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 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
from audio_specialist import audio_specialist_singleton

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_old(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()):
        
        # 1. Definição dos Parâmetros da Geração com base na Tese
        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)
        total_latents_brutos = total_frames_brutos // 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

        if total_latents_brutos <= latents_a_podar:
            raise gr.Error(f"A porcentagem de poda ({trim_percent}%) é muito alta. Reduza-a ou aumente a duração.")

        DEJAVU_FRAME_TARGET = frames_a_podar - 1
        DESTINATION_FRAME_TARGET = total_frames_brutos - 1
        
        logger.info("--- CONFIGURAÇÃO DA GERAÇÃO ADUC-SDR ---")
        logger.info(f"Total de Latents por Geração Exploratória (V_bruto): {total_latents_brutos} ({total_frames_brutos} frames)")
        logger.info(f"Latents a serem descartados (Poda Causal): {latents_a_podar} ({frames_a_podar} frames)")
        logger.info(f"Chunks Latentes do Eco Causal (C): {ECO_LATENT_CHUNKS}")
        logger.info(f"Frame alvo do Déjà-Vu (D): {DEJAVU_FRAME_TARGET}")
        logger.info(f"Frame alvo do Destino (K): {DESTINATION_FRAME_TARGET}")
        logger.info("------------------------------------------")

        # 2. Inicialização do Estado
        base_ltx_params = {"guidance_scale": 2.0, "stg_scale": 0.025, "rescaling_scale": 0.15, "num_inference_steps": 20, "image_cond_noise_scale": 0.00}
        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
        dejavu_latent_for_next_loop = None
        
        latent_fragments[]
        
        if len(keyframe_paths) < 2:
            raise gr.Error(f"A geração requer no mínimo 2 keyframes. Você forneceu {len(keyframe_paths)}.")
        
        num_transitions_to_generate = len(keyframe_paths) - 1
        
        # 3. Loop Principal de Geração de Fragmentos
        for i in range(num_transitions_to_generate):
            fragment_index = i + 1
            logger.info(f"--- INICIANDO FRAGMENTO {fragment_index}/{num_transitions_to_generate} ---")
            progress(fragment_index / num_transitions_to_generate, desc=f"Produzindo Transição {fragment_index}/{num_transitions_to_generate}")
            
            # 3.1. Consulta ao Maestro (Γ) para obter a intenção (Pᵢ)
            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}"

            # 3.2. Montagem das Âncoras para a Fórmula Canônica Ψ({C, D, K}, P)
            conditioning_items = []
            logger.info("  [Ψ.1] Montando âncoras causais...")
            
            if eco_latent_for_next_loop is None:
               logger.info("    - Primeiro fragmento: Usando Keyframe inicial como âncora de partida.")
               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:
               logger.info("    - Âncora 1: Eco Causal (C) - Herança do passado.")
               conditioning_items.append(LatentConditioningItem(eco_latent_for_next_loop, 0, 1.0))
               logger.info("    - Âncora 2: Déjà-Vu (D) - Memória de um futuro idealizado.")
               conditioning_items.append(LatentConditioningItem(dejavu_latent_for_next_loop, DEJAVU_FRAME_TARGET, handler_strength))
            
            logger.info("    - Âncora 3: Destino (K) - Âncora geométrica/narrativa.")
            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))

            # 3.3. Execução da Câmera (Ψ): Geração Exploratória para criar V_bruto
            logger.info(f"  [Ψ.2] Câmera (Ψ) executando a geração exploratória de {total_latents_brutos} chunks latentes...")
            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)
            logger.info(f"    - Geração concluída. Tensor latente bruto (V_bruto) criado com shape: {latents_brutos.shape}.")

            # 3.4. Execução do Destilador (Δ): Implementação do Ciclo de Poda Causal (com workaround empírico)
            logger.info(f"  [Δ] Destilador (Δ) executando o Ciclo de Poda Causal...")

            last_trim = latents_brutos[:, :, -(latents_a_podar+1):, :, :].clone()
            eco_latent_for_next_loop = last_trim[:, :, :2, :, :].clone()   
            dejavu_latent_for_next_loop = last_trim[:, :, -1:, :, :].clone()
            
            latents_video = latents_brutos[:, :, :-(latents_a_podar-1), :, :].clone()
            latents_video = latents_video[:, :, 1:, :, :]

            logger.info(f"  [Δ] latents_video {latents_video.shape}")
            logger.info(f"    - (Δ.1) Déjà-Vu (D) destilado. Shape: {dejavu_latent_for_next_loop.shape}")
            logger.info(f"    - (Δ.2) Eco Causal (C) extraído. Shape: {eco_latent_for_next_loop.shape}")

            if transition_type == "cut":
                logger.warning("  - DECISÃO DO MAESTRO: Corte ('cut'). Resetando a memória causal (Eco e Déjà-Vu).")
                eco_latent_for_next_loop = None
                dejavu_latent_for_next_loop = None

            if use_upscaler:
                latents_video = self.upscale_latents(latents_video)
            
            latent_fragments.append(latents_video)
        
        current_step += 1
        progress(current_step / TOTAL_STEPS, desc="Unificação Causal (Concatenação)...")
        
        logger.info("--- CONCATENANDO TODOS OS FRAGMENTOS LATENTES ---")
        tensors_para_concatenar = []
        for idx, tensor_frag in enumerate(latent_fragments):
            # Move cada tensor para o dispositivo de destino antes de adicioná-lo à lista.
            target_device = self.device
            tensor_on_target_device = tensor_frag.to(target_device)
            if idx < len(latent_fragments) - 1:
                tensors_para_concatenar.append(tensor_on_target_device[:, :, :-1, :, :]) 
            else:
                tensors_para_concatenar.append(tensor_on_target_device)

        processed_latents = torch.cat(tensors_para_concatenar, dim=2)
        logger.info(f"Concatenação concluída. Shape final do tensor latente: {final_concatenated_latents.shape}")
        
        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.")

        
        base_name = f"movie_{int(time.time())}"
        current_step += 1
        progress(current_step / TOTAL_STEPS, desc="Renderização (em lotes)...")
        
        if use_audio:
             video_path = self._generate_video_and_audio_from_latents(processed_latents, global_prompt, base_name)
        else:
            video_path = os.path.join(self.workspace_dir, f"{base_name}_silent.mp4")
            logger.info("Etapa de sonoplastia desativada.")
            pixel_tensor = self.latents_to_pixels(processed_latents)
            self.save_video_from_tensor(pixel_tensor, video_path, fps=24)
        
        if use_hd:
            current_step += 1
            progress(current_step / TOTAL_STEPS, desc="Masterização Final (HD)...")
            try:
                hd_specialist_singleton.process_video(
                    input_video_path=video_path,
                    output_video_path=video_path,
                    prompt=" "
                )
            except Exception as e:
                logger.error(f"Falha na masterização HD: {e}. Usando vídeo de qualidade padrão.")
        else:
            logger.info("Etapa de edicao HD desativada.")
            
        
        logger.info(f"Processo concluído! Vídeo final salvo em: {video_path}")
        yield {"final_path": video_path}


    def _generate_video_and_audio_from_latents(self, latent_tensor, audio_prompt, 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()
        
        try:
            result = subprocess.run(
                ["ffprobe", "-v", "error", "-show_entries", "format=duration", "-of", "default=noprint_wrappers=1:nokey=1", silent_video_path],
                capture_output=True, text=True, check=True)
            frag_duration = float(result.stdout.strip())
        except (subprocess.CalledProcessError, ValueError, FileNotFoundError):
             logger.warning(f"ffprobe falhou em {os.path.basename(silent_video_path)}. Calculando duração manualmente.")
             num_pixel_frames = latent_tensor.shape[2] * 8
             frag_duration = num_pixel_frames / 24.0

        video_with_audio_path = audio_specialist_singleton.generate_audio_for_video(
            video_path=silent_video_path, prompt=audio_prompt,
            duration_seconds=frag_duration)
        
        if os.path.exists(silent_video_path):
             os.remove(silent_video_path)
        return video_with_audio_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):
        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)
        gc.collect()
        torch.cuda.empty_cache()
        return new_full_latents

        
    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