app.py

import gradio as gr
import numpy as np
import random
import torch
import spaces

from PIL import Image
from diffusers import FlowMatchEulerDiscreteScheduler, QwenImageEditPlusPipeline
# from optimization import optimize_pipeline_
# from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
# from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
# from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3

from huggingface_hub import InferenceClient
import math

import os
import base64
from io import BytesIO
import json

SYSTEM_PROMPT = '''
# Edit Instruction Rewriter
You are a professional edit instruction rewriter. Your task is to generate a precise, concise, and visually achievable professional-level edit instruction based on the user-provided instruction and the image to be edited.  

Please strictly follow the rewriting rules below:

## 1. General Principles
- Keep the rewritten prompt **concise and comprehensive**. Avoid overly long sentences and unnecessary descriptive language.  
- If the instruction is contradictory, vague, or unachievable, prioritize reasonable inference and correction, and supplement details when necessary.  
- Keep the main part of the original instruction unchanged, only enhancing its clarity, rationality, and visual feasibility.  
- All added objects or modifications must align with the logic and style of the scene in the input images.  
- If multiple sub-images are to be generated, describe the content of each sub-image individually.  

## 2. Task-Type Handling Rules

### 1. Add, Delete, Replace Tasks
- If the instruction is clear (already includes task type, target entity, position, quantity, attributes), preserve the original intent and only refine the grammar.  
- If the description is vague, supplement with minimal but sufficient details (category, color, size, orientation, position, etc.). For example:  
    > Original: "Add an animal"  
    > Rewritten: "Add a light-gray cat in the bottom-right corner, sitting and facing the camera"  
- Remove meaningless instructions: e.g., "Add 0 objects" should be ignored or flagged as invalid.  
- For replacement tasks, specify "Replace Y with X" and briefly describe the key visual features of X.  

### 2. Text Editing Tasks
- All text content must be enclosed in English double quotes `" "`. Keep the original language of the text, and keep the capitalization.  
- Both adding new text and replacing existing text are text replacement tasks, For example:  
    - Replace "xx" to "yy"  
    - Replace the mask / bounding box to "yy"  
    - Replace the visual object to "yy"  
- Specify text position, color, and layout only if user has required.  
- If font is specified, keep the original language of the font.  

### 3. Human Editing Tasks
- Make the smallest changes to the given user's prompt.  
- If changes to background, action, expression, camera shot, or ambient lighting are required, please list each modification individually.
- **Edits to makeup or facial features / expression must be subtle, not exaggerated, and must preserve the subject's identity consistency.**
    > Original: "Add eyebrows to the face"  
    > Rewritten: "Slightly thicken the person's eyebrows with little change, look natural."

### 4. Style Conversion or Enhancement Tasks
- If a style is specified, describe it concisely using key visual features. For example:  
    > Original: "Disco style"  
    > Rewritten: "1970s disco style: flashing lights, disco ball, mirrored walls, vibrant colors"  
- For style reference, analyze the original image and extract key characteristics (color, composition, texture, lighting, artistic style, etc.), integrating them into the instruction.  
- **Colorization tasks (including old photo restoration) must use the fixed template:**  
  "Restore and colorize the old photo."  
- Clearly specify the object to be modified. For example:  
    > Original: Modify the subject in Picture 1 to match the style of Picture 2.  
    > Rewritten: Change the girl in Picture 1 to the ink-wash style of Picture 2 — rendered in black-and-white watercolor with soft color transitions.

### 5. Material Replacement
- Clearly specify the object and the material. For example: "Change the material of the apple to papercut style."
- For text material replacement, use the fixed template:
    "Change the material of text "xxxx" to laser style"

### 6. Logo/Pattern Editing
- Material replacement should preserve the original shape and structure as much as possible. For example:
   > Original: "Convert to sapphire material"  
   > Rewritten: "Convert the main subject in the image to sapphire material, preserving similar shape and structure"
- When migrating logos/patterns to new scenes, ensure shape and structure consistency. For example:
   > Original: "Migrate the logo in the image to a new scene"  
   > Rewritten: "Migrate the logo in the image to a new scene, preserving similar shape and structure"

### 7. Multi-Image Tasks
- Rewritten prompts must clearly point out which image's element is being modified. For example:  
    > Original: "Replace the subject of picture 1 with the subject of picture 2"  
    > Rewritten: "Replace the girl of picture 1 with the boy of picture 2, keeping picture 2's background unchanged"  
- For stylization tasks, describe the reference image's style in the rewritten prompt, while preserving the visual content of the source image.  

## 3. Rationale and Logic Check
- Resolve contradictory instructions: e.g., "Remove all trees but keep all trees" requires logical correction.
- Supplement missing critical information: e.g., if position is unspecified, choose a reasonable area based on composition (near subject, blank space, center/edge, etc.).

# Output Format Example
```json
{
   "Rewritten": "..."
}
'''

def polish_prompt_hf(original_prompt, img_list):
    """
    Rewrites the prompt using a Hugging Face InferenceClient.
    Supports multiple images via img_list.
    """
    # Ensure HF_TOKEN is set
    api_key = os.environ.get("inference_providers")
    if not api_key:
        print("Warning: HF_TOKEN not set. Falling back to original prompt.")
        return original_prompt
    prompt = f"{SYSTEM_PROMPT}\n\nUser Input: {original_prompt}\n\nRewritten Prompt:"
    system_prompt = "you are a helpful assistant, you should provide useful answers to users."
    try:
        # Initialize the client
        client = InferenceClient(
            provider="nebius",
            api_key=api_key,
        )

        # Convert list of images to base64 data URLs
        image_urls = []
        if img_list is not None:
            # Ensure img_list is actually a list
            if not isinstance(img_list, list):
                img_list = [img_list]
            
            for img in img_list:
                image_url = None
                # If img is a PIL Image
                if hasattr(img, 'save'):  # Check if it's a PIL Image
                    buffered = BytesIO()
                    img.save(buffered, format="PNG")
                    img_base64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
                    image_url = f"data:image/png;base64,{img_base64}"
                # If img is already a file path (string)
                elif isinstance(img, str):
                    with open(img, "rb") as image_file:
                        img_base64 = base64.b64encode(image_file.read()).decode('utf-8')
                    image_url = f"data:image/png;base64,{img_base64}"
                else:
                    print(f"Warning: Unexpected image type: {type(img)}, skipping...")
                    continue
                
                if image_url:
                    image_urls.append(image_url)

        # Build the content array with text first, then all images
        content = [
            {
                "type": "text",
                "text": prompt
            }
        ]
        
        # Add all images to the content
        for image_url in image_urls:
            content.append({
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
            })

        # Format the messages for the chat completions API
        messages = [
            {"role": "system", "content": system_prompt},
            {
                "role": "user",
                "content": content
            }
        ]

        # Call the API
        completion = client.chat.completions.create(
            model="Qwen/Qwen2.5-VL-72B-Instruct",
            messages=messages,
        )
        
        # Parse the response
        result = completion.choices[0].message.content
        
        # Try to extract JSON if present
        if '"Rewritten"' in result:
            try:
                # Clean up the response
                result = result.replace('```json', '').replace('```', '')
                result_json = json.loads(result)
                polished_prompt = result_json.get('Rewritten', result)
            except:
                polished_prompt = result
        else:
            polished_prompt = result
            
        polished_prompt = polished_prompt.strip().replace("\n", " ")
        return polished_prompt
        
    except Exception as e:
        print(f"Error during API call to Hugging Face: {e}")
        # Fallback to original prompt if enhancement fails
        return original_prompt 



def encode_image(pil_image):
    import io
    buffered = io.BytesIO()
    pil_image.save(buffered, format="PNG")
    return base64.b64encode(buffered.getvalue()).decode("utf-8")

# --- Model Loading ---
dtype = torch.bfloat16
device = "cuda" if torch.cuda.is_available() else "cpu"

# Scheduler configuration for Lightning
scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

# Initialize scheduler with Lightning config
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

# Load the model pipeline
pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2511", 
                                                 scheduler=scheduler,
                                                 torch_dtype=dtype).to(device)
pipe.load_lora_weights(
        "lightx2v/Qwen-Image-Edit-2511-Lightning", 
        weight_name="Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors",adapter_name="fast"
)
pipe.load_lora_weights(
        "lilylilith/AnyPose", 
        weight_name="2511-AnyPose-base-000006250.safetensors",adapter_name="base"
)
pipe.load_lora_weights(
        "lilylilith/AnyPose", 
        weight_name="2511-AnyPose-helper-00006000.safetensors",adapter_name="helper"
)
pipe.set_adapters(["fast", "base", "helper"], adapter_weights=[1., 0.7, 0.7])
pipe.fuse_lora(adapter_names=["fast", "base", "helper"], lora_scale=1)
pipe.unload_lora_weights()

# # Apply the same optimizations from the first version
# pipe.transformer.__class__ = QwenImageTransformer2DModel
# pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())

# # --- Ahead-of-time compilation ---
# optimize_pipeline_(pipe, image=[Image.new("RGB", (1024, 1024)), Image.new("RGB", (1024, 1024))], prompt="prompt")

# --- UI Constants and Helpers ---
MAX_SEED = np.iinfo(np.int32).max

DEFAULT_LORA_PROMPT = """
Make the person in image 1 do the exact same pose of the person in image 2. 
Changing the style and background of the image of the person in image 1 is undesirable, so don't do it. 
The new pose should be pixel accurate to the pose we are trying to copy. 
The position of the arms and head and legs should be the same as the pose we are trying to copy. 
Change the field of view and angle to match exactly image 2. Head tilt and eye gaze pose should match the person in image 2. 
Remove the background of image 2, and replace it with the background of image 1.
"""

def use_output_as_input(output_images):
    """Convert output images to input format for the reference image"""
    if output_images is None or len(output_images) == 0:
        return None
    return output_images[0]

# --- Main Inference Function (with hardcoded negative prompt) ---
@spaces.GPU()
def infer(
    reference_image,
    pose_image,
    prompt,
    seed=42,
    randomize_seed=False,
    true_guidance_scale=1.0,
    num_inference_steps=4,
    height=None,
    width=None,
    rewrite_prompt=True,
    num_images_per_prompt=1,
    progress=gr.Progress(track_tqdm=True),
):
    """
    Run image-editing inference using the Qwen-Image-Edit pipeline.

    Parameters:
        reference_image: Reference image (PIL or path-based).
        pose_image: Pose image (PIL or path-based).
        prompt (str): Editing instruction (may be rewritten by LLM if enabled).
        seed (int): Random seed for reproducibility.
        randomize_seed (bool): If True, overrides seed with a random value.
        true_guidance_scale (float): CFG scale used by Qwen-Image.
        num_inference_steps (int): Number of diffusion steps.
        height (int | None): Optional output height override.
        width (int | None): Optional output width override.
        rewrite_prompt (bool): Whether to rewrite the prompt using Qwen-2.5-VL.
        num_images_per_prompt (int): Number of images to generate.
        progress: Gradio progress callback.

    Returns:
        tuple: (generated_images, seed_used, UI_visibility_update)
    """
    
    # Hardcode the negative prompt as requested
    negative_prompt = " "
    
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    # Set up the generator for reproducibility
    generator = torch.Generator(device=device).manual_seed(seed)
    
    # Load input images into PIL Images
    pil_images = []
    
    # Process reference image (first)
    if reference_image is not None:
        try:
            if isinstance(reference_image, Image.Image):
                pil_images.append(reference_image.convert("RGB"))
            elif isinstance(reference_image, str):
                pil_images.append(Image.open(reference_image).convert("RGB"))
            elif hasattr(reference_image, "name"):
                pil_images.append(Image.open(reference_image.name).convert("RGB"))
        except Exception:
            pass
    
    # Process pose image (second)
    if pose_image is not None:
        try:
            if isinstance(pose_image, Image.Image):
                pil_images.append(pose_image.convert("RGB"))
            elif isinstance(pose_image, str):
                pil_images.append(Image.open(pose_image).convert("RGB"))
            elif hasattr(pose_image, "name"):
                pil_images.append(Image.open(pose_image.name).convert("RGB"))
        except Exception:
            pass

    if height==256 and width==256:
        height, width = None, None
    print(f"Calling pipeline with prompt: '{prompt}'")
    print(f"Negative Prompt: '{negative_prompt}'")
    print(f"Seed: {seed}, Steps: {num_inference_steps}, Guidance: {true_guidance_scale}, Size: {width}x{height}")
    if rewrite_prompt and len(pil_images) > 0:
        prompt = polish_prompt_hf(prompt, pil_images)
        print(f"Rewritten Prompt: {prompt}")
    

    # Generate the image
    image = pipe(
        image=pil_images if len(pil_images) > 0 else None,
        prompt=prompt,
        height=height,
        width=width,
        negative_prompt=negative_prompt,
        num_inference_steps=num_inference_steps,
        generator=generator,
        true_cfg_scale=true_guidance_scale,
        num_images_per_prompt=num_images_per_prompt,
    ).images

    # Return images, seed, and make button visible
    return image, seed, gr.update(visible=False)

# --- Examples and UI Layout ---
examples = []

css = """
#col-container {
    margin: 0 auto;
    max-width: 1024px;
}
#logo-title {
    text-align: center;
}
#logo-title img {
    width: 400px;
}
#edit_text{margin-top: -62px !important}
"""

with gr.Blocks(css=css, theme=gr.themes.Citrus()) as demo:
    with gr.Column(elem_id="col-container"):
        gr.HTML("""
        <div id="logo-title">
            <h1 style="color: #5b47d1;">Qwen Edit Any Pose 🕺 </h1>
            <h2 style="font-style: italic;color: #5b47d1;margin-top: -10px !important;">Fast 4-step pose transfer with Qwen Edit 2511 & AnyPose LoRA</h2>
        </div>
        """)
        gr.Markdown("""
        Transfer any pose from a reference image to your subject using [Qwen-Image-Edit-2511](https://huggingface.co/Qwen/Qwen-Image-Edit-2511) with [lilylilith/AnyPose LoRA](https://huggingface.co/lilylilith/AnyPose) and [lightx2v Lightning LoRA](https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning) for fast inference.
        [Learn more](https://github.com/QwenLM/Qwen-Image) about the Qwen-Image series.
        """)
        with gr.Row():
            with gr.Column(scale=1):
                with gr.Row():
                    reference_image = gr.Image(label="Reference Image", type="pil", interactive=True)
                    pose_image = gr.Image(label="Pose Image", type="pil", interactive=True)
                
                with gr.Row():
                    prompt = gr.Text(
                            label="Prompt",
                            value=DEFAULT_LORA_PROMPT,
                            show_label=False,
                            visible=False
                    )
                    run_button = gr.Button("Edit Pose", variant="primary")

            with gr.Column(scale=1):
                result = gr.Gallery(label="Result", show_label=False, type="pil", interactive=False)
                # Add this button right after the result gallery - initially hidden
                use_output_btn = gr.Button("↗️ Use as input", variant="secondary", size="sm", visible=False)

        with gr.Accordion("Advanced Settings", open=False):
            # Negative prompt UI element is removed here

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )

            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

            with gr.Row():

                true_guidance_scale = gr.Slider(
                    label="True guidance scale",
                    minimum=1.0,
                    maximum=10.0,
                    step=0.1,
                    value=1.0
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=40,
                    step=1,
                    value=4,
                )
                
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=2048,
                    step=8,
                    value=None,
                )
                
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=2048,
                    step=8,
                    value=None,
                )
                
                
                rewrite_prompt = gr.Checkbox(label="Rewrite prompt", value=False, visible=False)

        gr.Examples(
        examples=[
            ["s-l1200.jpg","High-Lunge_Andrew-Clark.jpg"],
        ],
        inputs=[
            reference_image,
            pose_image,
        ],
        outputs=[result, seed, use_output_btn],
        fn=infer,
        cache_examples=True,
        cache_mode="lazy",
    )
    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[
            reference_image,
            pose_image,
            prompt,
            seed,
            randomize_seed,
            true_guidance_scale,
            num_inference_steps,
            height,
            width,
            rewrite_prompt,
        ],
        outputs=[result, seed, use_output_btn],  # Added use_output_btn to outputs
    )

    # Add the new event handler for the "Use Output as Input" button
    use_output_btn.click(
        fn=use_output_as_input,
        inputs=[result],
        outputs=[reference_image]
    )

if __name__ == "__main__":
    demo.launch(mcp_server=True)