Update api/ltx_server_refactored_complete.py

api/ltx_server_refactored_complete.py

@@ -283,31 +283,57 @@ class VideoService:
         downscaled_height = self._align(int(height_padded * downscale_factor), vae_scale_factor)
         downscaled_width = self._align(int(width_padded * downscale_factor), vae_scale_factor)
 
+        # 1. Começa com a configuração padrão
         first_pass_config = self.config.get("first_pass", {}).copy()
+        
+        # 2. Aplica os overrides da UI, se existirem
         if kwargs.get("ltx_configs_override"):
-            self._apply_ui_overrides(first_pass_config)
+            self._apply_ui_overrides(first_pass_config, kwargs.get("ltx_configs_override"))
 
+        # 3. Monta o dicionário de argumentos SEM conditioning_items primeiro
         pipeline_kwargs = {
-            "prompt": kwargs['prompt'], "negative_prompt": kwargs['negative_prompt'],
-            "height": downscaled_height, "width": downscaled_width, "num_frames": kwargs['num_frames'],
-            "frame_rate": DEFAULT_FPS, "generator": torch.Generator(device=self.main_device).manual_seed(kwargs['seed']),
-            "output_type": "latent", **first_pass_config
+            "prompt": kwargs['prompt'],
+            "negative_prompt": kwargs['negative_prompt'],
+            "height": downscaled_height,
+            "width": downscaled_width,
+            "num_frames": kwargs['num_frames'],
+            "frame_rate": int(DEFAULT_FPS),
+            "generator": torch.Generator(device=self.main_device).manual_seed(kwargs['seed']),
+            "output_type": "latent",
+            #"conditioning_items": conditioning_items if conditioning_items else None,
+            "media_items": None,
+            "decode_timestep": self.config["decode_timestep"],
+            "decode_noise_scale": self.config["decode_noise_scale"],
+            "stochastic_sampling": self.config["stochastic_sampling"],
+            "image_cond_noise_scale": 0.01,
+            "is_video": True,
+            "vae_per_channel_normalize": True,
+            "mixed_precision": (self.config["precision"] == "mixed_precision"),
+            "offload_to_cpu": False,
+            "enhance_prompt": False,
+            #"skip_layer_strategy": SkipLayerStrategy.AttentionValues,
+            **first_pass_config
         }
-
-        logging.info(f"\n[Info] pipeline_kwargs:\n {pipeline_kwargs}\n\n")
-        self._log_conditioning_items(kwargs.get['conditioning_items'])
         
-        pipeline_kwargs = {
-            "conditioning_items": kwargs['conditioning_items'],
-            **pipeline_kwargs
-        }
+        # --- Bloco de Logging para Depuração ---
+        # 4. Loga os argumentos do pipeline (sem os tensores de condição)
+        logging.info(f"\n[Info] Pipeline Arguments (BASE):\n {json.dumps(pipeline_kwargs, indent=2, default=str)}\n")
+        
+        # Loga os conditioning_items separadamente com a nossa função helper
+        conditioning_items_list = kwargs.get('conditioning_items')
+        self._log_conditioning_items(conditioning_items_list)
+        # --- Fim do Bloco de Logging ---
 
+        # 5. Adiciona os conditioning_items ao dicionário
+        pipeline_kwargs['conditioning_items'] = conditioning_items_list
         
+        # 6. Executa o pipeline com o dicionário completo
         with torch.autocast(device_type=self.main_device.type, dtype=self.runtime_autocast_dtype, enabled="cuda" in self.main_device.type):
             latents_raw = self.pipeline(**pipeline_kwargs).images
         
         return latents_raw.to(self.main_device)
 
+    
     @log_function_io
     def _finalize_generation(self, temp_latent_paths: List[Path], base_filename: str, seed: int) -> Tuple[str, str, int]:
         """Consolidates latents, decodes them to video, and saves final artifacts."""