Update deformes4D_engine.py
Browse files- deformes4D_engine.py +8 -62
deformes4D_engine.py
CHANGED
|
@@ -126,55 +126,7 @@ class Deformes4DEngine:
|
|
| 126 |
raise gr.Error(f"Falha na montagem final do vídeo. Detalhes: {e.stderr}")
|
| 127 |
return output_path
|
| 128 |
|
| 129 |
-
|
| 130 |
-
def _render_and_post_process_latents(self,
|
| 131 |
-
low_res_latents: torch.Tensor,
|
| 132 |
-
base_name: str,
|
| 133 |
-
motion_prompt_for_refine: str,
|
| 134 |
-
final_resolution: tuple
|
| 135 |
-
) -> str:
|
| 136 |
-
logger.info(f"--- INICIANDO PIPELINE DE PÓS-PRODUÇÃO PARA '{base_name}' ---")
|
| 137 |
-
|
| 138 |
-
high_res_latents = upscaler_specialist_singleton.upscale(low_res_latents)
|
| 139 |
-
|
| 140 |
-
_, _, _, refined_h_latent, refined_w_latent = high_res_latents.shape
|
| 141 |
-
video_h = refined_h_latent * self.ltx_manager.workers[0].pipeline.vae_scale_factor
|
| 142 |
-
video_w = refined_w_latent * self.ltx_manager.workers[0].pipeline.vae_scale_factor
|
| 143 |
-
num_latent_frames = high_res_latents.shape[2]
|
| 144 |
-
num_video_frames = num_latent_frames * self.ltx_manager.workers[0].pipeline.video_scale_factor
|
| 145 |
-
if isinstance(self.vae, CausalVideoAutoencoder):
|
| 146 |
-
num_video_frames -= 1
|
| 147 |
-
|
| 148 |
-
refine_kwargs = {
|
| 149 |
-
'height': video_h, 'width': video_w, 'video_total_frames': num_video_frames, 'video_fps': 24,
|
| 150 |
-
'current_fragment_index': int(time.time()), 'motion_prompt': motion_prompt_for_refine,
|
| 151 |
-
'conditioning_items_data': [],
|
| 152 |
-
'denoise_strength': 0.4, 'refine_steps': 10
|
| 153 |
-
}
|
| 154 |
-
|
| 155 |
-
final_latents, _ = self.ltx_manager.refine_latents(high_res_latents, **refine_kwargs)
|
| 156 |
-
|
| 157 |
-
untrimmed_video_path = os.path.join(self.workspace_dir, f"{base_name}_untrimmed.mp4")
|
| 158 |
-
trimmed_video_path = os.path.join(self.workspace_dir, f"{base_name}.mp4")
|
| 159 |
-
|
| 160 |
-
pixel_tensor = self.latents_to_pixels(final_latents)
|
| 161 |
-
|
| 162 |
-
logger.info(f"Redimensionando vídeo final de {pixel_tensor.shape[-2:]} para {final_resolution}")
|
| 163 |
-
pixel_tensor_resized = torch.nn.functional.interpolate(
|
| 164 |
-
pixel_tensor.squeeze(0), size=final_resolution, mode='bilinear', align_corners=False
|
| 165 |
-
).unsqueeze(0)
|
| 166 |
-
|
| 167 |
-
self.save_video_from_tensor(pixel_tensor_resized, untrimmed_video_path, fps=24)
|
| 168 |
-
del pixel_tensor, final_latents, high_res_latents, pixel_tensor_resized
|
| 169 |
-
gc.collect()
|
| 170 |
-
torch.cuda.empty_cache()
|
| 171 |
-
|
| 172 |
-
success = self._trim_last_frame_ffmpeg(untrimmed_video_path, trimmed_video_path)
|
| 173 |
-
|
| 174 |
-
if os.path.exists(untrimmed_video_path):
|
| 175 |
-
os.remove(untrimmed_video_path)
|
| 176 |
-
|
| 177 |
-
return trimmed_video_path if success else untrimmed_video_path
|
| 178 |
|
| 179 |
def _generate_video_and_audio(self, silent_video_path: str, audio_prompt: str, base_name: str) -> str:
|
| 180 |
try:
|
|
@@ -283,21 +235,15 @@ class Deformes4DEngine:
|
|
| 283 |
|
| 284 |
progress((num_transitions_to_generate + 1) / (num_transitions_to_generate + 2), desc="Pós-produção (Upscale e Refinamento)...")
|
| 285 |
base_name = f"final_movie_hq_{int(time.time())}"
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
|
| 290 |
-
|
|
|
|
| 291 |
)
|
| 292 |
-
|
| 293 |
-
#progress((num_transitions_to_generate + 1.5) / (num_transitions_to_generate + 2), desc="Gerando paisagem sonora...")
|
| 294 |
-
#video_with_audio_path = self._generate_video_and_audio(
|
| 295 |
-
# silent_video_path=silent_video_path,
|
| 296 |
-
# audio_prompt=global_prompt,
|
| 297 |
-
# base_name=base_name
|
| 298 |
-
#)
|
| 299 |
|
| 300 |
-
yield {"final_path":
|
| 301 |
|
| 302 |
def _generate_latent_tensor_internal(self, conditioning_items, ltx_params, target_resolution, total_frames_to_generate):
|
| 303 |
kwargs = {
|
|
|
|
| 126 |
raise gr.Error(f"Falha na montagem final do vídeo. Detalhes: {e.stderr}")
|
| 127 |
return output_path
|
| 128 |
|
| 129 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
|
| 131 |
def _generate_video_and_audio(self, silent_video_path: str, audio_prompt: str, base_name: str) -> str:
|
| 132 |
try:
|
|
|
|
| 235 |
|
| 236 |
progress((num_transitions_to_generate + 1) / (num_transitions_to_generate + 2), desc="Pós-produção (Upscale e Refinamento)...")
|
| 237 |
base_name = f"final_movie_hq_{int(time.time())}"
|
| 238 |
+
|
| 239 |
+
progress((num_transitions_to_generate + 1.5) / (num_transitions_to_generate + 2), desc="Gerando paisagem sonora...")
|
| 240 |
+
video_with_audio_path = self._generate_video_and_audio(
|
| 241 |
+
silent_video_path=silent_video_path,
|
| 242 |
+
audio_prompt=global_prompt,
|
| 243 |
+
base_name=base_name
|
| 244 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 245 |
|
| 246 |
+
yield {"final_path": video_with_audio_path}
|
| 247 |
|
| 248 |
def _generate_latent_tensor_internal(self, conditioning_items, ltx_params, target_resolution, total_frames_to_generate):
|
| 249 |
kwargs = {
|