Merge code

app.py

@@ -574,6 +574,333 @@ def process(input_image, prompt,
             yield output_filename, gr.update(visible=False), gr.update(), '', gr.update(interactive=True), gr.update(interactive=False)
             break
 
+# 20250506 pftq: Modified worker to accept video input and clean frame count
+@spaces.GPU()
+@torch.no_grad()
+def worker_video(input_video, prompt, n_prompt, seed, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
+
+    stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Starting ...'))))
+
+    try:
+        # Clean GPU
+        if not high_vram:
+            unload_complete_models(
+                text_encoder, text_encoder_2, image_encoder, vae, transformer
+            )
+
+        # Text encoding
+        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Text encoding ...'))))
+
+        if not high_vram:
+            fake_diffusers_current_device(text_encoder, gpu)  # since we only encode one text - that is one model move and one encode, offload is same time consumption since it is also one load and one encode.
+            load_model_as_complete(text_encoder_2, target_device=gpu)
+
+        llama_vec, clip_l_pooler = encode_prompt_conds(prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
+
+        if cfg == 1:
+            llama_vec_n, clip_l_pooler_n = torch.zeros_like(llama_vec), torch.zeros_like(clip_l_pooler)
+        else:
+            llama_vec_n, clip_l_pooler_n = encode_prompt_conds(n_prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
+
+        llama_vec, llama_attention_mask = crop_or_pad_yield_mask(llama_vec, length=512)
+        llama_vec_n, llama_attention_mask_n = crop_or_pad_yield_mask(llama_vec_n, length=512)
+
+        # 20250506 pftq: Processing input video instead of image
+        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Video processing ...'))))
+
+        # 20250506 pftq: Encode video
+        #H, W = 640, 640  # Default resolution, will be adjusted
+        #height, width = find_nearest_bucket(H, W, resolution=640)
+        #start_latent, input_image_np, history_latents, fps = video_encode(input_video, vae, height, width, vae_batch_size=16, device=gpu)
+        start_latent, input_image_np, video_latents, fps, height, width, input_video_pixels  = video_encode(input_video, resolution, no_resize, vae, vae_batch_size=vae_batch, device=gpu)
+
+        #Image.fromarray(input_image_np).save(os.path.join(outputs_folder, f'{job_id}.png'))
+
+        # CLIP Vision
+        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'CLIP Vision encoding ...'))))
+
+        if not high_vram:
+            load_model_as_complete(image_encoder, target_device=gpu)
+
+        image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
+        image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
+
+        # Dtype
+        llama_vec = llama_vec.to(transformer.dtype)
+        llama_vec_n = llama_vec_n.to(transformer.dtype)
+        clip_l_pooler = clip_l_pooler.to(transformer.dtype)
+        clip_l_pooler_n = clip_l_pooler_n.to(transformer.dtype)
+        image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
+
+        total_latent_sections = (total_second_length * fps) / (latent_window_size * 4)
+        total_latent_sections = int(max(round(total_latent_sections), 1))
+
+        for idx in range(batch):
+            if batch > 1:
+                print(f"Beginning video {idx+1} of {batch} with seed {seed} ")
+
+            #job_id = generate_timestamp()
+            job_id = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")+f"_framepackf1-videoinput_{width}-{total_second_length}sec_seed-{seed}_steps-{steps}_distilled-{gs}_cfg-{cfg}" # 20250506 pftq: easier to read timestamp and filename
+
+            # Sampling
+            stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...'))))
+
+            rnd = torch.Generator("cpu").manual_seed(seed)
+
+            # 20250506 pftq: Initialize history_latents with video latents
+            history_latents = video_latents.cpu()
+            total_generated_latent_frames = history_latents.shape[2]
+            # 20250506 pftq: Initialize history_pixels to fix UnboundLocalError
+            history_pixels = None
+            previous_video = None
+
+            # 20250507 pftq: hot fix for initial video being corrupted by vae encoding, issue with ghosting because of slight differences
+            #history_pixels = input_video_pixels
+            #save_bcthw_as_mp4(vae_decode(video_latents, vae).cpu(), os.path.join(outputs_folder, f'{job_id}_input_video.mp4'), fps=fps, crf=mp4_crf) # 20250507 pftq: test fast movement corrupted by vae encoding if vae batch size too low
+
+            for section_index in range(total_latent_sections):
+                if stream.input_queue.top() == 'end':
+                    stream.output_queue.push(('end', None))
+                    return
+
+                print(f'section_index = {section_index}, total_latent_sections = {total_latent_sections}')
+
+                if not high_vram:
+                    unload_complete_models()
+                    move_model_to_device_with_memory_preservation(transformer, target_device=gpu, preserved_memory_gb=gpu_memory_preservation)
+
+                if use_teacache:
+                    transformer.initialize_teacache(enable_teacache=True, num_steps=steps)
+                else:
+                    transformer.initialize_teacache(enable_teacache=False)
+
+                def callback(d):
+                    preview = d['denoised']
+                    preview = vae_decode_fake(preview)
+
+                    preview = (preview * 255.0).detach().cpu().numpy().clip(0, 255).astype(np.uint8)
+                    preview = einops.rearrange(preview, 'b c t h w -> (b h) (t w) c')
+
+                    if stream.input_queue.top() == 'end':
+                        stream.output_queue.push(('end', None))
+                        raise KeyboardInterrupt('User ends the task.')
+
+                    current_step = d['i'] + 1
+                    percentage = int(100.0 * current_step / steps)
+                    hint = f'Sampling {current_step}/{steps}'
+                    desc = f'Total frames: {int(max(0, total_generated_latent_frames * 4 - 3))}, Video length: {max(0, (total_generated_latent_frames * 4 - 3) / fps) :.2f} seconds (FPS-{fps}), Seed: {seed}, Video {idx+1} of {batch}. The video is generating part {section_index+1} of {total_latent_sections}...'
+                    stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint))))
+                    return
+
+                # 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 = history_latents[:, :, -total_context_frames:, :, :] if total_context_frames > 0 else history_latents[:, :, :fallback_frame_count, :, :]
+                if total_context_frames > 0:
+                    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
+                        clean_latents_4x = splits[split_idx] if num_4x_frames > 0 else history_latents[:, :, :fallback_frame_count, :, :]
+                        if clean_latents_4x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
+                            clean_latents_4x = torch.cat([clean_latents_4x, clean_latents_4x[:, :, -1:, :, :]], dim=2)[:, :, :2, :, :]
+                        split_idx += 1 if num_4x_frames > 0 else 0
+                        clean_latents_2x = splits[split_idx] if num_2x_frames > 0 and split_idx < len(splits) else history_latents[:, :, :fallback_frame_count, :, :]
+                        if clean_latents_2x.shape[2] < 2:  # 20250507 pftq: edge case for <=1 sec videos
+                            clean_latents_2x = torch.cat([clean_latents_2x, clean_latents_2x[:, :, -1:, :, :]], dim=2)[:, :, :2, :, :]
+                        split_idx += 1 if num_2x_frames > 0 else 0
+                        clean_latents_1x = splits[split_idx] if effective_clean_frames > 0 and split_idx < len(splits) else history_latents[:, :, :fallback_frame_count, :, :]
+                    else:
+                        clean_latents_4x = clean_latents_2x = clean_latents_1x = history_latents[:, :, :fallback_frame_count, :, :]
+                else:
+                    clean_latents_4x = clean_latents_2x = clean_latents_1x = history_latents[:, :, :fallback_frame_count, :, :]
+
+                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)
+
+                generated_latents = sample_hunyuan(
+                    transformer=transformer,
+                    sampler='unipc',
+                    width=width,
+                    height=height,
+                    frames=max_frames,
+                    real_guidance_scale=cfg,
+                    distilled_guidance_scale=gs,
+                    guidance_rescale=rs,
+                    num_inference_steps=steps,
+                    generator=rnd,
+                    prompt_embeds=llama_vec,
+                    prompt_embeds_mask=llama_attention_mask,
+                    prompt_poolers=clip_l_pooler,
+                    negative_prompt_embeds=llama_vec_n,
+                    negative_prompt_embeds_mask=llama_attention_mask_n,
+                    negative_prompt_poolers=clip_l_pooler_n,
+                    device=gpu,
+                    dtype=torch.bfloat16,
+                    image_embeddings=image_encoder_last_hidden_state,
+                    latent_indices=latent_indices,
+                    clean_latents=clean_latents,
+                    clean_latent_indices=clean_latent_indices,
+                    clean_latents_2x=clean_latents_2x,
+                    clean_latent_2x_indices=clean_latent_2x_indices,
+                    clean_latents_4x=clean_latents_4x,
+                    clean_latent_4x_indices=clean_latent_4x_indices,
+                    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 = min(latent_window_size * 4 - 3, history_pixels.shape[2])
+
+                  #if section_index == 0:
+                    #extra_latents = 1  # Add up to 2 extra latent frames for smoother overlap to initial video
+                    #extra_pixel_frames = extra_latents * 4  # Approx. 4 pixel frames per latent
+                    #overlapped_frames = min(overlapped_frames + extra_pixel_frames, history_pixels.shape[2], section_latent_frames * 4)
+
+                  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()
+
+                output_filename = os.path.join(outputs_folder, f'{job_id}_{total_generated_latent_frames}.mp4')
+
+                # 20250506 pftq: Use input video FPS for output
+                save_bcthw_as_mp4(history_pixels, output_filename, fps=fps, crf=mp4_crf)
+                print(f"Latest video saved: {output_filename}")
+                # 20250508 pftq: Save prompt to mp4 metadata comments
+                set_mp4_comments_imageio_ffmpeg(output_filename, f"Prompt: {prompt} | Negative Prompt: {n_prompt}");
+                print(f"Prompt saved to mp4 metadata comments: {output_filename}")
+
+                # 20250506 pftq: Clean up previous partial files
+                if previous_video is not None and os.path.exists(previous_video):
+                    try:
+                        os.remove(previous_video)
+                        print(f"Previous partial video deleted: {previous_video}")
+                    except Exception as e:
+                        print(f"Error deleting previous partial video {previous_video}: {e}")
+                previous_video = output_filename
+
+                print(f'Decoded. Current latent shape {real_history_latents.shape}; pixel shape {history_pixels.shape}')
+
+                stream.output_queue.push(('file', output_filename))
+
+            seed = (seed + 1) % np.iinfo(np.int32).max
+
+    except:
+        traceback.print_exc()
+
+        if not high_vram:
+            unload_complete_models(
+                text_encoder, text_encoder_2, image_encoder, vae, transformer
+            )
+
+    stream.output_queue.push(('end', None))
+    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, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
+    global total_second_length_debug_value
+    if total_second_length_debug_value is not None:
+        return min(total_second_length_debug_value * 60 * 10, 600)
+    return total_second_length * 60 * 10
+
+# 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, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch):
+    global stream, high_vram, input_video_debug_value, prompt_debug_value, total_second_length_debug_value
+
+    if torch.cuda.device_count() == 0:
+        gr.Warning('Set this space to GPU config to make it work.')
+        return None, None, None, None, None, None
+
+    if input_video_debug_value is not None:
+        input_video = input_video_debug_value
+        input_video_debug_value = None
+
+    if prompt_debug_value is not None:
+        prompt = prompt_debug_value
+        prompt_debug_value = None
+
+    if total_second_length_debug_value is not None:
+        total_second_length = total_second_length_debug_value
+        total_second_length_debug_value = None
+
+    if randomize_seed:
+        seed = random.randint(0, np.iinfo(np.int32).max)
+
+    # 20250506 pftq: Updated assertion for video input
+    assert input_video is not None, 'No input video!'
+
+    yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
+
+    # 20250507 pftq: Even the H100 needs offloading if the video dimensions are 720p or higher
+    if high_vram and (no_resize or resolution>640):
+        print("Disabling high vram mode due to no resize and/or potentially higher resolution...")
+        high_vram = False
+        vae.enable_slicing()
+        vae.enable_tiling()
+        DynamicSwapInstaller.install_model(transformer, device=gpu)
+        DynamicSwapInstaller.install_model(text_encoder, device=gpu)
+
+    # 20250508 pftq: automatically set distilled cfg to 1 if cfg is used
+    if cfg > 1:
+        gs = 1
+
+    stream = AsyncStream()
+
+    # 20250506 pftq: Pass num_clean_frames, vae_batch, etc
+    async_run(worker_video, input_video, prompt, n_prompt, seed, batch, resolution, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, no_resize, mp4_crf, num_clean_frames, vae_batch)
+
+    output_filename = None
+
+    while True:
+        flag, data = stream.output_queue.next()
+
+        if flag == 'file':
+            output_filename = data
+            yield output_filename, gr.update(), gr.update(), gr.update(), gr.update(interactive=False), gr.update(interactive=True)
+
+        if flag == 'progress':
+            preview, desc, html = data
+            #yield gr.update(), gr.update(visible=True, value=preview), desc, html, gr.update(interactive=False), gr.update(interactive=True)
+            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.', '', gr.update(interactive=True), gr.update(interactive=False)
+            break
+
 
 def end_process():
     stream.input_queue.push('end')