.guidance_rescale > 0: + noise_pred_video = rescale_noise_cfg( + noise_pred_video_g, noise_pred_video, guidance_rescale=self.guidance_rescale + ) + else: + noise_pred_video = noise_pred_video_g + + if self.audio_guidance_rescale > 0: + noise_pred_audio = rescale_noise_cfg( + noise_pred_audio_g, noise_pred_audio, guidance_rescale=self.audio_guidance_rescale + ) + else: + noise_pred_audio = noise_pred_audio_g + + # compute the previous noisy sample x_t -> x_t-1 + noise_pred_video = self._unpack_latents( + noise_pred_video, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + # Convert back to velocity for scheduler + noise_pred_video = self.convert_x0_to_velocity(latents, noise_pred_video, i, self.scheduler) + noise_pred_audio = self.convert_x0_to_velocity(audio_latents, noise_pred_audio, i, audio_scheduler) + + noise_pred_video = noise_pred_video[:, :, 1:] + noise_latents = latents[:, :, 1:] + pred_latents = self.scheduler.step(noise_pred_video, t, noise_latents, return_dict=False)[0] + + latents = torch.cat([latents[:, :, :1], pred_latents], dim=2) + latents = self._pack_latents( + latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size + ) + + # NOTE: for now duplicate scheduler for audio latents in case self.scheduler sets internal state in + # the step method (such as _step_index) + audio_latents = audio_scheduler.step(noise_pred_audio, t, audio_latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + latents = self._unpack_latents( + latents, + latent_num_frames, + latent_height, + latent_width, + self.transformer_spatial_patch_size, + self.transformer_temporal_patch_size, + ) + + audio_latents = self._denormalize_audio_latents( + audio_latents, self.audio_vae.latents_mean, self.audio_vae.latents_std + ) + audio_latents = self._unpack_audio_latents(audio_latents, audio_num_frames, num_mel_bins=latent_mel_bins) + + if output_type == "latent": + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + video = latents + audio = audio_latents + else: + latents = latents.to(prompt_embeds.dtype) + + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + + latents = latents.to(self.vae.dtype) + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + audio_latents = audio_latents.to(self.audio_vae.dtype) + generated_mel_spectrograms = self.audio_vae.decode(audio_latents, return_dict=False)[0] + audio = self.vocoder(generated_mel_spectrograms) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video, audio) + + return LTX2PipelineOutput(frames=video, audio=audio) diff --git a/diffusers/pipelines/ltx2/pipeline_ltx2_latent_upsample.py b/diffusers/pipelines/ltx2/pipeline_ltx2_latent_upsample.py new file mode 100644 index 0000000000000000000000000000000000000000..4d085ca8f4ca78c0368e0e0a0579fd2e80ca0c25 --- /dev/null +++ b/diffusers/pipelines/ltx2/pipeline_ltx2_latent_upsample.py @@ -0,0 +1,425 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import torch + +from ...image_processor import PipelineImageInput +from ...models import AutoencoderKLLTX2Video +from ...utils import get_logger, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..ltx.pipeline_output import LTXPipelineOutput +from ..pipeline_utils import DiffusionPipeline +from .latent_upsampler import LTX2LatentUpsamplerModel + + +logger = get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import LTX2ImageToVideoPipeline, LTX2LatentUpsamplePipeline + >>> from diffusers.utils import encode_video + >>> from diffusers.pipelines.ltx2.latent_upsampler import LTX2LatentUpsamplerModel + >>> from diffusers.utils import load_image + + >>> pipe = LTX2ImageToVideoPipeline.from_pretrained("Lightricks/LTX-2", torch_dtype=torch.bfloat16) + >>> pipe.enable_model_cpu_offload() + + >>> image = load_image( + ... "https://huggingface.co/datasets/a-r-r-o-w/tiny-meme-dataset-captioned/resolve/main/images/8.png" + ... ) + >>> prompt = "A young girl stands calmly in the foreground, looking directly at the camera, as a house fire rages in the background." + >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted" + + >>> frame_rate = 24.0 + >>> video, audio = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... width=768, + ... height=512, + ... num_frames=121, + ... frame_rate=frame_rate, + ... num_inference_steps=40, + ... guidance_scale=4.0, + ... output_type="pil", + ... return_dict=False, + ... ) + + >>> latent_upsampler = LTX2LatentUpsamplerModel.from_pretrained( + ... "Lightricks/LTX-2", subfolder="latent_upsampler", torch_dtype=torch.bfloat16 + ... ) + >>> upsample_pipe = LTX2LatentUpsamplePipeline(vae=pipe.vae, latent_upsampler=latent_upsampler) + >>> upsample_pipe.vae.enable_tiling() + >>> upsample_pipe.to(device="cuda", dtype=torch.bfloat16) + + >>> video = upsample_pipe( + ... video=video, + ... width=768, + ... height=512, + ... output_type="np", + ... return_dict=False, + ... )[0] + + >>> encode_video( + ... video[0], + ... fps=frame_rate, + ... audio=audio[0].float().cpu(), + ... audio_sample_rate=pipe.vocoder.config.output_sampling_rate, # should be 24000 + ... output_path="video.mp4", + ... ) + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class LTX2LatentUpsamplePipeline(DiffusionPipeline): + model_cpu_offload_seq = "vae->latent_upsampler" + + def __init__( + self, + vae: AutoencoderKLLTX2Video, + latent_upsampler: LTX2LatentUpsamplerModel, + ) -> None: + super().__init__() + + self.register_modules(vae=vae, latent_upsampler=latent_upsampler) + + self.vae_spatial_compression_ratio = ( + self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32 + ) + self.vae_temporal_compression_ratio = ( + self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8 + ) + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio) + + def prepare_latents( + self, + video: torch.Tensor | None = None, + batch_size: int = 1, + num_frames: int = 121, + height: int = 512, + width: int = 768, + spatial_patch_size: int = 1, + temporal_patch_size: int = 1, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + if latents.ndim == 3: + # Convert token seq [B, S, D] to latent video [B, C, F, H, W] + latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1 + latent_height = height // self.vae_spatial_compression_ratio + latent_width = width // self.vae_spatial_compression_ratio + latents = self._unpack_latents( + latents, latent_num_frames, latent_height, latent_width, spatial_patch_size, temporal_patch_size + ) + return latents.to(device=device, dtype=dtype) + + video = video.to(device=device, dtype=self.vae.dtype) + if isinstance(generator, list): + if len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + init_latents = [ + retrieve_latents(self.vae.encode(video[i].unsqueeze(0)), generator[i]) for i in range(batch_size) + ] + else: + init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), generator) for vid in video] + + init_latents = torch.cat(init_latents, dim=0).to(dtype) + # NOTE: latent upsampler operates on the unnormalized latents, so don't normalize here + # init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std) + return init_latents + + def adain_filter_latent(self, latents: torch.Tensor, reference_latents: torch.Tensor, factor: float = 1.0): + """ + Applies Adaptive Instance Normalization (AdaIN) to a latent tensor based on statistics from a reference latent + tensor. + + Args: + latent (`torch.Tensor`): + Input latents to normalize + reference_latents (`torch.Tensor`): + The reference latents providing style statistics. + factor (`float`): + Blending factor between original and transformed latent. Range: -10.0 to 10.0, Default: 1.0 + + Returns: + torch.Tensor: The transformed latent tensor + """ + result = latents.clone() + + for i in range(latents.size(0)): + for c in range(latents.size(1)): + r_sd, r_mean = torch.std_mean(reference_latents[i, c], dim=None) # index by original dim order + i_sd, i_mean = torch.std_mean(result[i, c], dim=None) + + result[i, c] = ((result[i, c] - i_mean) / i_sd) * r_sd + r_mean + + result = torch.lerp(latents, result, factor) + return result + + def tone_map_latents(self, latents: torch.Tensor, compression: float) -> torch.Tensor: + """ + Applies a non-linear tone-mapping function to latent values to reduce their dynamic range in a perceptually + smooth way using a sigmoid-based compression. + + This is useful for regularizing high-variance latents or for conditioning outputs during generation, especially + when controlling dynamic behavior with a `compression` factor. + + Args: + latents : torch.Tensor + Input latent tensor with arbitrary shape. Expected to be roughly in [-1, 1] or [0, 1] range. + compression : float + Compression strength in the range [0, 1]. + - 0.0: No tone-mapping (identity transform) + - 1.0: Full compression effect + + Returns: + torch.Tensor + The tone-mapped latent tensor of the same shape as input. + """ + # Remap [0-1] to [0-0.75] and apply sigmoid compression in one shot + scale_factor = compression * 0.75 + abs_latents = torch.abs(latents) + + # Sigmoid compression: sigmoid shifts large values toward 0.2, small values stay ~1.0 + # When scale_factor=0, sigmoid term vanishes, when scale_factor=0.75, full effect + sigmoid_term = torch.sigmoid(4.0 * scale_factor * (abs_latents - 1.0)) + scales = 1.0 - 0.8 * scale_factor * sigmoid_term + + filtered = latents * scales + return filtered + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._denormalize_latents + def _denormalize_latents( + latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0 + ) -> torch.Tensor: + # Denormalize latents across the channel dimension [B, C, F, H, W] + latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents * latents_std / scaling_factor + latents_mean + return latents + + @staticmethod + # Copied from diffusers.pipelines.ltx2.pipeline_ltx2.LTX2Pipeline._unpack_latents + def _unpack_latents( + latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1 + ) -> torch.Tensor: + # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions) + # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of + # what happens in the `_pack_latents` method. + batch_size = latents.size(0) + latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size) + latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3) + return latents + + def check_inputs(self, video, height, width, latents, tone_map_compression_ratio): + if height % self.vae_spatial_compression_ratio != 0 or width % self.vae_spatial_compression_ratio != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if video is not None and latents is not None: + raise ValueError("Only one of `video` or `latents` can be provided.") + if video is None and latents is None: + raise ValueError("One of `video` or `latents` has to be provided.") + + if not (0 <= tone_map_compression_ratio <= 1): + raise ValueError("`tone_map_compression_ratio` must be in the range [0, 1]") + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + video: list[PipelineImageInput] | None = None, + height: int = 512, + width: int = 768, + num_frames: int = 121, + spatial_patch_size: int = 1, + temporal_patch_size: int = 1, + latents: torch.Tensor | None = None, + latents_normalized: bool = False, + decode_timestep: float | list[float] = 0.0, + decode_noise_scale: float | list[float] | None = None, + adain_factor: float = 0.0, + tone_map_compression_ratio: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + video (`list[PipelineImageInput]`, *optional*) + The video to be upsampled (such as a LTX 2.0 first stage output). If not supplied, `latents` should be + supplied. + height (`int`, *optional*, defaults to `512`): + The height in pixels of the input video (not the generated video, which will have a larger resolution). + width (`int`, *optional*, defaults to `768`): + The width in pixels of the input video (not the generated video, which will have a larger resolution). + num_frames (`int`, *optional*, defaults to `121`): + The number of frames in the input video. + spatial_patch_size (`int`, *optional*, defaults to `1`): + The spatial patch size of the video latents. Used when `latents` is supplied if unpacking is necessary. + temporal_patch_size (`int`, *optional*, defaults to `1`): + The temporal patch size of the video latents. Used when `latents` is supplied if unpacking is + necessary. + latents (`torch.Tensor`, *optional*): + Pre-generated video latents. This can be supplied in place of the `video` argument. Can either be a + patch sequence of shape `(batch_size, seq_len, hidden_dim)` or a video latent of shape `(batch_size, + latent_channels, latent_frames, latent_height, latent_width)`. + latents_normalized (`bool`, *optional*, defaults to `False`) + If `latents` are supplied, whether the `latents` are normalized using the VAE latent mean and std. If + `True`, the `latents` will be denormalized before being supplied to the latent upsampler. + decode_timestep (`float`, defaults to `0.0`): + The timestep at which generated video is decoded. + decode_noise_scale (`float`, defaults to `None`): + The interpolation factor between random noise and denoised latents at the decode timestep. + adain_factor (`float`, *optional*, defaults to `0.0`): + Adaptive Instance Normalization (AdaIN) blending factor between the upsampled and original latents. + Should be in [-10.0, 10.0]; supplying 0.0 (the default) means that AdaIN is not performed. + tone_map_compression_ratio (`float`, *optional*, defaults to `0.0`): + The compression strength for tone mapping, which will reduce the dynamic range of the latent values. + This is useful for regularizing high-variance latents or for conditioning outputs during generation. + Should be in [0, 1], where 0.0 (the default) means tone mapping is not applied and 1.0 corresponds to + the full compression effect. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple. + + Examples: + + Returns: + [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is the upsampled video. + """ + + self.check_inputs( + video=video, + height=height, + width=width, + latents=latents, + tone_map_compression_ratio=tone_map_compression_ratio, + ) + + if video is not None: + # Batched video input is not yet tested/supported. TODO: take a look later + batch_size = 1 + else: + batch_size = latents.shape[0] + device = self._execution_device + + if video is not None: + num_frames = len(video) + if num_frames % self.vae_temporal_compression_ratio != 1: + num_frames = ( + num_frames // self.vae_temporal_compression_ratio * self.vae_temporal_compression_ratio + 1 + ) + video = video[:num_frames] + logger.warning( + f"Video length expected to be of the form `k * {self.vae_temporal_compression_ratio} + 1` but is {len(video)}. Truncating to {num_frames} frames." + ) + video = self.video_processor.preprocess_video(video, height=height, width=width) + video = video.to(device=device, dtype=torch.float32) + + latents_supplied = latents is not None + latents = self.prepare_latents( + video=video, + batch_size=batch_size, + num_frames=num_frames, + height=height, + width=width, + spatial_patch_size=spatial_patch_size, + temporal_patch_size=temporal_patch_size, + dtype=torch.float32, + device=device, + generator=generator, + latents=latents, + ) + + if latents_supplied and latents_normalized: + latents = self._denormalize_latents( + latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor + ) + latents = latents.to(self.latent_upsampler.dtype) + latents_upsampled = self.latent_upsampler(latents) + + if adain_factor > 0.0: + latents = self.adain_filter_latent(latents_upsampled, latents, adain_factor) + else: + latents = latents_upsampled + + if tone_map_compression_ratio > 0.0: + latents = self.tone_map_latents(latents, tone_map_compression_ratio) + + if output_type == "latent": + video = latents + else: + if not self.vae.config.timestep_conditioning: + timestep = None + else: + noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype) + if not isinstance(decode_timestep, list): + decode_timestep = [decode_timestep] * batch_size + if decode_noise_scale is None: + decode_noise_scale = decode_timestep + elif not isinstance(decode_noise_scale, list): + decode_noise_scale = [decode_noise_scale] * batch_size + + timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype) + decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[ + :, None, None, None, None + ] + latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise + + video = self.vae.decode(latents, timestep, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return LTXPipelineOutput(frames=video) diff --git a/diffusers/pipelines/ltx2/pipeline_output.py b/diffusers/pipelines/ltx2/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..02891219ee6768bb0a76b2a36c11e3541b51081d --- /dev/null +++ b/diffusers/pipelines/ltx2/pipeline_output.py @@ -0,0 +1,23 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class LTX2PipelineOutput(BaseOutput): + r""" + Output class for LTX pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + audio (`torch.Tensor`, `np.ndarray`): + TODO + """ + + frames: torch.Tensor + audio: torch.Tensor diff --git a/diffusers/pipelines/ltx2/utils.py b/diffusers/pipelines/ltx2/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..52d446c4681967a14c6796f6d552f10337454bad --- /dev/null +++ b/diffusers/pipelines/ltx2/utils.py @@ -0,0 +1,155 @@ +# Copyright 2026 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# Pre-trained sigma values for distilled model are taken from +# https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py +DISTILLED_SIGMA_VALUES = [1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875] + +# Reduced schedule for super-resolution stage 2 (subset of distilled values) +STAGE_2_DISTILLED_SIGMA_VALUES = [0.909375, 0.725, 0.421875] + + +# Default negative prompt from +# https://github.com/Lightricks/LTX-2/blob/ae855f8538843825f9015a419cf4ba5edaf5eec2/packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py#L131-L143 +DEFAULT_NEGATIVE_PROMPT = ( + "blurry, out of focus, overexposed, underexposed, low contrast, washed out colors, excessive noise, " + "grainy texture, poor lighting, flickering, motion blur, distorted proportions, unnatural skin tones, " + "deformed facial features, asymmetrical face, missing facial features, extra limbs, disfigured hands, " + "wrong hand count, artifacts around text, inconsistent perspective, camera shake, incorrect depth of " + "field, background too sharp, background clutter, distracting reflections, harsh shadows, inconsistent " + "lighting direction, color banding, cartoonish rendering, 3D CGI look, unrealistic materials, uncanny " + "valley effect, incorrect ethnicity, wrong gender, exaggerated expressions, wrong gaze direction, " + "mismatched lip sync, silent or muted audio, distorted voice, robotic voice, echo, background noise, " + "off-sync audio, incorrect dialogue, added dialogue, repetitive speech, jittery movement, awkward " + "pauses, incorrect timing, unnatural transitions, inconsistent framing, tilted camera, flat lighting, " + "inconsistent tone, cinematic oversaturation, stylized filters, or AI artifacts." +) + + +# System prompts for prompt enhancement +# https://github.com/Lightricks/LTX-2/blob/ae855f8538843825f9015a419cf4ba5edaf5eec2/packages/ltx-core/src/ltx_core/text_encoders/gemma/encoders/prompts/gemma_t2v_system_prompt.txt#L1 +# Disable line-too-long rule in ruff to keep the prompts exactly the same (e.g. in terms of newlines) +# Supported in ruff>=0.15.0 +# ruff: disable[E501] +T2V_DEFAULT_SYSTEM_PROMPT = """ +You are a Creative Assistant. Given a user's raw input prompt describing a scene or concept, expand it into a detailed +video generation prompt with specific visuals and integrated audio to guide a text-to-video model. + +#### Guidelines +- Strictly follow all aspects of the user's raw input: include every element requested (style, visuals, motions, + actions, camera movement, audio). + - If the input is vague, invent concrete details: lighting, textures, materials, scene settings, etc. + - For characters: describe gender, clothing, hair, expressions. DO NOT invent unrequested characters. +- Use active language: present-progressive verbs ("is walking," "speaking"). If no action specified, describe natural + movements. +- Maintain chronological flow: use temporal connectors ("as," "then," "while"). +- Audio layer: Describe complete soundscape (background audio, ambient sounds, SFX, speech/music when requested). + Integrate sounds chronologically alongside actions. Be specific (e.g., "soft footsteps on tile"), not vague (e.g., + "ambient sound is present"). +- Speech (only when requested): + - For ANY speech-related input (talking, conversation, singing, etc.), ALWAYS include exact words in quotes with + voice characteristics (e.g., "The man says in an excited voice: 'You won't believe what I just saw!'"). + - Specify language if not English and accent if relevant. +- Style: Include visual style at the beginning: "Style: