Generation time

app.py

@@ -12,6 +12,7 @@ import einops
 import safetensors.torch as sf
 import numpy as np
 import random
+import time
 import math
 # 20250506 pftq: Added for video input loading
 import decord
@@ -422,6 +423,37 @@ def worker(input_image, prompts, n_prompt, seed, resolution, total_second_length
         clean_latent_indices_start, clean_latent_4x_indices, clean_latent_2x_indices, clean_latent_1x_indices, latent_indices = indices.split([1, 16, 2, 1, latent_window_size], dim=1)
         clean_latent_indices = torch.cat([clean_latent_indices_start, clean_latent_1x_indices], dim=1)
 
+        def post_process(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):
+            total_generated_latent_frames += int(generated_latents.shape[2])
+            history_latents = torch.cat([history_latents, generated_latents.to(history_latents)], dim=2)
+
+            if not high_vram:
+                offload_model_from_device_for_memory_preservation(transformer, target_device=gpu, preserved_memory_gb=8)
+                load_model_as_complete(vae, target_device=gpu)
+
+            if history_pixels is None:
+                real_history_latents = history_latents[:, :, -total_generated_latent_frames:, :, :]
+                history_pixels = vae_decode(real_history_latents, vae).cpu()
+            else:
+                section_latent_frames = latent_window_size * 2
+                overlapped_frames = latent_window_size * 4 - 3
+
+                real_history_latents = history_latents[:, :, max(-section_latent_frames, -total_generated_latent_frames):, :, :]
+                history_pixels = soft_append_bcthw(history_pixels, vae_decode(real_history_latents, vae).cpu(), overlapped_frames)
+
+            if not high_vram:
+                unload_complete_models()
+
+            if enable_preview or section_index == total_latent_sections - 1:
+                output_filename = os.path.join(outputs_folder, f'{job_id}_{total_generated_latent_frames}.mp4')
+
+                save_bcthw_as_mp4(history_pixels, output_filename, fps=30, crf=mp4_crf)
+
+                print(f'Decoded. Current latent shape pixel shape {history_pixels.shape}')
+
+                stream.output_queue.push(('file', output_filename))
+            return [total_generated_latent_frames, history_latents, history_pixels]
+
         for section_index in range(total_latent_sections):
             if stream.input_queue.top() == 'end':
                 stream.output_queue.push(('end', None))
@@ -475,35 +507,7 @@ def worker(input_image, prompts, n_prompt, seed, resolution, total_second_length
                 callback=callback,
             )
 
-            total_generated_latent_frames += int(generated_latents.shape[2])
-            history_latents = torch.cat([history_latents, generated_latents.to(history_latents)], dim=2)
-
-            if not high_vram:
-                offload_model_from_device_for_memory_preservation(transformer, target_device=gpu, preserved_memory_gb=8)
-                load_model_as_complete(vae, target_device=gpu)
-
-            real_history_latents = history_latents[:, :, -total_generated_latent_frames:, :, :]
-
-            if history_pixels is None:
-                history_pixels = vae_decode(real_history_latents, vae).cpu()
-            else:
-                section_latent_frames = latent_window_size * 2
-                overlapped_frames = latent_window_size * 4 - 3
-
-                current_pixels = vae_decode(real_history_latents[:, :, -section_latent_frames:], vae).cpu()
-                history_pixels = soft_append_bcthw(history_pixels, current_pixels, overlapped_frames)
-
-            if not high_vram:
-                unload_complete_models()
-
-            if enable_preview or section_index == total_latent_sections - 1:
-                output_filename = os.path.join(outputs_folder, f'{job_id}_{total_generated_latent_frames}.mp4')
-
-                save_bcthw_as_mp4(history_pixels, output_filename, fps=30, crf=mp4_crf)
-
-                print(f'Decoded. Current latent shape {real_history_latents.shape}; pixel shape {history_pixels.shape}')
-
-                stream.output_queue.push(('file', output_filename))
+            [total_generated_latent_frames, history_latents, history_pixels] = post_process(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)
     except:
         traceback.print_exc()
 
@@ -516,8 +520,7 @@ def worker(input_image, prompts, n_prompt, seed, resolution, total_second_length
     return
 
 def get_duration(input_image, prompt, generation_mode, n_prompt, randomize_seed, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf):
-    return total_second_length * 60 * (0.9 if use_teacache else 1.5) * (2**((resolution - 640) / 640)) * (1 + ((steps - 25) / 100))
-
+    return total_second_length * 60 * (0.9 if use_teacache else 1.5) * (1 + ((steps - 25) / 100))
 
 @spaces.GPU(duration=get_duration)
 def process(input_image, prompt,
@@ -537,6 +540,7 @@ def process(input_image, prompt,
             use_teacache=False,
             mp4_crf=16
            ):
+    start = time.time()
     global stream
 
     if torch.cuda.device_count() == 0:
@@ -575,7 +579,17 @@ def process(input_image, prompt,
             yield gr.update(), gr.update(visible=True, value=preview), desc, html, gr.update(interactive=False), gr.update(interactive=True)
 
         if flag == 'end':
-            yield output_filename, gr.update(visible=False), gr.update(), 'To make all your generated scenes consistent, you can then apply a face swap on the main character.', gr.update(interactive=True), gr.update(interactive=False)
+            end = time.time()
+            secondes = int(end - start)
+            minutes = math.floor(secondes / 60)
+            secondes = secondes - (minutes * 60)
+            hours = math.floor(minutes / 60)
+            minutes = minutes - (hours * 60)
+            yield output_filename, gr.update(visible=False), gr.update(), "The video has been generated in " + \
+            ((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.", gr.update(interactive=True), gr.update(interactive=False)
             break
 
 # 20250506 pftq: Modified worker to accept video input and clean frame count
@@ -663,6 +677,63 @@ 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):
+            # 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
+            adjusted_latent_frames = max(1, (max_pixel_frames + 3) // 4)  # Convert back to latent frames
+            # Adjust num_clean_frames to match original behavior: num_clean_frames=2 means 1 frame for clean_latents_1x
+            effective_clean_frames = max(0, num_clean_frames - 1)
+            effective_clean_frames = min(effective_clean_frames, available_frames - 2) if available_frames > 2 else 0 # 20250507 pftq: changed 1 to 2 for edge case for <=1 sec videos
+            num_2x_frames = min(2, max(1, available_frames - effective_clean_frames - 1)) if available_frames > effective_clean_frames + 1 else 0 # 20250507 pftq: subtracted 1 for edge case for <=1 sec videos
+            num_4x_frames = min(16, max(1, available_frames - effective_clean_frames - num_2x_frames)) if available_frames > effective_clean_frames + num_2x_frames else 0 # 20250507 pftq: Edge case for <=1 sec
+
+            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, sum([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
+            )
+            clean_latent_indices = torch.cat([clean_latent_indices_start, clean_latent_1x_indices], 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
+            context_frames = clean_latents_4x = clean_latents_2x = clean_latents_1x = history_latents[:, :, :fallback_frame_count, :, :]
+
+            if total_context_frames > 0:
+                context_frames = history_latents[:, :, -total_context_frames:, :, :]
+                split_sizes = [num_4x_frames, num_2x_frames, effective_clean_frames]
+                split_sizes = [s for s in split_sizes if s > 0]  # Remove zero sizes
+                if split_sizes:
+                    splits = context_frames.split(split_sizes, dim=2)
+                    split_idx = 0
+                    
+                    if num_4x_frames > 0:
+                        clean_latents_4x = splits[split_idx]
+                        split_idx = 1
+                    if clean_latents_4x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
+                        print("Edge case for <=1 sec videos 4x")
+                        clean_latents_4x = clean_latents_4x.expand(-1, -1, 2, -1, -1)
+                        
+                    if num_2x_frames > 0 and split_idx < len(splits):
+                        clean_latents_2x = splits[split_idx]
+                        if clean_latents_2x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
+                            print("Edge case for <=1 sec videos 2x")
+                            clean_latents_2x = clean_latents_2x.expand(-1, -1, 2, -1, -1)
+                        split_idx += 1
+                    elif clean_latents_2x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
+                        clean_latents_2x = clean_latents_4x
+                        
+                    if effective_clean_frames > 0 and split_idx < len(splits):
+                        clean_latents_1x = splits[split_idx]
+
+            clean_latents = torch.cat([start_latent.to(history_latents), 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)
+            return [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]
+
         for idx in range(batch):
             if batch > 1:
                 print(f"Beginning video {idx+1} of {batch} with seed {seed} ")
@@ -701,60 +772,7 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
                 else:
                     transformer.initialize_teacache(enable_teacache=False)
 
-                # 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
-                adjusted_latent_frames = max(1, (max_pixel_frames + 3) // 4)  # Convert back to latent frames
-                # Adjust num_clean_frames to match original behavior: num_clean_frames=2 means 1 frame for clean_latents_1x
-                effective_clean_frames = max(0, num_clean_frames - 1) if num_clean_frames > 1 else 0
-                effective_clean_frames = min(effective_clean_frames, available_frames - 2) if available_frames > 2 else 0 # 20250507 pftq: changed 1 to 2 for edge case for <=1 sec videos
-                num_2x_frames = min(2, max(1, available_frames - effective_clean_frames - 1)) if available_frames > effective_clean_frames + 1 else 0 # 20250507 pftq: subtracted 1 for edge case for <=1 sec videos
-                num_4x_frames = min(16, max(1, available_frames - effective_clean_frames - num_2x_frames)) if available_frames > effective_clean_frames + num_2x_frames else 0 # 20250507 pftq: Edge case for <=1 sec
-
-                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, sum([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
-                )
-                clean_latent_indices = torch.cat([clean_latent_indices_start, clean_latent_1x_indices], 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
-                context_frames = clean_latents_4x = clean_latents_2x = clean_latents_1x = history_latents[:, :, :fallback_frame_count, :, :]
-
-                if total_context_frames > 0:
-                    context_frames = history_latents[:, :, -total_context_frames:, :, :]
-                    split_sizes = [num_4x_frames, num_2x_frames, effective_clean_frames]
-                    split_sizes = [s for s in split_sizes if s > 0]  # Remove zero sizes
-                    if split_sizes:
-                        splits = context_frames.split(split_sizes, dim=2)
-                        split_idx = 0
-                        
-                        if num_4x_frames > 0:
-                            clean_latents_4x = splits[split_idx]
-                            split_idx = 1
-                        if clean_latents_4x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
-                            print("Edge case for <=1 sec videos 4x")
-                            clean_latents_4x = clean_latents_4x.expand(-1, -1, 2, -1, -1)
-                            
-                        if num_2x_frames > 0 and split_idx < len(splits):
-                            clean_latents_2x = splits[split_idx]
-                            if clean_latents_2x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
-                                print("Edge case for <=1 sec videos 2x")
-                                clean_latents_2x = clean_latents_2x.expand(-1, -1, 2, -1, -1)
-                            split_idx += 1
-                        elif clean_latents_2x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
-                            clean_latents_2x = clean_latents_4x
-                            
-                        if effective_clean_frames > 0 and split_idx < len(splits):
-                            clean_latents_1x = splits[split_idx]
-
-                clean_latents = torch.cat([start_latent.to(history_latents), 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)
+                [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)
 
                 generated_latents = sample_hunyuan(
                     transformer=transformer,
@@ -801,8 +819,7 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
                   section_latent_frames = latent_window_size * 2
                   overlapped_frames = min(latent_window_size * 4 - 3, history_pixels.shape[2])
 
-                  current_pixels = vae_decode(real_history_latents[:, :, -section_latent_frames:], vae).cpu()
-                  history_pixels = soft_append_bcthw(history_pixels, current_pixels, overlapped_frames)
+                  history_pixels = soft_append_bcthw(history_pixels, vae_decode(real_history_latents[:, :, -section_latent_frames:], vae).cpu(), overlapped_frames)
 
                 if not high_vram:
                     unload_complete_models()
@@ -844,11 +861,12 @@ def worker_video(input_video, prompts, n_prompt, seed, batch, resolution, total_
     return
 
 def get_duration_video(input_video, prompt, n_prompt, randomize_seed, 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):
-    return total_second_length * 60 * (0.9 if use_teacache else 2.3) * (2**((resolution - 640) / 640)) * (1 + ((steps - 25) / 100))
+    return total_second_length * 60 * (0.9 if use_teacache else 2.3) * (1 + ((steps - 25) / 100))
 
 # 20250506 pftq: Modified process to pass clean frame count, etc from video_encode
 @spaces.GPU(duration=get_duration_video)
 def process_video(input_video, prompt, n_prompt, randomize_seed, 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):
+    start = time.time()
     global stream, high_vram
 
     if torch.cuda.device_count() == 0:
@@ -899,7 +917,18 @@ def process_video(input_video, prompt, n_prompt, randomize_seed, seed, batch, re
             yield output_filename, gr.update(visible=True, value=preview), desc, html, gr.update(interactive=False), gr.update(interactive=True) # 20250506 pftq: Keep refreshing the video in case it got hidden when the tab was in the background
 
         if flag == 'end':
-            yield output_filename, gr.update(visible=False), desc+' Video complete. To make all your generated scenes consistent, you can then apply a face swap on the main character.', '', gr.update(interactive=True), gr.update(interactive=False)
+            end = time.time()
+            secondes = int(end - start)
+            minutes = math.floor(secondes / 60)
+            secondes = secondes - (minutes * 60)
+            hours = math.floor(minutes / 60)
+            minutes = minutes - (hours * 60)
+            yield output_filename, gr.update(visible=False), desc + \
+            " The video has been generated in " + \
+            ((str(hours) + " h, ") if hours != 0 else "") + \
+            ((str(minutes) + " min, ") if hours != 0 or minutes != 0 else "") + \
+            str(secondes) + " sec. " + \
+            " Video complete. 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.", '', gr.update(interactive=True), gr.update(interactive=False)
             break
 
 def end_process():
@@ -985,6 +1014,7 @@ with block:
                     timed_prompt.change(fn=handle_timed_prompt_change, inputs=[timed_prompt_id, timed_prompt], outputs=[final_prompt])
 
             final_prompt = gr.Textbox(label="Final prompt", value='', info='Use ; to separate in time')
+            prompt_hint = gr.HTML("Video extension barely follows the prompt; to force to follow the prompt, you have to set the Distilled CFG Scale to 3.0 and the Context Frames to 2 but the video quality will be poor.")
             total_second_length = gr.Slider(label="Video Length to Generate (seconds)", minimum=1, maximum=120, value=2, step=0.1)
 
             with gr.Row():
@@ -994,24 +1024,24 @@ with block:
 
             with gr.Accordion("Advanced settings", open=False):
                 enable_preview = gr.Checkbox(label='Enable preview', value=True, info='Display a preview around each second generated but it costs 2 sec. for each second generated.')
-                use_teacache = gr.Checkbox(label='Use TeaCache', value=False, info='Faster speed, but often makes hands and fingers slightly worse.')
+                use_teacache = gr.Checkbox(label='Use TeaCache', value=False, info='Faster speed and no break in brightness, but often makes hands and fingers slightly worse.')
 
-                n_prompt = gr.Textbox(label="Negative Prompt", value="Missing arm, unrealistic position, impossible contortion, blurred, blurry", info='Requires using normal CFG (undistilled) instead of Distilled (set Distilled=1 and CFG > 1).')
+                n_prompt = gr.Textbox(label="Negative Prompt", value="Missing arm, unrealistic position, impossible contortion, visible bone, muscle contraction, blurred, blurry", info='Requires using normal CFG (undistilled) instead of Distilled (set Distilled=1 and CFG > 1).')
 
                 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=25, step=1, info='Increase for more quality, especially if using high non-distilled CFG. Changing this value is not recommended.')
+                steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=25, 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.')
 
                 with gr.Row():
                     no_resize = gr.Checkbox(label='Force Original Video Resolution (no Resizing)', value=False, info='Might run out of VRAM (720p requires > 24GB VRAM).')
                     resolution = gr.Dropdown([
-                        640,
-                        672,
-                        704,
-                        768,
-                        832,
-                        864,
-                        960
-                    ], value=640, label="Resolution (max width or height)")
+                        ["409,600 px (working)", 640],
+                        ["451,584 px (working)", 672],
+                        ["495,616 px (VRAM pb on HF)", 704],
+                        ["589,824 px (not tested)", 768],
+                        ["692,224 px (not tested)", 832],
+                        ["746,496 px (not tested)", 864],
+                        ["921,600 px (not tested)", 960]
+                    ], value=672, label="Resolution (width x height)", info="Do not affect the generation time")
 
                 # 20250506 pftq: Reduced default distilled guidance scale to improve adherence to input video
                 cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=1.0, step=0.01, info='Use this instead of Distilled for more detail/control + Negative Prompt (make sure Distilled set to 1). Doubles render time. Should not change.')
@@ -1049,157 +1079,74 @@ with block:
     ips = [input_image, final_prompt, generation_mode, n_prompt, randomize_seed, seed, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, enable_preview, use_teacache, mp4_crf]
     ips_video = [input_video, final_prompt, n_prompt, randomize_seed, 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 save_preferences(preferences, value):
-        preferences["generation-mode"] = value
-        return preferences
-    
-    def load_preferences(saved_prefs):
-        saved_prefs = init_preferences(saved_prefs)
-        return saved_prefs["generation-mode"]
-    
-    def init_preferences(saved_prefs):
-        if saved_prefs is None:
-            saved_prefs = default_local_storage
-        return saved_prefs
-    
-    def check_parameters(generation_mode, input_image, input_video):
-        if generation_mode == "image" and input_image is None:
-            raise gr.Error("Please provide an image to extend.")
-        if generation_mode == "video" and input_video is None:
-            raise gr.Error("Please provide a video to extend.")
-        return gr.update(interactive=True)
-
-    prompt_number.change(fn=handle_prompt_number_change, inputs=[], outputs=[])
-    timeless_prompt.change(fn=handle_timeless_prompt_change, inputs=[timeless_prompt], outputs=[final_prompt])
-    start_button.click(fn = check_parameters, inputs = [
-        generation_mode, input_image, input_video
-    ], outputs = [end_button], queue = False, show_progress = False).success(fn=process, inputs=ips, outputs=[result_video, preview_image, progress_desc, progress_bar, start_button, end_button])
-    start_button_video.click(fn = check_parameters, inputs = [
-        generation_mode, input_image, input_video
-    ], outputs = [end_button], queue = False, show_progress = False).success(fn=process_video, inputs=ips_video, outputs=[result_video, preview_image, progress_desc, progress_bar, start_button_video, end_button])
-    end_button.click(fn=end_process)
-
-    generation_mode.change(fn = save_preferences, inputs = [
-        local_storage,
-        generation_mode,
-    ], outputs = [
-       local_storage
-    ])
-
-    with gr.Row(elem_id="image_examples", visible=False):
-        gr.Examples(
+    gr.Examples(
+        label = "Examples from image",
         examples = [
                 [
                     "./img_examples/Example1.png", # input_image
                     "A dolphin emerges from the water, photorealistic, realistic, intricate details, 8k, insanely detailed",
                     "image", # generation_mode
-                    "Missing arm, unrealistic position, impossible contortion, blurred, blurry", # n_prompt
+                    "Missing arm, unrealistic position, impossible contortion, visible bone, muscle contraction, blurred, blurry", # n_prompt
                     True, # randomize_seed
                     42, # seed
                     672, # resolution
                     1, # total_second_length
                     9, # latent_window_size
-                    50, # steps
+                    25, # steps
                     1.0, # cfg
                     10.0, # gs
                     0.0, # rs
                     6, # gpu_memory_preservation
                     False, # enable_preview
-                    False, # use_teacache
+                    True, # use_teacache
                     16 # mp4_crf
                 ],
                 [
-                    "./img_examples/Example1.png", # input_image
-                    "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",
+                    "./img_examples/Example2.webp", # input_image
+                    "A black man on the left and an Asian 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 black man on the left and an Asian 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, unrealistic position, impossible contortion, blurred, blurry", # n_prompt
+                    "Missing arm, unrealistic position, impossible contortion, visible bone, muscle contraction, blurred, blurry", # n_prompt
                     True, # randomize_seed
                     42, # seed
                     672, # resolution
-                    1, # total_second_length
+                    2, # total_second_length
                     9, # latent_window_size
-                    35, # steps
+                    25, # steps
                     1.0, # cfg
                     10.0, # gs
                     0.0, # rs
                     6, # gpu_memory_preservation
                     False, # enable_preview
-                    False, # use_teacache
+                    True, # use_teacache
                     16 # mp4_crf
                 ],
-            ],
-        run_on_click = True,
-        fn = process,
-	    inputs = ips,
-        outputs = [result_video, preview_image, progress_desc, progress_bar, start_button, end_button],
-        cache_examples = torch.cuda.device_count() > 0,
-    )
-
-    with gr.Row(elem_id="video_examples", visible=False):
-        gr.Examples(
-        examples = [
                 [
-                    "./img_examples/Example1.mp4", # input_video
-                    "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, unrealistic position, blurred, blurry", # n_prompt
+                    "./img_examples/Example2.webp", # input_image
+                    "A black man on the left and an Asian 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 black man on the left and an Asian 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, unrealistic position, impossible contortion, visible bone, muscle contraction, blurred, blurry", # n_prompt
                     True, # randomize_seed
                     42, # seed
-                    1, # batch
                     672, # resolution
-                    1, # total_second_length
+                    2, # total_second_length
                     9, # latent_window_size
-                    50, # steps
+                    25, # steps
                     1.0, # cfg
                     10.0, # gs
                     0.0, # rs
                     6, # gpu_memory_preservation
                     False, # enable_preview
-                    False, # use_teacache
-                    False, # no_resize
-                    16, # mp4_crf
-                    5, # num_clean_frames
-                    default_vae
+                    True, # use_teacache
+                    16 # mp4_crf
                 ],
                 [
-                    "./img_examples/Example1.mp4", # input_video
-                    "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, unrealistic position, blurred, blurry", # n_prompt
-                    True, # randomize_seed
-                    42, # seed
-                    1, # batch
-                    640, # resolution
-                    1, # total_second_length
-                    9, # latent_window_size
-                    35, # steps
-                    1.0, # cfg
-                    10.0, # gs
-                    0.0, # rs
-                    6, # gpu_memory_preservation
-                    False, # enable_preview
-                    False, # use_teacache
-                    False, # no_resize
-                    16, # mp4_crf
-                    5, # num_clean_frames
-                    default_vae
-                ],
-            ],
-        run_on_click = True,
-        fn = process_video,
-	    inputs = ips_video,
-	    outputs = [result_video, preview_image, progress_desc, progress_bar, start_button_video, end_button],
-        cache_examples = torch.cuda.device_count() > 0,
-    )
-
-    gr.Examples(
-        examples = [
-                [
-                    "./img_examples/Example1.png", # input_image
-                    "A dolphin emerges from the water, photorealistic, realistic, intricate details, 8k, insanely detailed",
+                    "./img_examples/Example3.jpg", # input_image
+                    "A boy is walking to the right, full view, full-length view, cartoon",
                     "image", # generation_mode
-                    "Missing arm, unrealistic position, impossible contortion, blurred, blurry", # n_prompt
+                    "Missing arm, unrealistic position, impossible contortion, visible bone, muscle contraction, blurred, blurry", # n_prompt
                     True, # randomize_seed
                     42, # seed
-                    640, # resolution
+                    672, # resolution
                     1, # total_second_length
                     9, # latent_window_size
                     25, # steps
@@ -1208,7 +1155,7 @@ with block:
                     0.0, # rs
                     6, # gpu_memory_preservation
                     False, # enable_preview
-                    False, # use_teacache
+                    True, # use_teacache
                     16 # mp4_crf
                 ]
             ],
@@ -1220,15 +1167,16 @@ with block:
     )
 
     gr.Examples(
+        label = "Examples from video",
         examples = [
                 [
                     "./img_examples/Example1.mp4", # input_video
                     "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, unrealistic position, blurred, blurry", # n_prompt
+                    "Missing arm, unrealistic position, impossible contortion, visible bone, muscle contraction, blurred, blurry", # n_prompt
                     True, # randomize_seed
                     42, # seed
                     1, # batch
-                    640, # resolution
+                    672, # resolution
                     1, # total_second_length
                     9, # latent_window_size
                     25, # steps
@@ -1237,7 +1185,7 @@ with block:
                     0.0, # rs
                     6, # gpu_memory_preservation
                     False, # enable_preview
-                    False, # use_teacache
+                    True, # use_teacache
                     False, # no_resize
                     16, # mp4_crf
                     5, # num_clean_frames
@@ -1250,20 +1198,57 @@ with block:
 	    outputs = [result_video, preview_image, progress_desc, progress_bar, start_button_video, end_button],
         cache_examples = False,
     )
+
+    def save_preferences(preferences, value):
+        preferences["generation-mode"] = value
+        return preferences
+    
+    def load_preferences(saved_prefs):
+        saved_prefs = init_preferences(saved_prefs)
+        return saved_prefs["generation-mode"]
+    
+    def init_preferences(saved_prefs):
+        if saved_prefs is None:
+            saved_prefs = default_local_storage
+        return saved_prefs
+    
+    def check_parameters(generation_mode, input_image, input_video):
+        if generation_mode == "image" and input_image is None:
+            raise gr.Error("Please provide an image to extend.")
+        if generation_mode == "video" and input_video is None:
+            raise gr.Error("Please provide a video to extend.")
+        return gr.update(interactive=True)
     
     def handle_generation_mode_change(generation_mode_data):
         if generation_mode_data == "text":
-            return [gr.update(visible = True), gr.update(visible = False), gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False)]
+            return [gr.update(visible = True), gr.update(visible = False), gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False)]
         elif generation_mode_data == "image":
-            return [gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False)]
+            return [gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False), gr.update(visible = False)]
         elif generation_mode_data == "video":
-            return [gr.update(visible = False), gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True)]
+            return [gr.update(visible = False), gr.update(visible = False), gr.update(visible = True), gr.update(visible = False), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True)]
 
     
+    prompt_number.change(fn=handle_prompt_number_change, inputs=[], outputs=[])
+    timeless_prompt.change(fn=handle_timeless_prompt_change, inputs=[timeless_prompt], outputs=[final_prompt])
+    start_button.click(fn = check_parameters, inputs = [
+        generation_mode, input_image, input_video
+    ], outputs = [end_button], queue = False, show_progress = False).success(fn=process, inputs=ips, outputs=[result_video, preview_image, progress_desc, progress_bar, start_button, end_button])
+    start_button_video.click(fn = check_parameters, inputs = [
+        generation_mode, input_image, input_video
+    ], outputs = [end_button], queue = False, show_progress = False).success(fn=process_video, inputs=ips_video, outputs=[result_video, preview_image, progress_desc, progress_bar, start_button_video, end_button])
+    end_button.click(fn=end_process)
+
+    generation_mode.change(fn = save_preferences, inputs = [
+        local_storage,
+        generation_mode,
+    ], outputs = [
+       local_storage
+    ])
+
     generation_mode.change(
         fn=handle_generation_mode_change,
         inputs=[generation_mode],
-        outputs=[text_to_video_hint, input_image, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch]
+        outputs=[text_to_video_hint, input_image, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch, prompt_hint]
     )
 
     # Update display when the page loads
@@ -1271,7 +1256,7 @@ with block:
         fn=handle_generation_mode_change, inputs = [
         generation_mode
     ], outputs = [
-       text_to_video_hint, input_image, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch
+       text_to_video_hint, input_image, input_video, start_button, start_button_video, no_resize, batch, num_clean_frames, vae_batch, prompt_hint
     ]
     )