deformes4D_engine.py

# 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 random
import gc

from ltx_manager_helpers import ltx_manager_singleton
from gemini_helpers import gemini_singleton 
from ltx_video.models.autoencoders.vae_encode import vae_encode, vae_decode

logger = logging.getLogger(__name__)

@dataclass
class LatentConditioningItem:
    latent_tensor: torch.Tensor
    media_frame_number: int
    conditioning_strength: float

class Deformes4DEngine:
    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 (SDR Executor) 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

    def save_latent_tensor(self, tensor: torch.Tensor, path: str):
        torch.save(tensor.cpu(), path)
        logger.info(f"Tensor latente salvo em: {path}")

    def load_latent_tensor(self, path: str) -> torch.Tensor:
        tensor = torch.load(path, map_location=self.device)
        logger.info(f"Tensor latente carregado de: {path} para o dispositivo {self.device}")
        return tensor

    @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 save_video_from_tensor(self, video_tensor: torch.Tensor, path: str, fps: int = 24):
        if video_tensor is None or video_tensor.ndim != 5 or video_tensor.shape[2] == 0:
            logger.warning("Tentativa de salvar um tensor de vídeo inválido. Abortando.")
            return
        video_tensor = video_tensor.squeeze(0).permute(1, 2, 3, 0)
        video_tensor = (video_tensor.clamp(-1, 1) + 1) / 2.0
        video_np = (video_tensor.detach().cpu().float().numpy() * 255).astype(np.uint8)
        with imageio.get_writer(path, fps=fps, codec='libx264', quality=8) as writer:
            for frame in video_np: writer.append_data(frame)
        logger.info(f"Vídeo salvo em: {path}")

    def _preprocess_image_for_latent_conversion(self, image: Image.Image, target_resolution: tuple) -> Image.Image:
        if image.size != target_resolution:
            logger.info(f"  - AÇÃO: Redimensionando imagem de {image.size} para {target_resolution} antes da conversão para latente.")
            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)
        
    def _generate_video_from_latents(self, latent_tensor, base_name):
        silent_video_path = os.path.join(self.workspace_dir, f"{base_name}_silent.mp4")
        pixel_tensor = self.latents_to_pixels(latent_tensor)
        self.save_video_from_tensor(pixel_tensor, silent_video_path, fps=24)
        del pixel_tensor; gc.collect()
        return silent_video_path

    def _generate_latent_tensor_internal(self, conditioning_items, ltx_params, target_resolution, total_frames_to_generate):
        final_ltx_params = {**ltx_params, 'width': target_resolution[0], 'height': target_resolution[1], 'video_total_frames': total_frames_to_generate, 'video_fps': 24, 'current_fragment_index': int(time.time()), 'conditioning_items_data': conditioning_items}
        new_full_latents, _ = self.ltx_manager.generate_latent_fragment(**final_ltx_params)
        return new_full_latents

    def concatenate_videos_ffmpeg(self, video_paths: list[str], output_path: str) -> str:
        if not video_paths: raise gr.Error("Nenhum fragmento de vídeo para montar.")
        list_file_path = os.path.join(self.workspace_dir, "concat_list.txt")
        with open(list_file_path, 'w', encoding='utf-8') as f:
            for path in video_paths: f.write(f"file '{os.path.abspath(path)}'\n")
        cmd_list = ['ffmpeg', '-y', '-f', 'concat', '-safe', '0', '-i', list_file_path, '-c', 'copy', output_path]
        logger.info("Executando concatenação FFmpeg...")
        try:
            subprocess.run(cmd_list, check=True, capture_output=True, text=True)
        except subprocess.CalledProcessError as e:
            logger.error(f"Erro no FFmpeg: {e.stderr}")
            raise gr.Error(f"Falha na montagem final do vídeo. Detalhes: {e.stderr}")
        return output_path
    
    def generate_full_movie(self, keyframes: list, global_prompt: str, storyboard: list, 
                            # Novos parâmetros de controle vindos da UI
                            n_chunks_to_generate: int,
                            video_end_chunk: int,
                            eco_start_chunk: int,
                            handler_start_chunk: int,
                            handler_frame_target: int,
                            # Parâmetros restantes
                            handler_strength: float, destination_convergence_strength: float, 
                            video_resolution: int, use_continuity_director: bool, 
                            progress: gr.Progress = gr.Progress()):
        
        # --- Lógica de Geração Dinâmica Baseada nos Parâmetros da UI ---
        FRAMES_TO_GENERATE = (n_chunks_to_generate - 1) * 8 + 1
        
        VIDEO_CHUNK_INDICES = slice(0, video_end_chunk)
        ECO_CHUNK_INDICES = slice(eco_start_chunk, eco_start_chunk + 2)
        HANDLER_CHUNK_INDICES = slice(handler_start_chunk, handler_start_chunk + 2)
        
        HANDLER_FRAME_TARGET = handler_frame_target
        DESTINATION_FRAME_TARGET = FRAMES_TO_GENERATE - 1

        # Validação de sanidade para garantir que as fatias não se sobreponham de forma incorreta
        if not (video_end_chunk <= eco_start_chunk and 
                (eco_start_chunk + 2) <= handler_start_chunk and 
                (handler_start_chunk + 2) <= n_chunks_to_generate):
            raise gr.Error(f"Configuração de Chunks inválida! Verifique as sobreposições e o total de chunks.")

        logger.info("="*60)
        logger.info("MODO DE GERAÇÃO: Estratégia de Cauda Longa Dinâmica")
        logger.info(f"  - Geração Bruta por Fragmento: {n_chunks_to_generate} chunks ({FRAMES_TO_GENERATE} frames)")
        logger.info(f"  - Clipe Final por Fragmento: Chunks {VIDEO_CHUNK_INDICES.start} a {VIDEO_CHUNK_INDICES.stop-1}")
        logger.info(f"  - Guia de Eco (Memória): Chunks {ECO_CHUNK_INDICES.start} a {ECO_CHUNK_INDICES.stop-1}")
        logger.info(f"  - Guia de Handler (Evolução): Chunks {HANDLER_CHUNK_INDICES.start} a {HANDLER_CHUNK_INDICES.stop-1}")
        logger.info(f"  - Ponto de Aplicação do Handler: Frame {HANDLER_FRAME_TARGET}")
        logger.info("="*60)
        
        base_ltx_params = {"guidance_scale": 1.0, "stg_scale": 0.0, "rescaling_scale": 0.15, "num_inference_steps": 20}
        keyframe_paths = [item[0] if isinstance(item, tuple) else item for item in keyframes]
        video_clips_paths, story_history = [], ""
        target_resolution_tuple = (video_resolution, video_resolution) 
        
        eco_latent_for_next_loop = None
        handler_latent_for_next_loop = None
        
        if len(keyframe_paths) < 3:
            raise gr.Error(f"O modelo de geração requer no mínimo 3 keyframes (Passado, Presente, Futuro). Você forneceu {len(keyframe_paths)}.")
        
        num_transitions_to_generate = len(keyframe_paths) - 2
        
        for i in range(num_transitions_to_generate):
            start_keyframe_index = i + 1
            
            logger.info(f"--- INICIANDO FRAGMENTO {i+1}/{num_transitions_to_generate} ---")
            progress((i + 1) / num_transitions_to_generate, desc=f"Produzindo Transição {i+1}/{num_transitions_to_generate}")
            
            past_keyframe_path = keyframe_paths[start_keyframe_index - 1]
            start_keyframe_path = keyframe_paths[start_keyframe_index]
            destination_keyframe_path = keyframe_paths[start_keyframe_index + 1]
            future_story_prompt = storyboard[start_keyframe_index + 1] if (start_keyframe_index + 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[start_keyframe_index - 1], storyboard[start_keyframe_index], future_story_prompt
            )
            _, motion_prompt = decision["transition_type"], decision["motion_prompt"]
            story_history += f"\n- Ato {i+1}: {motion_prompt}"

            conditioning_items = []
            logger.info("  [0. PREPARAÇÃO] Montando itens de condicionamento...")
            
            if i == 0:
               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))
               logger.info("    - Condicionamento inicial: Imagem K1 no frame 0.")
            else:
               conditioning_items.append(LatentConditioningItem(eco_latent_for_next_loop, 0, 1.0))
               conditioning_items.append(LatentConditioningItem(handler_latent_for_next_loop, HANDLER_FRAME_TARGET, handler_strength))
               logger.info(f"    - Guia de Memória (Eco) aplicada no frame 0 (shape: {eco_latent_for_next_loop.shape}).")
               logger.info(f"    - Guia de Evolução (Handler) aplicada no frame {HANDLER_FRAME_TARGET} (shape: {handler_latent_for_next_loop.shape}).")
            
            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))
            logger.info(f"    - Guia de Destino: Imagem K{start_keyframe_index + 1} no frame {DESTINATION_FRAME_TARGET}.")

            current_ltx_params = {**base_ltx_params, "motion_prompt": motion_prompt}
            new_full_latents = self._generate_latent_tensor_internal(conditioning_items, current_ltx_params, target_resolution_tuple, FRAMES_TO_GENERATE)
            logger.info(f"  [1. GERAÇÃO] Tensor latente bruto gerado com shape: {new_full_latents.shape}.")

            eco_latent_for_next_loop = new_full_latents[:, :, ECO_CHUNK_INDICES, :, :].clone()
            handler_latent_for_next_loop = new_full_latents[:, :, HANDLER_CHUNK_INDICES, :, :].clone()
            logger.info(f"  [GUIAS] Guias para a próxima iteração extraídas. Eco shape: {eco_latent_for_next_loop.shape}, Handler shape: {handler_latent_for_next_loop.shape}.")

            latents_for_video = new_full_latents[:, :, VIDEO_CHUNK_INDICES, :, :]
            logger.info(f"  [2. EDIÇÃO] Tensor final para vídeo extraído com {latents_for_video.shape[2]} chunks.")

            base_name = f"fragment_{i}_{int(time.time())}"
            video_path = self._generate_video_from_latents(latents_for_video, base_name)
            video_clips_paths.append(video_path)
            yield {"fragment_path": video_path}
                 
        final_movie_path = os.path.join(self.workspace_dir, f"final_movie_silent_{int(time.time())}.mp4")
        self.concatenate_videos_ffmpeg(video_clips_paths, final_movie_path)
        
        logger.info(f"Filme completo salvo em: {final_movie_path}")
        yield {"final_path": final_movie_path}

    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