Update api/ltx_server_refactored_complete.py

api/ltx_server_refactored_complete.py

@@ -210,7 +210,8 @@ class VideoService:
                 current_frames = self._align(current_frames, alignment_rule='n*8+1')
 
                 current_conditions = kwargs.get("initial_conditions", []) if i == 0 else []
-                if overlap_condition_item: current_conditions.append(overlap_condition_item)
+                if overlap_condition_item: 
+                    current_conditions.append(overlap_condition_item)
 
                 chunk_latents = self._generate_single_chunk_low(
                     prompt=chunk_prompt, num_frames=current_frames, seed=used_seed + i,
@@ -219,9 +220,28 @@ class VideoService:
                 if chunk_latents is None: raise RuntimeError(f"Failed to generate latents for scene {i+1}.")
 
                 if is_narrative and i < num_chunks - 1:
+                    # 1. Criar tensor overlap latente
                     overlap_latents = chunk_latents[:, :, -overlap_frames:, :, :].clone()
-                    overlap_condition_item = ConditioningItem(media_item=overlap_latents, media_frame_number=0, conditioning_strength=1.0)
-                
+                    logging.info(f"Criado overlap latente com shape: {list(overlap_latents.shape)}")
+
+                    # 2. DECODIFICA o latente de volta para um tensor de PIXEL
+                    logging.info("Decodificando latente de overlap para tensor de pixel...")
+                    overlap_pixel_tensor = vae_manager_singleton.decode(
+                        overlap_latents,
+                        decode_timestep=float(self.config.get("decode_timestep", 0.05))
+                    )
+                    # O resultado de decode() está na CPU, no formato (B, C, F, H, W) e [0, 1]
+                    # Precisamos normalizá-lo para [-1, 1] que é o que o pipeline espera.
+                    overlap_pixel_tensor_normalized = (overlap_pixel_tensor * 2.0) - 1.0
+                    logging.info(f"Tensor de pixel de overlap criado com shape: {list(overlap_pixel_tensor_normalized.shape)}")
+                    
+                    # 3. Cria o ConditioningItem com o TENSOR DE PIXEL, não com o latente.
+                    overlap_condition_item = ConditioningItem(
+                        media_item=overlap_pixel_tensor_normalized, 
+                        media_frame_number=0,conditioning_strength=1.0
+                    )
+                    
+                                
                 if i > 0:
                     chunk_latents = chunk_latents[:, :, overlap_frames:, :, :]