Video+end frame

app.py

@@ -10,9 +10,7 @@ except:
     class spaces():
         def GPU(*args, **kwargs):
             def decorator(function):
-                def new_function(*dummy_args, **dummy_kwargs):
-                    return function(*dummy_args, **dummy_kwargs)
-                return new_function
+                return lambda *dummy_args, **dummy_kwargs: function(*dummy_args, **dummy_kwargs)
             return decorator
 
 import gradio as gr
@@ -24,6 +22,7 @@ import numpy as np
 import random
 import time
 import math
+import gc
 # 20250506 pftq: Added for video input loading
 import decord
 # 20250506 pftq: Added for progress bars in video_encode
@@ -309,8 +308,67 @@ def set_mp4_comments_imageio_ffmpeg(input_file, comments):
         print(f"Error saving prompt to video metadata, ffmpeg may be required: "+str(e))
         return False
 
+# 20250507 pftq: New function to encode a single image (end frame)
 @torch.no_grad()
-def worker(input_image, end_image, image_position, prompts, n_prompt, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number):
+def image_encode(image_np, target_width, target_height, vae, image_encoder, feature_extractor, device="cuda"):
+    """
+    Encode a single image into a latent and compute its CLIP vision embedding.
+
+    Args:
+        image_np: Input image as numpy array.
+        target_width, target_height: Exact resolution to resize the image to (matches start frame).
+        vae: AutoencoderKLHunyuanVideo model.
+        image_encoder: SiglipVisionModel for CLIP vision encoding.
+        feature_extractor: SiglipImageProcessor for preprocessing.
+        device: Device for computation (e.g., "cuda").
+
+    Returns:
+        latent: Latent representation of the image (shape: [1, channels, 1, height//8, width//8]).
+        clip_embedding: CLIP vision embedding of the image.
+        processed_image_np: Processed image as numpy array (after resizing).
+    """
+    # 20250507 pftq: Process end frame with exact start frame dimensions
+    print("Processing end frame...")
+    try:
+        print(f"Using exact start frame resolution for end frame: {target_width}x{target_height}")
+
+        # Resize and preprocess image to match start frame
+        processed_image_np = resize_and_center_crop(image_np, target_width=target_width, target_height=target_height)
+
+        # Convert to tensor and normalize
+        image_pt = torch.from_numpy(processed_image_np).float() / 127.5 - 1
+        image_pt = image_pt.permute(2, 0, 1).unsqueeze(0).unsqueeze(2)  # Shape: [1, channels, 1, height, width]
+        image_pt = image_pt.to(device)
+
+        # Move VAE to device
+        vae.to(device)
+
+        # Encode to latent
+        latent = vae_encode(image_pt, vae)
+        print(f"image_encode vae output shape: {latent.shape}")
+
+        # Move image encoder to device
+        image_encoder.to(device)
+
+        # Compute CLIP vision embedding
+        clip_embedding = hf_clip_vision_encode(processed_image_np, feature_extractor, image_encoder).last_hidden_state
+
+        # Move models back to CPU and clear cache
+        if device == "cuda":
+            vae.to(cpu)
+            image_encoder.to(cpu)
+            torch.cuda.empty_cache()
+            print("VAE and image encoder moved back to CPU, CUDA cache cleared")
+
+        print(f"End latent shape: {latent.shape}")
+        return latent, clip_embedding, processed_image_np
+
+    except Exception as e:
+        print(f"Error in image_encode: {str(e)}")
+        raise
+
+@torch.no_grad()
+def worker(input_image, end_image, image_position, end_stillness, prompts, n_prompt, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number):
     def encode_prompt(prompt, n_prompt):
         llama_vec, clip_l_pooler = encode_prompt_conds(prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
 
@@ -405,6 +463,8 @@ def worker(input_image, end_image, image_position, prompts, n_prompt, seed, reso
             return [start_latent, image_encoder_last_hidden_state]
             
         [start_latent, image_encoder_last_hidden_state] = get_start_latent(input_image, height, width, vae, gpu, image_encoder, high_vram)
+        del input_image
+        del end_image
 
         # Dtype
 
@@ -500,7 +560,7 @@ def worker(input_image, end_image, image_position, prompts, n_prompt, seed, reso
             [llama_vec, clip_l_pooler, llama_vec_n, clip_l_pooler_n, llama_attention_mask, llama_attention_mask_n] = prompt_parameters[prompt_index]
             
             if prompt_index < len(prompt_parameters) - 1 or (prompt_index == total_latent_sections - 1):
-                prompt_parameters[prompt_index] = None
+                del prompt_parameters[prompt_index]
 
             if not high_vram:
                 unload_complete_models()
@@ -548,6 +608,13 @@ def worker(input_image, end_image, image_position, prompts, n_prompt, seed, reso
                 clean_latent_4x_indices=clean_latent_4x_indices,
                 callback=callback,
             )
+            del clean_latents
+            del clean_latents_2x
+            del clean_latents_4x
+            del latent_indices
+            del clean_latent_indices
+            del clean_latent_2x_indices
+            del clean_latent_4x_indices
 
             [total_generated_latent_frames, history_latents, history_pixels] = post_process(forward, generated_latents, total_generated_latent_frames, history_latents, high_vram, transformer, gpu, vae, history_pixels, latent_window_size, enable_preview, section_index, total_latent_sections, outputs_folder, mp4_crf, stream)
             
@@ -561,7 +628,8 @@ def worker(input_image, end_image, image_position, prompts, n_prompt, seed, reso
                     real_history_latents = history_latents[:, :, :total_generated_latent_frames, :, :]
                     zero_latents = history_latents[:, :, total_generated_latent_frames:, :, :]
                     history_latents = torch.cat([zero_latents, real_history_latents], dim=2)
-                    real_history_latents = zero_latents = None
+                    del real_history_latents
+                    del zero_latents
                     
                     forward = True
                     section_index = first_section_index
@@ -580,7 +648,7 @@ def worker(input_image, end_image, image_position, prompts, n_prompt, seed, reso
     return
 
 @torch.no_grad()
-def worker_start_end(input_image, end_image, image_position, prompts, n_prompt, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number):
+def worker_start_end(input_image, end_image, image_position, end_stillness, prompts, n_prompt, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number):
     def encode_prompt(prompt, n_prompt):
         llama_vec, clip_l_pooler = encode_prompt_conds(prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
 
@@ -689,6 +757,8 @@ def worker_start_end(input_image, end_image, image_position, prompts, n_prompt,
             return [start_latent, end_latent, image_encoder_last_hidden_state]
             
         [start_latent, end_latent, image_encoder_last_hidden_state] = get_start_latent(input_image, has_end_image, end_image, height, width, vae, gpu, image_encoder, high_vram)
+        del input_image
+        del end_image
 
         # Dtype
         image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
@@ -708,7 +778,7 @@ def worker_start_end(input_image, end_image, image_position, prompts, n_prompt,
         total_generated_latent_frames = 0
 
         if total_latent_sections > 4:
-            # In theory the latent_paddings should follow the above sequence, but it seems that duplicating some
+            # In theory the latent_paddings should follow the else sequence, but it seems that duplicating some
             # items looks better than expanding it when total_latent_sections > 4
             # One can try to remove below trick and just
             # use `latent_paddings = list(reversed(range(total_latent_sections)))` to compare
@@ -787,15 +857,15 @@ def worker_start_end(input_image, end_image, image_position, prompts, n_prompt,
             if len(prompt_parameters) > 0:
                 [llama_vec, clip_l_pooler, llama_vec_n, clip_l_pooler_n, llama_attention_mask, llama_attention_mask_n] = prompt_parameters.pop(len(prompt_parameters) - 1)
 
-            indices = torch.arange(1 + latent_padding_size + latent_window_size + 1 + 2 + 16).unsqueeze(0)
-            clean_latent_indices_pre, blank_indices, latent_indices, clean_latent_indices_post, clean_latent_2x_indices, clean_latent_4x_indices = indices.split([1, latent_padding_size, latent_window_size, 1, 2, 16], dim=1)
+            indices = torch.arange(1 + latent_padding_size + latent_window_size + 1 + (end_stillness if is_first_section else 0) + 2 + 16).unsqueeze(0)
+            clean_latent_indices_pre, blank_indices, latent_indices, clean_latent_indices_post, clean_latent_2x_indices, clean_latent_4x_indices = indices.split([1, latent_padding_size, latent_window_size, 1 + (end_stillness if is_first_section else 0), 2, 16], dim=1)
             clean_latent_indices = torch.cat([clean_latent_indices_pre, clean_latent_indices_post], dim=1)
             
             clean_latents_post, clean_latents_2x, clean_latents_4x = history_latents[:, :, :1 + 2 + 16, :, :].split([1, 2, 16], dim=2)
 
             # Use end image latent for the first section if provided
             if has_end_image and is_first_section:
-                clean_latents_post = end_latent
+                clean_latents_post = end_latent.expand(-1, -1, 1 + end_stillness, -1, -1)
 
             clean_latents = torch.cat([start_latent, clean_latents_post], dim=2)
 
@@ -838,6 +908,13 @@ def worker_start_end(input_image, end_image, image_position, prompts, n_prompt,
                 clean_latent_4x_indices=clean_latent_4x_indices,
                 callback=callback,
             )
+            del clean_latents
+            del clean_latents_2x
+            del clean_latents_4x
+            del latent_indices
+            del clean_latent_indices
+            del clean_latent_2x_indices
+            del clean_latent_4x_indices
 
             [total_generated_latent_frames, history_latents, history_pixels] = post_process(job_id, start_latent, generated_latents, total_generated_latent_frames, history_latents, high_vram, transformer, gpu, vae, history_pixels, latent_window_size, enable_preview, outputs_folder, mp4_crf, stream, is_last_section)
 
@@ -856,7 +933,7 @@ def worker_start_end(input_image, end_image, image_position, prompts, n_prompt,
 
 # 20250506 pftq: Modified worker to accept video input and clean frame count
 @torch.no_grad()
-def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
+def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
     def encode_prompt(prompt, n_prompt):
         llama_vec, clip_l_pooler = encode_prompt_conds(prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
 
@@ -882,6 +959,7 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
 
         # 20250506 pftq: Encode video
         start_latent, input_image_np, video_latents, fps, height, width = video_encode(input_video, resolution, no_resize, vae, vae_batch_size=vae_batch, device=gpu)
+        del input_video
         start_latent = start_latent.to(dtype=torch.float32, device=cpu)
         video_latents = video_latents.cpu()
 
@@ -919,12 +997,29 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
             load_model_as_complete(image_encoder, target_device=gpu)
 
         image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
+        del input_image_np
+
+        # 20250507 pftq: Process end frame if provided
+        if end_frame is not None:
+            if not high_vram:
+                load_model_as_complete(vae, target_device=gpu)
+
+            stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'End frame encoding ...'))))
+            end_latent = image_encode(
+                end_frame, target_width=width, target_height=height, vae=vae,
+                image_encoder=image_encoder, feature_extractor=feature_extractor, device=gpu
+            )[0]
+            del end_frame
+            end_latent = end_latent.to(dtype=torch.float32, device=cpu)
+        else:
+            end_latent = None
 
         # Clean GPU
         if not high_vram:
-            unload_complete_models(image_encoder)
+            unload_complete_models(image_encoder, vae)
 
         image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
+        del image_encoder_output
 
         # Dtype
         image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
@@ -951,7 +1046,13 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
             def callback(d):
                 return
 
-        def compute_latent(history_latents, latent_window_size, num_clean_frames, start_latent):
+        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
@@ -965,11 +1066,12 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
             total_context_frames = num_4x_frames + num_2x_frames + effective_clean_frames
             total_context_frames = min(total_context_frames, available_frames)  # 20250507 pftq: Edge case for <=1 sec videos
 
-            indices = torch.arange(0, 1 + num_4x_frames + num_2x_frames + effective_clean_frames + adjusted_latent_frames).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 = indices.split(
-                [1, num_4x_frames, num_2x_frames, effective_clean_frames, adjusted_latent_frames], dim=1 # 20250507 pftq: latent_window_size to adjusted_latent_frames for 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 + 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 + 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], dim=1)
+            clean_latent_indices = torch.cat([clean_latent_indices_start, clean_latent_1x_indices, clean_latent_indices_post], dim=1)
 
             # 20250506 pftq: Split history_latents dynamically based on available frames
             fallback_frame_count = 2 # 20250507 pftq: Changed 0 to 2 Edge case for <=1 sec videos
@@ -1002,7 +1104,10 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
                     if effective_clean_frames > 0 and split_idx < len(splits):
                         clean_latents_1x = splits[split_idx]
 
-            clean_latents = torch.cat([start_latent, clean_latents_1x], dim=2)
+            if end_latent is not None:
+                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)
 
             # 20250507 pftq: Fix for <=1 sec videos.
             max_frames = min(latent_window_size * 4 - 3, history_latents.shape[2] * 4)
@@ -1024,10 +1129,18 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
             history_latents = video_latents
             total_generated_latent_frames = history_latents.shape[2]
             # 20250506 pftq: Initialize history_pixels to fix UnboundLocalError
-            history_pixels = None
-            previous_video = None
+            history_pixels = previous_video = None
+
+            # 20250509 Generate backwards with end frame for better end frame anchoring
+            if total_latent_sections > 4:
+                latent_paddings = [3] + [2] * (total_latent_sections - 3) + [1, 0]
+            else:
+                latent_paddings = list(reversed(range(total_latent_sections)))
 
-            for section_index in range(total_latent_sections):
+            for section_index, latent_padding in enumerate(latent_paddings):
+                is_start_of_video = latent_padding == 0
+                is_end_of_video = latent_padding == latent_paddings[0]
+                latent_padding_size = latent_padding * latent_window_size
                 if stream.input_queue.top() == 'end':
                     stream.output_queue.push(('end', None))
                     return
@@ -1046,7 +1159,7 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
                 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, num_clean_frames, start_latent)
+                [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,
@@ -1077,6 +1190,13 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
                     clean_latent_4x_indices=clean_latent_4x_indices,
                     callback=callback,
                 )
+                del clean_latents
+                del clean_latents_2x
+                del clean_latents_4x
+                del latent_indices
+                del clean_latent_indices
+                del clean_latent_2x_indices
+                del clean_latent_4x_indices
 
                 total_generated_latent_frames += int(generated_latents.shape[2])
                 history_latents = torch.cat([history_latents, generated_latents.to(history_latents)], dim=2)
@@ -1134,17 +1254,17 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
     stream.output_queue.push(('end', None))
     return
 
-def get_duration(input_image, end_image, image_position, prompts, generation_mode, n_prompt, seed, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number):
+def get_duration(input_image, end_image, image_position, end_stillness, prompts, generation_mode, n_prompt, seed, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number):
     return allocation_time
 
 @spaces.GPU(duration=get_duration)
-def process_on_gpu(input_image, end_image, image_position, prompts, generation_mode, n_prompt, seed, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number
+def process_on_gpu(input_image, end_image, image_position, end_stillness, prompts, generation_mode, n_prompt, seed, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number
            ):
     start = time.time()
     global stream
     stream = AsyncStream()
 
-    async_run(worker_start_end if generation_mode == "start_end" else worker, input_image, end_image, image_position, prompts, n_prompt, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number)
+    async_run(worker_start_end if generation_mode == "start_end" else worker, input_image, end_image, image_position, end_stillness, prompts, n_prompt, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number)
 
     output_filename = None
 
@@ -1170,12 +1290,13 @@ def process_on_gpu(input_image, end_image, image_position, prompts, generation_m
             ((str(hours) + " h, ") if hours != 0 else "") + \
             ((str(minutes) + " min, ") if hours != 0 or minutes != 0 else "") + \
             str(secondes) + " sec. " + \
-            "You can upscale the result with RIFE. To make all your generated scenes consistent, you can then apply a face swap on the main character. If you do not see the generated video above, the process may have failed. See the logs for more information. If you see an error like ''NVML_SUCCESS == r INTERNAL ASSERT FAILED'', you probably haven't enough VRAM. Test an example or other options to compare. You can share your inputs to the original space or set your space in public for a peer review.", gr.update(interactive=True), gr.update(interactive=False), gr.update(visible = False)
+            "You can upscale the result with https://huggingface.co/spaces/Nick088/Real-ESRGAN_Pytorch. To make all your generated scenes consistent, you can then apply a face swap on the main character. If you do not see the generated video above, the process may have failed. See the logs for more information. If you see an error like ''NVML_SUCCESS == r INTERNAL ASSERT FAILED'', you probably haven't enough VRAM. Test an example or other options to compare. You can share your inputs to the original space or set your space in public for a peer review.", gr.update(interactive=True), gr.update(interactive=False), gr.update(visible = False)
             break
 
 def process(input_image,
             end_image,
             image_position=0,
+            end_stillness=1,
             prompt="",
             generation_mode="image",
             n_prompt="",
@@ -1209,17 +1330,20 @@ def process(input_image,
 
     prompts = prompt.split(";")
 
-    # assert input_image is not None, 'No input image!'
     if generation_mode == "text":
-        default_height, default_width = 640, 640
+        default_height, default_width = resolution, resolution
         input_image = np.ones((default_height, default_width, 3), dtype=np.uint8) * 255
         print("No input image provided. Using a blank white image.")
+    assert input_image is not None, 'No input image!'
+    assert (generation_mode != "start_end") or end_image is not None, 'No end image!'
 
     yield gr.update(label="Previewed Frames"), None, '', '', gr.update(interactive=False), gr.update(interactive=True), gr.skip()
 
+    gc.collect()
     yield from process_on_gpu(input_image,
             end_image,
             image_position,
+            end_stillness,
             prompts,
             generation_mode,
             n_prompt,
@@ -1239,17 +1363,17 @@ def process(input_image,
             fps_number
            )
 
-def get_duration_video(input_video, prompts, n_prompt, seed, batch, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
+def get_duration_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed, batch, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
     return allocation_time
 
 @spaces.GPU(duration=get_duration_video)
-def process_video_on_gpu(input_video, prompts, n_prompt, seed, batch, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
+def process_video_on_gpu(input_video, end_frame, end_stillness, prompts, n_prompt, seed, batch, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
     start = time.time()
     global stream
     stream = AsyncStream()
 
     # 20250506 pftq: Pass num_clean_frames, vae_batch, etc
-    async_run(worker_video, input_video, prompts, n_prompt, seed, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch)
+    async_run(worker_video, input_video, end_frame, end_stillness, prompts, n_prompt, seed, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch)
 
     output_filename = None
 
@@ -1276,10 +1400,10 @@ def process_video_on_gpu(input_video, prompts, n_prompt, seed, batch, resolution
             ((str(hours) + " h, ") if hours != 0 else "") + \
             ((str(minutes) + " min, ") if hours != 0 or minutes != 0 else "") + \
             str(secondes) + " sec. " + \
-            " You can upscale the result with RIFE. To make all your generated scenes consistent, you can then apply a face swap on the main character. If you do not see the generated video above, the process may have failed. See the logs for more information. If you see an error like ''NVML_SUCCESS == r INTERNAL ASSERT FAILED'', you probably haven't enough VRAM. Test an example or other options to compare. You can share your inputs to the original space or set your space in public for a peer review.", '', gr.update(interactive=True), gr.update(interactive=False), gr.update(visible = False)
+            " You can upscale the result with https://huggingface.co/spaces/Nick088/Real-ESRGAN_Pytorch. To make all your generated scenes consistent, you can then apply a face swap on the main character. If you do not see the generated video above, the process may have failed. See the logs for more information. If you see an error like ''NVML_SUCCESS == r INTERNAL ASSERT FAILED'', you probably haven't enough VRAM. Test an example or other options to compare. You can share your inputs to the original space or set your space in public for a peer review.", '', gr.update(interactive=True), gr.update(interactive=False), gr.update(visible = False)
             break
 
-def process_video(input_video, prompt, n_prompt, randomize_seed, seed, auto_allocation, allocation_time, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
+def process_video(input_video, end_frame, end_stillness, prompt, n_prompt, randomize_seed, seed, auto_allocation, allocation_time, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
     global high_vram
     if auto_allocation:
         allocation_time = min(total_second_length * 60 * (2.5 if use_teacache else 3.5) * (1 + ((steps - 25) / 25))**2, 600)
@@ -1312,7 +1436,8 @@ def process_video(input_video, prompt, n_prompt, randomize_seed, seed, auto_allo
     if cfg > 1:
         gs = 1
 
-    yield from process_video_on_gpu(input_video, prompts, n_prompt, seed, batch, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch)
+    gc.collect()
+    yield from process_video_on_gpu(input_video, end_frame, end_stillness, prompt, n_prompt, seed, batch, resolution, total_second_length, allocation_time, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch)
 
 def end_process():
     stream.input_queue.push('end')
@@ -1382,12 +1507,12 @@ with block:
     local_storage = gr.BrowserState(default_local_storage)
     with gr.Row():
         with gr.Column():
-            generation_mode = gr.Radio([["Text-to-Video", "text"], ["Image-to-Video", "image"], ["Start & end frames", "start_end"], ["Video Extension", "video"]], elem_id="generation-mode", label="Generation mode", value = "image")
+            generation_mode = gr.Radio([["Text-to-Video", "text"], ["Image-to-Video", "image"], ["Start & end frames", "start_end"], ["Video Extension", "video"]], elem_id="generation-mode", label="Input mode", value = "image")
             text_to_video_hint = gr.HTML("Text-to-Video badly works with a flash effect at the start. I discourage to use the Text-to-Video feature. You should rather generate an image with Flux and use Image-to-Video. You will save time.")
             input_image = gr.Image(sources='upload', type="numpy", label="Image", height=320)
-            end_image = gr.Image(sources='upload', type="numpy", label="End Frame (Optional)", height=320)
             image_position = gr.Slider(label="Image position", minimum=0, maximum=100, value=0, step=1, info='0=Video start; 100=Video end (lower quality)')
             input_video = gr.Video(sources='upload', label="Input Video", height=320)
+            end_image = gr.Image(sources='upload', type="numpy", label="End Frame (optional)", height=320)
             timeless_prompt = gr.Textbox(label="Timeless prompt", info='Used on the whole duration of the generation', value='', placeholder="The creature starts to move, fast motion, fixed camera, focus motion, consistent arm, consistent position, mute colors, insanely detailed")
             prompt_number = gr.Slider(label="Timed prompt number", minimum=0, maximum=1000, value=0, step=1, info='Prompts will automatically appear')
 
@@ -1414,6 +1539,7 @@ with block:
                 n_prompt = gr.Textbox(label="Negative Prompt", value="Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", info='Requires using normal CFG (undistilled) instead of Distilled (set Distilled=1 and CFG > 1).')
 
                 fps_number = gr.Slider(label="Frame per seconds", info="The model is trained for 30 fps so other fps may generate weird results", minimum=10, maximum=60, value=30, step=1)
+                end_stillness = gr.Slider(label="End stillness", minimum=0, maximum=100, value=1, step=1, info='0=Realistic end; >0=Matches exactly the end image (but the time seems to freeze)')
 
                 latent_window_size = gr.Slider(label="Latent Window Size", minimum=1, maximum=33, value=9, step=1, info='Generate more frames at a time (larger chunks). Less degradation and better blending but higher VRAM cost. Should not change.')
                 steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=30, step=1, info='Increase for more quality, especially if using high non-distilled CFG. If your animation has very few motion, you may have brutal brightness change; this can be fixed increasing the steps.')
@@ -1466,8 +1592,8 @@ with block:
             progress_desc = gr.Markdown('', elem_classes='no-generating-animation')
             progress_bar = gr.HTML('', elem_classes='no-generating-animation')
 
-    ips = [input_image, end_image, image_position, final_prompt, generation_mode, n_prompt, randomize_seed, seed, auto_allocation, allocation_time, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number]
-    ips_video = [input_video, final_prompt, n_prompt, randomize_seed, seed, auto_allocation, allocation_time, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch]
+    ips = [input_image, end_image, image_position, end_stillness, final_prompt, generation_mode, n_prompt, randomize_seed, seed, auto_allocation, allocation_time, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf, fps_number]
+    ips_video = [input_video, end_image, end_stillness, final_prompt, n_prompt, randomize_seed, seed, auto_allocation, allocation_time, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch]
 
     gr.Examples(
         label = "✍️ Examples from text",
@@ -1476,6 +1602,7 @@ with block:
                     None, # input_image
                     None, # end_image
                     0, # image_position
+                    1, # end_stillness
                     "Overcrowed street in Japan, photorealistic, realistic, intricate details, 8k, insanely detailed",
                     "text", # generation_mode
                     "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", # n_prompt
@@ -1511,6 +1638,7 @@ with block:
                     "./img_examples/Example1.png", # input_image
                     None, # end_image
                     0, # image_position
+                    1, # end_stillness
                     "A dolphin emerges from the water, photorealistic, realistic, intricate details, 8k, insanely detailed",
                     "image", # generation_mode
                     "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", # n_prompt
@@ -1535,6 +1663,7 @@ with block:
                     "./img_examples/Example2.webp", # input_image
                     None, # end_image
                     0, # image_position
+                    1, # end_stillness
                     "A man on the left and a woman on the right face each other ready to start a conversation, large space between the persons, full view, full-length view, 3D, pixar, 3D render, CGI. The man talks and the woman listens; A man on the left and a woman on the right face each other ready to start a conversation, large space between the persons, full view, full-length view, 3D, pixar, 3D render, CGI. The woman talks, the man stops talking and the man listens; A man on the left and a woman on the right face each other ready to start a conversation, large space between the persons, full view, full-length view, 3D, pixar, 3D render, CGI. The woman talks and the man listens",
                     "image", # generation_mode
                     "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", # n_prompt
@@ -1559,6 +1688,7 @@ with block:
                     "./img_examples/Example2.webp", # input_image
                     None, # end_image
                     0, # image_position
+                    1, # end_stillness
                     "A man on the left and a woman on the right face each other ready to start a conversation, large space between the persons, full view, full-length view, 3D, pixar, 3D render, CGI. The woman talks and the man listens; A man on the left and a woman on the right face each other ready to start a conversation, large space between the persons, full view, full-length view, 3D, pixar, 3D render, CGI. The man talks, the woman stops talking and the woman listens A man on the left and a woman on the right face each other ready to start a conversation, large space between the persons, full view, full-length view, 3D, pixar, 3D render, CGI. The man talks and the woman listens",
                     "image", # generation_mode
                     "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", # n_prompt
@@ -1583,6 +1713,7 @@ with block:
                     "./img_examples/Example3.jpg", # input_image
                     None, # end_image
                     0, # image_position
+                    1, # end_stillness
                     "एउटा केटा दायाँतिर हिँडिरहेको छ, पूर्ण दृश्य, पूर्ण-लम्बाइको दृश्य, कार्टुन",
                     "image", # generation_mode
                     "हात छुटेको, लामो हात, अवास्तविक स्थिति, असम्भव विकृति, देखिने हड्डी, मांसपेशी संकुचन, कमजोर फ्रेम, धमिलो, धमिलो, अत्यधिक चिल्लो", # n_prompt
@@ -1607,6 +1738,7 @@ with block:
                     "./img_examples/Example4.webp", # input_image
                     None, # end_image
                     100, # image_position
+                    1, # end_stillness
                     "A building starting to explode, photorealistic, realisitc, 8k, insanely detailed",
                     "image", # generation_mode
                     "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", # n_prompt
@@ -1642,9 +1774,10 @@ with block:
                     "./img_examples/Example5.png", # input_image
                     "./img_examples/Example6.png", # end_image
                     0, # image_position
-                    "A woman jumps out of the train and arrives on the ground, viewed from the outside, photorealistic, realistic, amateur photography, midday, insanely detailed, 8k",
+                    0, # end_stillness
+                    "A woman jumps out of the train and arrives on the ground, viewed from the outside, photorealistic, realistic, amateur photography, midday, insanely detailed, 8k", # prompt
                     "start_end", # generation_mode
-                    "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", # n_prompt
+                    "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth, jumpcut, crossfader, crossfading", # n_prompt
                     True, # randomize_seed
                     42, # seed
                     True, # auto_allocation
@@ -1675,8 +1808,36 @@ with block:
         examples = [
                 [
                     "./img_examples/Example1.mp4", # input_video
+                    None, # end_image
+                    1, # end_stillness
                     "View of the sea as far as the eye can see, from the seaside, a piece of land is barely visible on the horizon at the middle, the sky is radiant, reflections of the sun in the water, photorealistic, realistic, intricate details, 8k, insanely detailed",
-                    "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth", # n_prompt
+                    "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth, jumpcut, crossfader, crossfading", # n_prompt
+                    True, # randomize_seed
+                    42, # seed
+                    True, # auto_allocation
+                    180, # allocation_time
+                    1, # batch
+                    672, # resolution
+                    1, # total_second_length
+                    9, # latent_window_size
+                    30, # steps
+                    1.0, # cfg
+                    10.0, # gs
+                    0.0, # rs
+                    6, # gpu_memory_preservation
+                    False, # enable_preview
+                    True, # use_teacache
+                    False, # no_resize
+                    16, # mp4_crf
+                    5, # num_clean_frames
+                    default_vae
+                ],
+                [
+                    "./img_examples/Example1.mp4", # input_video
+                    "./img_examples/Example1.png", # end_image
+                    1, # end_stillness
+                    "View of the sea as far as the eye can see, from the seaside, a piece of land is barely visible on the horizon at the middle, the sky is radiant, reflections of the sun in the water, photorealistic, realistic, intricate details, 8k, insanely detailed",
+                    "Missing arm, long hand, unrealistic position, impossible contortion, visible bone, muscle contraction, poorly framed, blurred, blurry, over-smooth, jumpcut, crossfader, crossfading", # n_prompt
                     True, # randomize_seed
                     42, # seed
                     True, # auto_allocation
@@ -1734,6 +1895,7 @@ with block:
             gr.update(visible = False), # image_position
             gr.update(visible = False), # input_image
             gr.update(visible = False), # end_image
+            gr.update(visible = False), # end_stillness
             gr.update(visible = False), # input_video
             gr.update(visible = True),  # start_button
             gr.update(visible = False), # start_button_video
@@ -1750,6 +1912,7 @@ with block:
             gr.update(visible = True),  # image_position
             gr.update(visible = True),  # input_image
             gr.update(visible = False), # end_image
+            gr.update(visible = False), # end_stillness
             gr.update(visible = False), # input_video
             gr.update(visible = True),  # start_button
             gr.update(visible = False), # start_button_video
@@ -1766,6 +1929,7 @@ with block:
             gr.update(visible = False), # image_position
             gr.update(visible = True),  # input_image
             gr.update(visible = True),  # end_image
+            gr.update(visible = True), # end_stillness
             gr.update(visible = False), # input_video
             gr.update(visible = True),  # start_button
             gr.update(visible = False), # start_button_video
@@ -1781,7 +1945,8 @@ with block:
             gr.update(visible = False), # text_to_video_hint
             gr.update(visible = False), # image_position
             gr.update(visible = False), # input_image
-            gr.update(visible = False), # end_image
+            gr.update(visible = True), # end_image
+            gr.update(visible = True), # end_stillness
             gr.update(visible = True),  # input_video
             gr.update(visible = False), # start_button
             gr.update(visible = True),  # start_button_video
@@ -1813,7 +1978,7 @@ with block:
     generation_mode.change(
         fn=handle_generation_mode_change,
         inputs=[generation_mode],
-        outputs=[text_to_video_hint, image_position, input_image, end_image, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch, prompt_hint, fps_number]
+        outputs=[text_to_video_hint, image_position, input_image, end_image, end_stillness, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch, prompt_hint, fps_number]
     )
 
     # Update display when the page loads
@@ -1821,7 +1986,7 @@ with block:
         fn=handle_generation_mode_change, inputs = [
         generation_mode
     ], outputs = [
-       text_to_video_hint, image_position, input_image, end_image, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch, prompt_hint, fps_number
+       text_to_video_hint, image_position, input_image, end_image, end_stillness, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch, prompt_hint, fps_number
     ]
     )