Only at the end

app.py

@@ -995,7 +995,6 @@ def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed,
                 image_encoder=image_encoder, feature_extractor=feature_extractor, device=gpu
             )[:2]
             end_latent = end_latent.to(dtype=torch.float32, device=cpu)
-            end_latent = end_latent.expand(-1, -1, 1 + end_stillness, -1, -1)
         else:
             end_latent = end_clip_embedding = None
 
@@ -1030,7 +1029,13 @@ def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed,
             def callback(d):
                 return
 
-        def compute_latent(history_latents, latent_window_size, latent_padding_size, num_clean_frames, start_latent, end_latent, end_stillness):
+        def compute_latent(history_latents, latent_window_size, latent_padding_size, num_clean_frames, start_latent, end_latent, end_stillness, is_end_of_video):
+            if is_end_of_video:
+                local_end_stillness = end_stillness
+                local_end_latent = end_latent.expand(-1, -1, 1 + local_end_stillness, -1, -1)
+            else:
+                local_end_stillness = 0
+                local_end_latent = end_latent
             # 20250506 pftq: Use user-specified number of context frames, matching original allocation for num_clean_frames=2
             available_frames = history_latents.shape[2]  # Number of latent frames
             max_pixel_frames = min(latent_window_size * 4 - 3, available_frames * 4)  # Cap at available pixel frames
@@ -1045,9 +1050,9 @@ def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed,
             total_context_frames = min(total_context_frames, available_frames)  # 20250507 pftq: Edge case for <=1 sec videos
 
             post_frames = 100 # Single frame for end_latent, otherwise padding causes still image
-            indices = torch.arange(0, 1 + num_4x_frames + num_2x_frames + effective_clean_frames + adjusted_latent_frames + ((latent_padding_size + 1 + end_stillness) if end_latent is not None else 0)).unsqueeze(0) # 20250507 pftq: latent_window_size to adjusted_latent_frames for edge case for <=1 sec videos
+            indices = torch.arange(0, 1 + num_4x_frames + num_2x_frames + effective_clean_frames + adjusted_latent_frames + ((latent_padding_size + 1 + local_end_stillness) if end_latent is not None else 0)).unsqueeze(0) # 20250507 pftq: latent_window_size to adjusted_latent_frames for edge case for <=1 sec videos
             clean_latent_indices_start, clean_latent_4x_indices, clean_latent_2x_indices, clean_latent_1x_indices, latent_indices, blank_indices, clean_latent_indices_post = indices.split(
-                [1, num_4x_frames, num_2x_frames, effective_clean_frames, adjusted_latent_frames, latent_padding_size if end_latent is not None else 0, (1 + end_stillness) if end_latent is not None else 0], dim=1 # 20250507 pftq: latent_window_size to adjusted_latent_frames for edge case for <=1 sec videos
+                [1, num_4x_frames, num_2x_frames, effective_clean_frames, adjusted_latent_frames, latent_padding_size if end_latent is not None else 0, (1 + local_end_stillness) if end_latent is not None else 0], dim=1 # 20250507 pftq: latent_window_size to adjusted_latent_frames for edge case for <=1 sec videos
             )
             clean_latent_indices = torch.cat([clean_latent_indices_start, clean_latent_1x_indices, clean_latent_indices_post], dim=1)
 
@@ -1083,7 +1088,7 @@ def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed,
                         clean_latents_1x = splits[split_idx]
 
             if end_latent is not None:
-                clean_latents = torch.cat([start_latent, clean_latents_1x, end_latent], dim=2)
+                clean_latents = torch.cat([start_latent, clean_latents_1x, local_end_latent], dim=2)
             else:
                 clean_latents = torch.cat([start_latent, clean_latents_1x], dim=2)
 
@@ -1138,7 +1143,7 @@ def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed,
                 else:
                     transformer.initialize_teacache(enable_teacache=False)
 
-                [max_frames, clean_latents, clean_latents_2x, clean_latents_4x, latent_indices, clean_latents, clean_latent_indices, clean_latent_2x_indices, clean_latent_4x_indices] = compute_latent(history_latents, latent_window_size, latent_padding_size, num_clean_frames, start_latent, end_latent, end_stillness)
+                [max_frames, clean_latents, clean_latents_2x, clean_latents_4x, latent_indices, clean_latents, clean_latent_indices, clean_latent_2x_indices, clean_latent_4x_indices] = compute_latent(history_latents, latent_window_size, latent_padding_size, num_clean_frames, start_latent, end_latent, end_stillness, is_end_of_video)
 
                 generated_latents = sample_hunyuan(
                     transformer=transformer,