Update api/ltx_server.py

api/ltx_server.py

@@ -543,10 +543,10 @@ class VideoService:
                 start = (num_latente_por_chunk*i)
                 end = (start+num_latente_por_chunk+overlap)
                 if i+1 < n_chunks:
-                    chunk = latents_brutos[:, :, start:end, :, :].clone().detach()
+                    chunk = latents_brutos[:, :, start:end, :, :].detach()
                     print(f"[DEBUG] chunk{i+1}[:, :, {start}:{end}, :, :] = {chunk.shape[2]}")
                 else:
-                    chunk = latents_brutos[:, :, start:, :, :].clone().detach()
+                    chunk = latents_brutos[:, :, start:, :, :].detach()
                     print(f"[DEBUG] chunk{i+1}[:, :, {start}:, :, :] = {chunk.shape[2]}")
                 chunks.append(chunk)
                 i+=1
@@ -578,8 +578,8 @@ class VideoService:
         #if total % 2 == 1:  # ÍMPAR
             # Ex: 11 → primeira 0..5, segunda 5..10
         cut = total // 2
-        primeira = latents_brutos[:, :, :cut+1, :, :].clone()
-        segunda  = latents_brutos[:, :, cut:, :, :].clone()
+        primeira = latents_brutos[:, :, :cut+1, :, :].detach()
+        segunda  = latents_brutos[:, :, cut:, :, :].detach()
         
         
         return primeira, segunda
@@ -759,14 +759,20 @@ class VideoService:
         if mode == "image-to-video":
             start_tensor = self._prepare_conditioning_tensor(start_image_filepath, height, width, padding_values)
             conditioning_items.append(ConditioningItem(start_tensor, 0, 1.0))
+            conditioning_items1.append(ConditioningItem(start_tensor, 0, 1.0))
+         
             if middle_image_filepath and middle_frame_number is not None:
                 middle_tensor = self._prepare_conditioning_tensor(middle_image_filepath, height, width, padding_values)
                 safe_middle_frame = max(0, min(int(middle_frame_number), actual_num_frames - 1))
                 conditioning_items.append(ConditioningItem(middle_tensor, safe_middle_frame, float(middle_image_weight)))
+                conditioning_items1.append(ConditioningItem(middle_tensor, safe_middle_frame, float(middle_image_weight))
+         
             if end_image_filepath:
                 end_tensor = self._prepare_conditioning_tensor(end_image_filepath, height, width, padding_values)
                 last_frame_index = actual_num_frames - 1
                 conditioning_items.append(ConditioningItem(end_tensor, last_frame_index, float(end_image_weight)))
+                conditioning_items2.append(ConditioningItem(end_tensor, last_frame_index//2, 1.0))
+         
             print(f"[DEBUG] Conditioning items: {len(conditioning_items)}")
 
         call_kwargs = {
@@ -889,7 +895,21 @@ class VideoService:
                         print(f"[DEBUG] Parte: {num_latent_frames_part - 1} latentes -> {num_pixel_frames_part} frames de pixel (alvo)")
                        
                         second_pass_kwargs = call_kwargs.copy()
-                        second_pass_kwargs.update({
+
+                        if i==0:
+                          second_pass_kwargs.update({
+                           "conditioning_items": conditioning_items1
+                           "output_type": "latent",
+                           "width": second_pass_width,
+                           "height": second_pass_height,
+                           "num_frames": num_pixel_frames_part,
+                           "latents": latents, # O tensor upscaled
+                           "guidance_scale": float(guidance_scale),
+                           **second_pass_config
+                          })
+                        else:
+                          second_pass_kwargs.update({
+                           "conditioning_items": conditioning_items2
                            "output_type": "latent",
                            "width": second_pass_width,
                            "height": second_pass_height,
@@ -897,7 +917,7 @@ class VideoService:
                            "latents": latents, # O tensor upscaled
                            "guidance_scale": float(guidance_scale),
                            **second_pass_config
-                        })
+                          })
                     
                         print(f"[DEBUG] Second Pass: Refinando em {width_padded}x{height_padded}...")
                         final_latents = self.pipeline(**second_pass_kwargs).images