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 audio_specialist import audio_specialist_singleton
from ltx_manager_helpers import ltx_manager_singleton
from gemini_helpers import gemini_singleton 
from upscaler_specialist import upscaler_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: Câmera Ψ e Destilador Δ) 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
    def save_latent_tensor(self, tensor: torch.Tensor, path: str):
        torch.save(tensor.cpu(), path)

    def load_latent_tensor(self, path: str) -> torch.Tensor:
        return torch.load(path, map_location=self.device)

    @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: 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)

    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)
    
    def _get_video_frame_count(self, video_path: str) -> int | None:
        if not os.path.exists(video_path): return None
        cmd = ['ffprobe', '-v', 'error', '-select_streams', 'v:0', '-count_frames',
               '-show_entries', 'stream=nb_read_frames', '-of', 'default=nokey=1:noprint_wrappers=1', video_path]
        try:
            result = subprocess.run(cmd, check=True, capture_output=True, text=True, encoding='utf-8')
            return int(result.stdout.strip())
        except Exception: return None

    def _trim_last_frame_ffmpeg(self, input_path: str, output_path: str) -> bool:
        frame_count = self._get_video_frame_count(input_path)
        if frame_count is None or frame_count < 2:
            if os.path.exists(input_path): os.rename(input_path, output_path)
            return True
        vf_filter = f"select='lt(n,{frame_count - 1})',setpts=PTS-STARTPTS"
        cmd_list = ['ffmpeg', '-y', '-i', input_path, '-vf', vf_filter, '-an', output_path]
        try:
            subprocess.run(cmd_list, check=True, capture_output=True, text=True, encoding='utf-8')
            return True
        except subprocess.CalledProcessError: return False
            
    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]
        try:
            subprocess.run(cmd_list, check=True, capture_output=True, text=True)
        except subprocess.CalledProcessError as e:
            raise gr.Error(f"Falha na montagem final do vídeo. Detalhes: {e.stderr}")
        return output_path

 

    def _generate_video_and_audio(self, silent_video_path: str, audio_prompt: str, base_name: str) -> str:
        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)
            duration = float(result.stdout.strip())
        except Exception:
            frame_count = self._get_video_frame_count(silent_video_path)
            duration = (frame_count / 24.0) if frame_count else 0

        video_with_audio_path = audio_specialist_singleton.generate_audio_for_video(
            video_path=silent_video_path, prompt=audio_prompt,
            duration_seconds=duration)
        
        return video_with_audio_path

    # 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, 
                            video_resolution: int, use_continuity_director: bool, 
                            progress: gr.Progress = gr.Progress()):
        
        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 + 1:
            raise gr.Error(f"A combinação de duração e poda é muito agressiva.")

        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}
        keyframe_paths = [item[0] if isinstance(item, tuple) else item for item in keyframes]
        story_history = ""
        
        
        eco_latent_for_next_loop = None
        dejavu_latent_for_next_loop = None
        
        num_transitions_to_generate = len(keyframe_paths) - 1
        low_res_latent_fragments = []

        for i in range(num_transitions_to_generate):
            fragment_index = i + 1
            progress(i / (num_transitions_to_generate + 2), desc=f"Gerando Latentes do Fragmento {fragment_index}")
            
            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 = 768, 1152
            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)
            final_resolution_tuple = (expected_height, expected_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


            hig_res_latent_fragments = upscaler_specialist_singleton.upscale(latents_video)
        
            low_res_latent_fragments.append(hig_res_latent_fragments)


        
        
        progress((num_transitions_to_generate) / (num_transitions_to_generate + 2), desc="Concatenando latentes...")
        tensors_para_concatenar = []
        target_device = self.device
        for idx, tensor_frag in enumerate(low_res_latent_fragments):
            tensor_on_target_device = tensor_frag.to(target_device)
            if idx < len(low_res_latent_fragments) - 1:
                tensors_para_concatenar.append(tensor_on_target_device[:, :, :-1, :, :]) 
            else:
                tensors_para_concatenar.append(tensor_on_target_device)
        
        final_concatenated_latents = torch.cat(tensors_para_concatenar, dim=2)
        
        progress((num_transitions_to_generate + 1) / (num_transitions_to_generate + 2), desc="Pós-produção (Upscale e Refinamento)...")
        base_name = f"final_movie_hq_{int(time.time())}"


        # Pós-produção: Upscale + Refine
        high_quality_video_path = self._render_and_post_process(
            final_concatenated_latents,
            base_name=base_name,
            expected_height=720,
            expected_width=720,
            fps=24
        )

        
        #progress((num_transitions_to_generate + 1.5) / (num_transitions_to_generate + 2), desc="Gerando paisagem sonora...")
        #video_with_audio_path = self._generate_video_and_audio(
        #    silent_video_path=silent_video_path,
        #    audio_prompt=global_prompt,
        #    base_name=base_name
        #)
        
        yield {"final_path": high_quality_video_path}


    def _render_and_post_process(self, final_concatenated_latents: torch.Tensor, 
                             base_name: str, expected_height: int, expected_width: int, fps: int = 24) -> str:
        logger.info("Iniciando pós-processamento: upscale + refinamento...")

        # --- 1. Upscale ---
        upscaled_latents = upscaler_specialist_singleton.upscale(final_concatenated_latents)
        logger.info(f"Upscale concluído: shape {list(upscaled_latents.shape)}")

        # --- 2. Refinamento ---
        _, _, _, h, w = upscaled_latents.shape
        refined_latents, _ = ltx_manager_singleton.refine_latents(
            upscaled_latents,
            height=h,
            width=w,
            denoise_strength=0.35,   # levemente menor pra preservar nitidez
            refine_steps=12          # mais iterações pra polir detalhes
        )
        logger.info("Refinamento concluído.")

        # --- 3. Decodificação ---
        pixel_tensor = self.latents_to_pixels(refined_latents)

        # --- 4. Render final ---
        video_path = os.path.join(self.workspace_dir, f"{base_name}_HQ.mp4")
        self.save_video_from_tensor(pixel_tensor, video_path, fps=fps)
        logger.info(f"Vídeo final salvo em: {video_path}")

        return video_path

        
    
    def _generate_latent_tensor_internal(self, conditioning_items, ltx_params, target_resolution, total_frames_to_generate):
        kwargs = {
            **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
        }
        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