# deformes4D_engine.py # Copyright (C) 4 de Agosto de 2025 Carlos Rodrigues dos Santos # # MODIFICATIONS FOR ADUC-SDR: # Copyright (C) 2025 Carlos Rodrigues dos Santos. All rights reserved. # # This file is part of the ADUC-SDR project. It contains the core logic for # video fragment generation, latent manipulation, and dynamic editing, # governed by the ADUC orchestrator. # This component is licensed under the GNU Affero General Public License v3.0. import os import time import imageio import numpy as np import torch import logging from PIL import Image, ImageOps from dataclasses import dataclass import gradio as gr import subprocess import gc from ltx_manager_helpers import ltx_manager_singleton from gemini_helpers import gemini_singleton from upscaler_specialist import upscaler_specialist_singleton from hd_specialist import hd_specialist_singleton from ltx_video.models.autoencoders.causal_video_autoencoder import CausalVideoAutoencoder from ltx_video.models.autoencoders.vae_encode import vae_encode, vae_decode logger = logging.getLogger(__name__) @dataclass class LatentConditioningItem: """Representa uma âncora de condicionamento no espaço latente para a Câmera (Ψ).""" latent_tensor: torch.Tensor media_frame_number: int conditioning_strength: float class Deformes4DEngine: """ Implementa a Câmera (Ψ) e o Destilador (Δ) da arquitetura ADUC-SDR. Orquestra a geração, pós-produção latente e renderização final dos fragmentos de vídeo. """ def __init__(self, ltx_manager, workspace_dir="deformes_workspace"): self.ltx_manager = ltx_manager self.workspace_dir = workspace_dir self._vae = None self.device = 'cuda' if torch.cuda.is_available() else 'cpu' logger.info("Especialista Deformes4D (Executor ADUC-SDR) inicializado.") @property def vae(self): if self._vae is None: self._vae = self.ltx_manager.workers[0].pipeline.vae self._vae.to(self.device); self._vae.eval() return self._vae # --- MÉTODOS AUXILIARES --- @torch.no_grad() def pixels_to_latents(self, tensor: torch.Tensor) -> torch.Tensor: tensor = tensor.to(self.device, dtype=self.vae.dtype) return vae_encode(tensor, self.vae, vae_per_channel_normalize=True) @torch.no_grad() def latents_to_pixels(self, latent_tensor: torch.Tensor, decode_timestep: float = 0.05) -> torch.Tensor: latent_tensor = latent_tensor.to(self.device, dtype=self.vae.dtype) timestep_tensor = torch.tensor([decode_timestep] * latent_tensor.shape[0], device=self.device, dtype=latent_tensor.dtype) return vae_decode(latent_tensor, self.vae, is_video=True, timestep=timestep_tensor, vae_per_channel_normalize=True) def _preprocess_image_for_latent_conversion(self, image: Image.Image, target_resolution: tuple) -> Image.Image: if image.size != target_resolution: return ImageOps.fit(image, target_resolution, Image.Resampling.LANCZOS) return image def pil_to_latent(self, pil_image: Image.Image) -> torch.Tensor: image_np = np.array(pil_image).astype(np.float32) / 255.0 tensor = torch.from_numpy(image_np).permute(2, 0, 1).unsqueeze(0).unsqueeze(2) tensor = (tensor * 2.0) - 1.0 return self.pixels_to_latents(tensor) # --- NÚCLEO DA LÓGICA ADUC-SDR --- def generate_full_movie(self, keyframes: list, global_prompt: str, storyboard: list, seconds_per_fragment: float, trim_percent: int, handler_strength: float, destination_convergence_strength: float, use_upscaler: bool, use_refiner: bool, use_hd: bool, use_audio: bool, video_resolution: int, use_continuity_director: bool, progress: gr.Progress = gr.Progress()): num_transitions_to_generate = len(keyframes) - 1 TOTAL_STEPS = num_transitions_to_generate + 4 current_step = 0 FPS = 24 FRAMES_PER_LATENT_CHUNK = 8 ECO_LATENT_CHUNKS = 2 total_frames_brutos = self._quantize_to_multiple(int(seconds_per_fragment * FPS), FRAMES_PER_LATENT_CHUNK) frames_a_podar = self._quantize_to_multiple(int(total_frames_brutos * (trim_percent / 100)), FRAMES_PER_LATENT_CHUNK) latents_a_podar = frames_a_podar // FRAMES_PER_LATENT_CHUNK DEJAVU_FRAME_TARGET = frames_a_podar - 1 if frames_a_podar > 0 else 0 DESTINATION_FRAME_TARGET = total_frames_brutos - 1 base_ltx_params = {"guidance_scale": 2.0, "stg_scale": 0.025, "rescaling_scale": 0.15, "num_inference_steps": 20} # --- [INÍCIO DA CORREÇÃO] --- # Lógica robusta para extrair caminhos de arquivo da lista de keyframes. keyframe_paths = [] for item in keyframes: if isinstance(item, str): keyframe_paths.append(item) elif isinstance(item, tuple) and len(item) > 0: keyframe_paths.append(item[0]) # Assume que o caminho está no primeiro elemento da tupla elif hasattr(item, 'name'): keyframe_paths.append(item.name) else: logger.warning(f"Item na lista de keyframes com tipo inesperado e sem atributo '.name': {type(item)}") # --- [FIM DA CORREÇÃO] --- story_history = "" eco_latent_for_next_loop = None dejavu_latent_for_next_loop = None raw_latent_fragments = [] # --- ATO I: GERAÇÃO CAUSAL PURA (LOOP DE FRAGMENTOS) --- for i in range(num_transitions_to_generate): fragment_index = i + 1 current_step += 1 progress(current_step / TOTAL_STEPS, desc=f"Gerando Fragmento Causal {fragment_index}/{num_transitions_to_generate}") past_keyframe_path = keyframe_paths[i - 1] if i > 0 else keyframe_paths[i] start_keyframe_path = keyframe_paths[i] destination_keyframe_path = keyframe_paths[i + 1] future_story_prompt = storyboard[i + 1] if (i + 1) < len(storyboard) else "A cena final." decision = gemini_singleton.get_cinematic_decision( global_prompt, story_history, past_keyframe_path, start_keyframe_path, destination_keyframe_path, storyboard[i - 1] if i > 0 else "O início.", storyboard[i], future_story_prompt) transition_type, motion_prompt = decision["transition_type"], decision["motion_prompt"] story_history += f"\n- Ato {fragment_index}: {motion_prompt}" expected_height, expected_width = video_resolution, video_resolution downscale_factor = 2 / 3 downscaled_height = self._quantize_to_multiple(int(expected_height * downscale_factor), 8) downscaled_width = self._quantize_to_multiple(int(expected_width * downscale_factor), 8) target_resolution_tuple = (downscaled_height, downscaled_width) conditioning_items = [] if eco_latent_for_next_loop is None: img_start = self._preprocess_image_for_latent_conversion(Image.open(start_keyframe_path).convert("RGB"), target_resolution_tuple) conditioning_items.append(LatentConditioningItem(self.pil_to_latent(img_start), 0, 1.0)) else: conditioning_items.append(LatentConditioningItem(eco_latent_for_next_loop, 0, 1.0)) conditioning_items.append(LatentConditioningItem(dejavu_latent_for_next_loop, DEJAVU_FRAME_TARGET, handler_strength)) img_dest = self._preprocess_image_for_latent_conversion(Image.open(destination_keyframe_path).convert("RGB"), target_resolution_tuple) conditioning_items.append(LatentConditioningItem(self.pil_to_latent(img_dest), DESTINATION_FRAME_TARGET, destination_convergence_strength)) current_ltx_params = {**base_ltx_params, "motion_prompt": motion_prompt} latents_brutos, _ = self._generate_latent_tensor_internal(conditioning_items, current_ltx_params, target_resolution_tuple, total_frames_brutos) last_trim = latents_brutos[:, :, -(latents_a_podar+1):, :, :].clone() eco_latent_for_next_loop = last_trim[:, :, :ECO_LATENT_CHUNKS, :, :].clone() dejavu_latent_for_next_loop = last_trim[:, :, -1:, :, :].clone() latents_video = latents_brutos[:, :, :-(latents_a_podar-1), :, :].clone() latents_video = latents_video[:, :, 1:, :, :] if transition_type == "cut": eco_latent_for_next_loop, dejavu_latent_for_next_loop = None, None raw_latent_fragments.append(latents_video) # --- ATO II: PÓS-PRODUÇÃO LATENTE GLOBAL (CONDICIONAL) --- current_step += 1 progress(current_step / TOTAL_STEPS, desc="Unificação Causal (Concatenação)...") tensors_on_main_device = [frag.to(self.device) for frag in raw_latent_fragments] processed_latents = torch.cat(tensors_on_main_device, dim=2) del raw_latent_fragments, tensors_on_main_device; gc.collect(); torch.cuda.empty_cache() if use_refiner: current_step += 1 progress(current_step / TOTAL_STEPS, desc="Polimento Global (Denoise)...") processed_latents = self.refine_latents( processed_latents, motion_prompt="", guidance_scale=1.0 ) logger.info(f"Polimento global aplicado. Shape: {processed_latents.shape}") else: logger.info("Etapa de refinamento desativada.") # --- ATO III: RENDERIZAÇÃO E FINALIZAÇÃO --- base_name = f"movie_{int(time.time())}" current_step += 1 progress(current_step / TOTAL_STEPS, desc="Renderização (em lotes)...") intermediate_video_path = os.path.join(self.workspace_dir, f"{base_name}_intermediate.mp4") with imageio.get_writer(intermediate_video_path, fps=FPS, codec='libx264', quality=8, output_params=['-pix_fmt', 'yuv420p']) as writer: chunk_size = 15 if use_upscaler else 30 latent_chunks = torch.split(processed_latents, chunk_size, dim=2) for i, latent_chunk in enumerate(latent_chunks): logger.info(f"Processando e renderizando lote {i+1}/{len(latent_chunks)}...") processed_chunk = self.upscale_latents(latent_chunk) if use_upscaler else latent_chunk pixel_tensor_chunk = self.latents_to_pixels(processed_chunk) pixel_tensor_chunk = pixel_tensor_chunk.squeeze(0).permute(1, 2, 3, 0) pixel_tensor_chunk = (pixel_tensor_chunk.clamp(-1, 1) + 1) / 2.0 video_np_chunk = (pixel_tensor_chunk.detach().cpu().float().numpy() * 255).astype(np.uint8) for frame in video_np_chunk: writer.append_data(frame) del latent_chunk, processed_chunk, pixel_tensor_chunk, video_np_chunk gc.collect() torch.cuda.empty_cache() del processed_latents; gc.collect(); torch.cuda.empty_cache() logger.info(f"Vídeo intermediário renderizado em: {intermediate_video_path}") final_video_path = os.path.join(self.workspace_dir, f"{base_name}_FINAL.mp4") if use_hd: current_step += 1 progress(current_step / TOTAL_STEPS, desc="Masterização Final (HD)...") try: hd_specialist_singleton.process_video( input_video_path=intermediate_video_path, output_video_path=final_video_path, prompt=global_prompt ) except Exception as e: logger.error(f"Falha na masterização HD: {e}. Usando vídeo de qualidade padrão.") os.rename(intermediate_video_path, final_video_path) else: logger.info("Etapa de masterização HD desativada.") os.rename(intermediate_video_path, final_video_path) if use_audio: logger.warning("Geração de áudio solicitada, mas está desativada nesta versão do código.") logger.info(f"Processo concluído! Vídeo final salvo em: {final_video_path}") yield {"final_path": final_video_path} def refine_latents(self, latents: torch.Tensor, fps: int = 24, denoise_strength: float = 0.35, refine_steps: int = 12, motion_prompt: str = "...", **kwargs) -> torch.Tensor: logger.info(f"Refinando tensor latente com shape {latents.shape}.") _, _, num_latent_frames, latent_h, latent_w = latents.shape video_scale_factor = getattr(self.vae.config, 'temporal_scale_factor', 8) vae_scale_factor = getattr(self.vae.config, 'spatial_downscale_factor', 8) pixel_height = latent_h * vae_scale_factor pixel_width = latent_w * vae_scale_factor pixel_frames = (num_latent_frames - 1) * video_scale_factor final_ltx_params = { "height": pixel_height, "width": pixel_width, "video_total_frames": pixel_frames, "video_fps": fps, "motion_prompt": motion_prompt, "current_fragment_index": int(time.time()), "denoise_strength": denoise_strength, "refine_steps": refine_steps, "guidance_scale": kwargs.get('guidance_scale', 2.0) } refined_latents_tensor, _ = self.ltx_manager.refine_latents(latents, **final_ltx_params) logger.info(f"Retornando tensor latente refinado com shape: {refined_latents_tensor.shape}") return refined_latents_tensor def upscale_latents(self, latents: torch.Tensor) -> torch.Tensor: logger.info(f"Realizando upscale em tensor latente com shape {latents.shape}.") return upscaler_specialist_singleton.upscale(latents) def _generate_latent_tensor_internal(self, conditioning_items, ltx_params, target_resolution, total_frames_to_generate): kwargs = { **ltx_params, 'width': target_resolution[1], 'height': target_resolution[0], 'video_total_frames': total_frames_to_generate, 'video_fps': 24, 'current_fragment_index': int(time.time()), 'conditioning_items_data': conditioning_items } return self.ltx_manager.generate_latent_fragment(**kwargs) def _quantize_to_multiple(self, n, m): if m == 0: return n quantized = int(round(n / m) * m) return m if n > 0 and quantized == 0 else quantized