block.launch(mcp_server=False, ssr_mode=False)

app.py

@@ -1,755 +1,753 @@
-from diffusers_helper.hf_login import login
-
-import os
-
-os.environ['HF_HOME'] = os.path.abspath(os.path.realpath(os.path.join(os.path.dirname(__file__), './hf_download')))
-
-import spaces
-import gradio as gr
-import torch
-import traceback
-import einops
-import safetensors.torch as sf
-import numpy as np
-import argparse
-import random
-import math
-# 20250506 pftq: Added for video input loading
-import decord
-# 20250506 pftq: Added for progress bars in video_encode
-from tqdm import tqdm
-# 20250506 pftq: Normalize file paths for Windows compatibility
-import pathlib
-# 20250506 pftq: for easier to read timestamp
-from datetime import datetime
-# 20250508 pftq: for saving prompt to mp4 comments metadata
-import imageio_ffmpeg
-import tempfile
-import shutil
-import subprocess
-
-from PIL import Image
-from diffusers import AutoencoderKLHunyuanVideo
-from transformers import LlamaModel, CLIPTextModel, LlamaTokenizerFast, CLIPTokenizer
-from diffusers_helper.hunyuan import encode_prompt_conds, vae_decode, vae_encode, vae_decode_fake
-from diffusers_helper.utils import save_bcthw_as_mp4, crop_or_pad_yield_mask, soft_append_bcthw, resize_and_center_crop, state_dict_weighted_merge, state_dict_offset_merge, generate_timestamp
-from diffusers_helper.models.hunyuan_video_packed import HunyuanVideoTransformer3DModelPacked
-from diffusers_helper.pipelines.k_diffusion_hunyuan import sample_hunyuan
-from diffusers_helper.memory import cpu, gpu, get_cuda_free_memory_gb, move_model_to_device_with_memory_preservation, offload_model_from_device_for_memory_preservation, fake_diffusers_current_device, DynamicSwapInstaller, unload_complete_models, load_model_as_complete
-from diffusers_helper.thread_utils import AsyncStream, async_run
-from diffusers_helper.gradio.progress_bar import make_progress_bar_css, make_progress_bar_html
-from transformers import SiglipImageProcessor, SiglipVisionModel
-from diffusers_helper.clip_vision import hf_clip_vision_encode
-from diffusers_helper.bucket_tools import find_nearest_bucket
-from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, HunyuanVideoTransformer3DModel, HunyuanVideoPipeline
-
-
-if torch.cuda.device_count() > 0:
-    free_mem_gb = get_cuda_free_memory_gb(gpu)
-    high_vram = free_mem_gb > 80
-    
-    print(f'Free VRAM {free_mem_gb} GB')
-    print(f'High-VRAM Mode: {high_vram}')
-    
-    text_encoder = LlamaModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder', torch_dtype=torch.float16).cpu()
-    text_encoder_2 = CLIPTextModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder_2', torch_dtype=torch.float16).cpu()
-    tokenizer = LlamaTokenizerFast.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer')
-    tokenizer_2 = CLIPTokenizer.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer_2')
-    vae = AutoencoderKLHunyuanVideo.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='vae', torch_dtype=torch.float16).cpu()
-    
-    feature_extractor = SiglipImageProcessor.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='feature_extractor')
-    image_encoder = SiglipVisionModel.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='image_encoder', torch_dtype=torch.float16).cpu()
-    
-    transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained('lllyasviel/FramePack_F1_I2V_HY_20250503', torch_dtype=torch.bfloat16).cpu()
-    
-    vae.eval()
-    text_encoder.eval()
-    text_encoder_2.eval()
-    image_encoder.eval()
-    transformer.eval()
-    
-    if not high_vram:
-        vae.enable_slicing()
-        vae.enable_tiling()
-    
-    transformer.high_quality_fp32_output_for_inference = True
-    print('transformer.high_quality_fp32_output_for_inference = True')
-    
-    transformer.to(dtype=torch.bfloat16)
-    vae.to(dtype=torch.float16)
-    image_encoder.to(dtype=torch.float16)
-    text_encoder.to(dtype=torch.float16)
-    text_encoder_2.to(dtype=torch.float16)
-    
-    vae.requires_grad_(False)
-    text_encoder.requires_grad_(False)
-    text_encoder_2.requires_grad_(False)
-    image_encoder.requires_grad_(False)
-    transformer.requires_grad_(False)
-    
-    if not high_vram:
-        # DynamicSwapInstaller is same as huggingface's enable_sequential_offload but 3x faster
-        DynamicSwapInstaller.install_model(transformer, device=gpu)
-        DynamicSwapInstaller.install_model(text_encoder, device=gpu)
-    else:
-        text_encoder.to(gpu)
-        text_encoder_2.to(gpu)
-        image_encoder.to(gpu)
-        vae.to(gpu)
-        transformer.to(gpu)
-
-stream = AsyncStream()
-
-outputs_folder = './outputs/'
-os.makedirs(outputs_folder, exist_ok=True)
-
-input_image_debug_value = prompt_debug_value = total_second_length_debug_value = None
-
-@spaces.GPU()
-@torch.no_grad()
-def video_encode(video_path, resolution, no_resize, vae, vae_batch_size=16, device="cuda", width=None, height=None):
-    """
-    Encode a video into latent representations using the VAE.
-
-    Args:
-        video_path: Path to the input video file.
-        vae: AutoencoderKLHunyuanVideo model.
-        height, width: Target resolution for resizing frames.
-        vae_batch_size: Number of frames to process per batch.
-        device: Device for computation (e.g., "cuda").
-
-    Returns:
-        start_latent: Latent of the first frame (for compatibility with original code).
-        input_image_np: First frame as numpy array (for CLIP vision encoding).
-        history_latents: Latents of all frames (shape: [1, channels, frames, height//8, width//8]).
-        fps: Frames per second of the input video.
-    """
-    # 20250506 pftq: Normalize video path for Windows compatibility
-    video_path = str(pathlib.Path(video_path).resolve())
-    print(f"Processing video: {video_path}")
-
-    # 20250506 pftq: Check CUDA availability and fallback to CPU if needed
-    if device == "cuda" and not torch.cuda.is_available():
-        print("CUDA is not available, falling back to CPU")
-        device = "cpu"
-
-    try:
-        # 20250506 pftq: Load video and get FPS
-        print("Initializing VideoReader...")
-        vr = decord.VideoReader(video_path)
-        fps = vr.get_avg_fps()  # Get input video FPS
-        num_real_frames = len(vr)
-        print(f"Video loaded: {num_real_frames} frames, FPS: {fps}")
-
-        # Truncate to nearest latent size (multiple of 4)
-        latent_size_factor = 4
-        num_frames = (num_real_frames // latent_size_factor) * latent_size_factor
-        if num_frames != num_real_frames:
-            print(f"Truncating video from {num_real_frames} to {num_frames} frames for latent size compatibility")
-        num_real_frames = num_frames
-
-        # 20250506 pftq: Read frames
-        print("Reading video frames...")
-        frames = vr.get_batch(range(num_real_frames)).asnumpy()  # Shape: (num_real_frames, height, width, channels)
-        print(f"Frames read: {frames.shape}")
-
-        # 20250506 pftq: Get native video resolution
-        native_height, native_width = frames.shape[1], frames.shape[2]
-        print(f"Native video resolution: {native_width}x{native_height}")
-
-        # 20250506 pftq: Use native resolution if height/width not specified, otherwise use provided values
-        target_height = native_height if height is None else height
-        target_width = native_width if width is None else width
-
-        # 20250506 pftq: Adjust to nearest bucket for model compatibility
-        if not no_resize:
-            target_height, target_width = find_nearest_bucket(target_height, target_width, resolution=resolution)
-            print(f"Adjusted resolution: {target_width}x{target_height}")
-        else:
-            print(f"Using native resolution without resizing: {target_width}x{target_height}")
-
-        # 20250506 pftq: Preprocess frames to match original image processing
-        processed_frames = []
-        for i, frame in enumerate(frames):
-            #print(f"Preprocessing frame {i+1}/{num_frames}")
-            frame_np = resize_and_center_crop(frame, target_width=target_width, target_height=target_height)
-            processed_frames.append(frame_np)
-        processed_frames = np.stack(processed_frames)  # Shape: (num_real_frames, height, width, channels)
-        print(f"Frames preprocessed: {processed_frames.shape}")
-
-        # 20250506 pftq: Save first frame for CLIP vision encoding
-        input_image_np = processed_frames[0]
-
-        # 20250506 pftq: Convert to tensor and normalize to [-1, 1]
-        print("Converting frames to tensor...")
-        frames_pt = torch.from_numpy(processed_frames).float() / 127.5 - 1
-        frames_pt = frames_pt.permute(0, 3, 1, 2)  # Shape: (num_real_frames, channels, height, width)
-        frames_pt = frames_pt.unsqueeze(0)  # Shape: (1, num_real_frames, channels, height, width)
-        frames_pt = frames_pt.permute(0, 2, 1, 3, 4)  # Shape: (1, channels, num_real_frames, height, width)
-        print(f"Tensor shape: {frames_pt.shape}")
-
-        # 20250507 pftq: Save pixel frames for use in worker
-        input_video_pixels = frames_pt.cpu()
-
-        # 20250506 pftq: Move to device
-        print(f"Moving tensor to device: {device}")
-        frames_pt = frames_pt.to(device)
-        print("Tensor moved to device")
-
-        # 20250506 pftq: Move VAE to device
-        print(f"Moving VAE to device: {device}")
-        vae.to(device)
-        print("VAE moved to device")
-
-        # 20250506 pftq: Encode frames in batches
-        print(f"Encoding input video frames in VAE batch size {vae_batch_size} (reduce if memory issues here or if forcing video resolution)")
-        latents = []
-        vae.eval()
-        with torch.no_grad():
-            for i in tqdm(range(0, frames_pt.shape[2], vae_batch_size), desc="Encoding video frames", mininterval=0.1):
-                #print(f"Encoding batch {i//vae_batch_size + 1}: frames {i} to {min(i + vae_batch_size, frames_pt.shape[2])}")
-                batch = frames_pt[:, :, i:i + vae_batch_size]  # Shape: (1, channels, batch_size, height, width)
-                try:
-                    # 20250506 pftq: Log GPU memory before encoding
-                    if device == "cuda":
-                        free_mem = torch.cuda.memory_allocated() / 1024**3
-                        #print(f"GPU memory before encoding: {free_mem:.2f} GB")
-                    batch_latent = vae_encode(batch, vae)
-                    # 20250506 pftq: Synchronize CUDA to catch issues
-                    if device == "cuda":
-                        torch.cuda.synchronize()
-                        #print(f"GPU memory after encoding: {torch.cuda.memory_allocated() / 1024**3:.2f} GB")
-                    latents.append(batch_latent)
-                    #print(f"Batch encoded, latent shape: {batch_latent.shape}")
-                except RuntimeError as e:
-                    print(f"Error during VAE encoding: {str(e)}")
-                    if device == "cuda" and "out of memory" in str(e).lower():
-                        print("CUDA out of memory, try reducing vae_batch_size or using CPU")
-                    raise
-
-        # 20250506 pftq: Concatenate latents
-        print("Concatenating latents...")
-        history_latents = torch.cat(latents, dim=2)  # Shape: (1, channels, frames, height//8, width//8)
-        print(f"History latents shape: {history_latents.shape}")
-
-        # 20250506 pftq: Get first frame's latent
-        start_latent = history_latents[:, :, :1]  # Shape: (1, channels, 1, height//8, width//8)
-        print(f"Start latent shape: {start_latent.shape}")
-
-        # 20250506 pftq: Move VAE back to CPU to free GPU memory
-        if device == "cuda":
-            vae.to(cpu)
-            torch.cuda.empty_cache()
-            print("VAE moved back to CPU, CUDA cache cleared")
-
-        return start_latent, input_image_np, history_latents, fps, target_height, target_width, input_video_pixels
-
-    except Exception as e:
-        print(f"Error in video_encode: {str(e)}")
-        raise
-
-# 20250508 pftq: for saving prompt to mp4 metadata comments
-def set_mp4_comments_imageio_ffmpeg(input_file, comments):
-    try:
-        # Get the path to the bundled FFmpeg binary from imageio-ffmpeg
-        ffmpeg_path = imageio_ffmpeg.get_ffmpeg_exe()
-
-        # Check if input file exists
-        if not os.path.exists(input_file):
-            print(f"Error: Input file {input_file} does not exist")
-            return False
-
-        # Create a temporary file path
-        temp_file = tempfile.NamedTemporaryFile(suffix='.mp4', delete=False).name
-
-        # FFmpeg command using the bundled binary
-        command = [
-            ffmpeg_path,                   # Use imageio-ffmpeg's FFmpeg
-            '-i', input_file,              # input file
-            '-metadata', f'comment={comments}',  # set comment metadata
-            '-c:v', 'copy',                # copy video stream without re-encoding
-            '-c:a', 'copy',                # copy audio stream without re-encoding
-            '-y',                          # overwrite output file if it exists
-            temp_file                      # temporary output file
-        ]
-
-        # Run the FFmpeg command
-        result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
-
-        if result.returncode == 0:
-            # Replace the original file with the modified one
-            shutil.move(temp_file, input_file)
-            print(f"Successfully added comments to {input_file}")
-            return True
-        else:
-            # Clean up temp file if FFmpeg fails
-            if os.path.exists(temp_file):
-                os.remove(temp_file)
-            print(f"Error: FFmpeg failed with message:\n{result.stderr}")
-            return False
-
-    except Exception as e:
-        # Clean up temp file in case of other errors
-        if 'temp_file' in locals() and os.path.exists(temp_file):
-            os.remove(temp_file)
-        print(f"Error saving prompt to video metadata, ffmpeg may be required: "+str(e))
-        return False
-
-@torch.no_grad()
-def worker(input_image, prompts, n_prompt, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf):
-    def encode_prompt(prompt, n_prompt):
-        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)
-
-        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)
-        return [llama_vec, clip_l_pooler, llama_vec_n, clip_l_pooler_n, llama_attention_mask, llama_attention_mask_n]
-
-    total_latent_sections = (total_second_length * 30) / (latent_window_size * 4)
-    total_latent_sections = int(max(round(total_latent_sections), 1))
-
-    job_id = generate_timestamp()
-
-    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)
-
-        prompt_parameters = []
-        
-        for prompt_part in prompts:
-            prompt_parameters.append(encode_prompt(prompt_part, n_prompt))
-
-        # Processing input image
-
-        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Image processing ...'))))
-
-        H, W, C = input_image.shape
-        height, width = find_nearest_bucket(H, W, resolution=640)
-        input_image_np = resize_and_center_crop(input_image, target_width=width, target_height=height)
-
-        Image.fromarray(input_image_np).save(os.path.join(outputs_folder, f'{job_id}.png'))
-
-        input_image_pt = torch.from_numpy(input_image_np).float() / 127.5 - 1
-        input_image_pt = input_image_pt.permute(2, 0, 1)[None, :, None]
-
-        # VAE encoding
-
-        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'VAE encoding ...'))))
-
-        if not high_vram:
-            load_model_as_complete(vae, target_device=gpu)
-
-        start_latent = vae_encode(input_image_pt, vae)
-
-        # 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
-
-        image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
-
-        # Sampling
-
-        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...'))))
-
-        rnd = torch.Generator("cpu").manual_seed(seed)
-
-        history_latents = torch.zeros(size=(1, 16, 16 + 2 + 1, height // 8, width // 8), dtype=torch.float32).cpu()
-        history_pixels = None
-
-        history_latents = torch.cat([history_latents, start_latent.to(history_latents)], dim=2)
-        total_generated_latent_frames = 1
-
-        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 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(0)
-
-            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 generated frames: {int(max(0, total_generated_latent_frames * 4 - 3))}, Video length: {max(0, (total_generated_latent_frames * 4 - 3) / 30) :.2f} seconds (FPS-30). The video is being extended now ...'
-                stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint))))
-                return
-
-            indices = torch.arange(0, sum([1, 16, 2, 1, latent_window_size])).unsqueeze(0)
-            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)
-
-            clean_latents_4x, clean_latents_2x, clean_latents_1x = history_latents[:, :, -sum([16, 2, 1]):, :, :].split([16, 2, 1], dim=2)
-            clean_latents = torch.cat([start_latent.to(history_latents), clean_latents_1x], dim=2)
-
-            generated_latents = sample_hunyuan(
-                transformer=transformer,
-                sampler='unipc',
-                width=width,
-                height=height,
-                frames=latent_window_size * 4 - 3,
-                real_guidance_scale=cfg,
-                distilled_guidance_scale=gs,
-                guidance_rescale=rs,
-                # shift=3.0,
-                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 = 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()
-
-            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))
-    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(input_image, prompt, t2v, n_prompt, randomize_seed, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf):
-    global total_second_length_debug_value
-
-    if total_second_length_debug_value is not None:
-        return min(total_second_length_debug_value * 60, 600)
-    return total_second_length * 60
-
-
-@spaces.GPU(duration=get_duration)
-def process(input_image, prompt, 
-            t2v=False, 
-            n_prompt="",
-            randomize_seed=True,
-            seed=31337, 
-            total_second_length=5, 
-            latent_window_size=9, 
-            steps=25, 
-            cfg=1.0, 
-            gs=10.0, 
-            rs=0.0, 
-            gpu_memory_preservation=6, 
-            use_teacache=True,
-            mp4_crf=16
-           ):
-    global stream, input_image_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_image_debug_value is not None or prompt_debug_value is not None or total_second_length_debug_value is not None:
-        print("Debug mode")
-        input_image = input_image_debug_value
-        prompt = prompt_debug_value
-        total_second_length = total_second_length_debug_value
-        input_image_debug_value = prompt_debug_value = total_second_length_debug_value = None
-
-    if randomize_seed:
-        seed = random.randint(0, np.iinfo(np.int32).max)
-
-    prompts = prompt.split(";")
-    
-    # assert input_image is not None, 'No input image!'
-    if t2v:
-        default_height, default_width = 640, 640
-        input_image = np.ones((default_height, default_width, 3), dtype=np.uint8) * 255
-        print("No input image provided. Using a blank white image.")
-    
-    yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
-
-    stream = AsyncStream()
-
-    async_run(worker, input_image, prompts, n_prompt, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf)
-
-    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)
-
-        if flag == 'end':
-            yield output_filename, gr.update(visible=False), gr.update(), '', gr.update(interactive=True), gr.update(interactive=False)
-            break
-
-
-def end_process():
-    stream.input_queue.push('end')
-
-
-css = make_progress_bar_css()
-block = gr.Blocks(css=css).queue()
-with block:
-    if torch.cuda.device_count() == 0:
-        with gr.Row():
-            gr.HTML("""
-    <p style="background-color: red;"><big><big><big><b>⚠️To use FramePack, <a href="https://huggingface.co/spaces/Fabrice-TIERCELIN/SUPIR?duplicate=true">duplicate this space</a> and set a GPU with 30 GB VRAM.</b>
-    
-    You can't use FramePack directly here because this space runs on a CPU, which is not enough for FramePack. Please provide <a href="https://huggingface.co/spaces/Fabrice-TIERCELIN/SUPIR/discussions/new">feedback</a> if you have issues.
-    </big></big></big></p>
-    """)
-    # 20250506 pftq: Updated title to reflect video input functionality
-    gr.Markdown('# Framepack F1 with Image Input or with Video Input (Video Extension)')
-    gr.Markdown(f"""### Video diffusion, but feels like image diffusion
-*FramePack F1 - a FramePack model that only predicts future frames from history frames*
-### *beta* FramePack Fill 🖋️- draw a mask over the input image to inpaint the video output
-adapted from the officical code repo [FramePack](https://github.com/lllyasviel/FramePack) by [lllyasviel](lllyasviel/FramePack_F1_I2V_HY_20250503) and [FramePack Studio](https://github.com/colinurbs/FramePack-Studio) 🙌🏻
-    """)
-    with gr.Row():
-        with gr.Column():
-            input_image = gr.Image(sources='upload', type="numpy", label="Image", height=320)
-            prompt = gr.Textbox(label="Prompt", value='')
-            t2v = gr.Checkbox(label="do text-to-video", value=False)
-
-            with gr.Row():
-                start_button = gr.Button(value="Start Generation", variant="primary")
-                end_button = gr.Button(value="End Generation", variant="stop", interactive=False)
-
-            total_second_length = gr.Slider(label="Generated Video Length (Seconds)", minimum=1, maximum=5, value=2, step=0.1)
-            with gr.Accordion("Advanced settings", open=False):
-                use_teacache = gr.Checkbox(label='Use TeaCache', value=True, info='Faster speed, but often makes hands and fingers slightly worse.')
-    
-                n_prompt = gr.Textbox(label="Negative Prompt", value="Missing arm, unrealistic position, blurred, blurry")  # Not used
-                randomize_seed = gr.Checkbox(label='Randomize seed', value=True, info='If checked, the seed is always different')
-                seed = gr.Slider(label="Seed", minimum=0, maximum=np.iinfo(np.int32).max, step=1, randomize=True)
-    
-                
-                latent_window_size = gr.Slider(label="Latent Window Size", minimum=1, maximum=33, value=9, step=1)  # Should not change
-                steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=25, step=1, info='Changing this value is not recommended.')
-    
-                cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=1.0, step=0.01)  # Should not change
-                gs = gr.Slider(label="Distilled CFG Scale", minimum=1.0, maximum=32.0, value=10.0, step=0.01, info='Changing this value is not recommended; 3=blurred motions& & unsharped; 10 focus motion')
-                rs = gr.Slider(label="CFG Re-Scale", minimum=0.0, maximum=1.0, value=0.0, step=0.01)  # Should not change
-    
-                gpu_memory_preservation = gr.Slider(label="GPU Inference Preserved Memory (GB) (larger means slower)", minimum=6, maximum=128, value=6, step=0.1, info="Set this number to a larger value if you encounter OOM. Larger value causes slower speed.")
-    
-                mp4_crf = gr.Slider(label="MP4 Compression", minimum=0, maximum=100, value=16, step=1, info="Lower means better quality. 0 is uncompressed. Change to 16 if you get black outputs. ")
-
-            with gr.Accordion("Debug", open=False):
-                input_image_debug = gr.Image(type="numpy", label="Image Debug", height=320)
-                prompt_debug = gr.Textbox(label="Prompt Debug", value='')
-                total_second_length_debug = gr.Slider(label="Additional Video Length to Generate (Seconds) Debug", minimum=1, maximum=120, value=5, step=0.1)
-
-        with gr.Column():
-            preview_image = gr.Image(label="Next Latents", height=200, visible=False)
-            result_video = gr.Video(label="Finished Frames", autoplay=True, show_share_button=False, height=512, loop=True)
-            progress_desc = gr.Markdown('', elem_classes='no-generating-animation')
-            progress_bar = gr.HTML('', elem_classes='no-generating-animation')
-
-    gr.HTML('<div style="text-align:center; margin-top:20px;">Share your results and find ideas at the <a href="https://x.com/search?q=framepack&f=live" target="_blank">FramePack Twitter (X) thread</a></div>')
-
-    ips = [input_image, prompt, t2v, n_prompt, randomize_seed, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf]
-    start_button.click(fn=process, inputs=ips, outputs=[result_video, preview_image, progress_desc, progress_bar, start_button, end_button])
-    end_button.click(fn=end_process)
-
-    with gr.Row(visible=False):
-        gr.Examples(
-        examples = [
-                [
-                    "./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",
-                    False, # t2v
-                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
-                    True, # randomize_seed
-                    42, # seed
-                    1, # total_second_length
-                    9, # latent_window_size
-                    25, # steps
-                    1.0, # cfg
-                    10.0, # gs
-                    0.0, # rs
-                    6, # gpu_memory_preservation
-                    True, # use_teacache
-                    16 # mp4_crf
-                ],
-                [
-                    "./img_examples/Example1.png", # input_image
-                    "A dolphin emerges from the water, photorealistic, realistic, intricate details, 8k, insanely detailed",
-                    False, # t2v
-                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
-                    True, # randomize_seed
-                    42, # seed
-                    1, # total_second_length
-                    9, # latent_window_size
-                    25, # steps
-                    1.0, # cfg
-                    10.0, # gs
-                    0.0, # rs
-                    6, # gpu_memory_preservation
-                    True, # use_teacache
-                    16 # mp4_crf
-                ],
-                [
-                    "./img_examples/Example1.png", # input_image
-                    "We are sinking, photorealistic, realistic, intricate details, 8k, insanely detailed",
-                    False, # t2v
-                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
-                    True, # randomize_seed
-                    42, # seed
-                    1, # total_second_length
-                    9, # latent_window_size
-                    25, # steps
-                    1.0, # cfg
-                    10.0, # gs
-                    0.0, # rs
-                    6, # gpu_memory_preservation
-                    True, # use_teacache
-                    16 # mp4_crf
-                ],
-                [
-                    "./img_examples/Example1.png", # input_image
-                    "A boat is passing, photorealistic, realistic, intricate details, 8k, insanely detailed",
-                    False, # t2v
-                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
-                    True, # randomize_seed
-                    42, # seed
-                    1, # total_second_length
-                    9, # latent_window_size
-                    25, # steps
-                    1.0, # cfg
-                    10.0, # gs
-                    0.0, # rs
-                    6, # gpu_memory_preservation
-                    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 = True,
-    )
-
-    gr.Markdown('## Guide')
-    gr.Markdown("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.")
-
-    
-    def handle_field_debug_change(input_image_debug_data, prompt_debug_data, total_second_length_debug_data):
-        global input_image_debug_value, prompt_debug_value, total_second_length_debug_value
-        input_image_debug_value = input_image_debug_data
-        prompt_debug_value = prompt_debug_data
-        total_second_length_debug_value = total_second_length_debug_data
-        return []
-    
-    input_image_debug.upload(
-        fn=handle_field_debug_change,
-        inputs=[input_image_debug, prompt_debug, total_second_length_debug],
-        outputs=[]
-    )
-    
-    prompt_debug.change(
-        fn=handle_field_debug_change,
-        inputs=[input_image_debug, prompt_debug, total_second_length_debug],
-        outputs=[]
-    )
-    
-    total_second_length_debug.change(
-        fn=handle_field_debug_change,
-        inputs=[input_image_debug, prompt_debug, total_second_length_debug],
-        outputs=[]
-    )
-
-block.launch(mcp_server=False)
+from diffusers_helper.hf_login import login
+
+import os
+
+os.environ['HF_HOME'] = os.path.abspath(os.path.realpath(os.path.join(os.path.dirname(__file__), './hf_download')))
+
+import spaces
+import gradio as gr
+import torch
+import traceback
+import einops
+import safetensors.torch as sf
+import numpy as np
+import argparse
+import random
+import math
+# 20250506 pftq: Added for video input loading
+import decord
+# 20250506 pftq: Added for progress bars in video_encode
+from tqdm import tqdm
+# 20250506 pftq: Normalize file paths for Windows compatibility
+import pathlib
+# 20250506 pftq: for easier to read timestamp
+from datetime import datetime
+# 20250508 pftq: for saving prompt to mp4 comments metadata
+import imageio_ffmpeg
+import tempfile
+import shutil
+import subprocess
+
+from PIL import Image
+from diffusers import AutoencoderKLHunyuanVideo
+from transformers import LlamaModel, CLIPTextModel, LlamaTokenizerFast, CLIPTokenizer
+from diffusers_helper.hunyuan import encode_prompt_conds, vae_decode, vae_encode, vae_decode_fake
+from diffusers_helper.utils import save_bcthw_as_mp4, crop_or_pad_yield_mask, soft_append_bcthw, resize_and_center_crop, state_dict_weighted_merge, state_dict_offset_merge, generate_timestamp
+from diffusers_helper.models.hunyuan_video_packed import HunyuanVideoTransformer3DModelPacked
+from diffusers_helper.pipelines.k_diffusion_hunyuan import sample_hunyuan
+from diffusers_helper.memory import cpu, gpu, get_cuda_free_memory_gb, move_model_to_device_with_memory_preservation, offload_model_from_device_for_memory_preservation, fake_diffusers_current_device, DynamicSwapInstaller, unload_complete_models, load_model_as_complete
+from diffusers_helper.thread_utils import AsyncStream, async_run
+from diffusers_helper.gradio.progress_bar import make_progress_bar_css, make_progress_bar_html
+from transformers import SiglipImageProcessor, SiglipVisionModel
+from diffusers_helper.clip_vision import hf_clip_vision_encode
+from diffusers_helper.bucket_tools import find_nearest_bucket
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, HunyuanVideoTransformer3DModel, HunyuanVideoPipeline
+
+
+if torch.cuda.device_count() > 0:
+    free_mem_gb = get_cuda_free_memory_gb(gpu)
+    high_vram = free_mem_gb > 80
+    
+    print(f'Free VRAM {free_mem_gb} GB')
+    print(f'High-VRAM Mode: {high_vram}')
+    
+    text_encoder = LlamaModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder', torch_dtype=torch.float16).cpu()
+    text_encoder_2 = CLIPTextModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder_2', torch_dtype=torch.float16).cpu()
+    tokenizer = LlamaTokenizerFast.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer')
+    tokenizer_2 = CLIPTokenizer.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer_2')
+    vae = AutoencoderKLHunyuanVideo.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='vae', torch_dtype=torch.float16).cpu()
+    
+    feature_extractor = SiglipImageProcessor.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='feature_extractor')
+    image_encoder = SiglipVisionModel.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='image_encoder', torch_dtype=torch.float16).cpu()
+    
+    transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained('lllyasviel/FramePack_F1_I2V_HY_20250503', torch_dtype=torch.bfloat16).cpu()
+    
+    vae.eval()
+    text_encoder.eval()
+    text_encoder_2.eval()
+    image_encoder.eval()
+    transformer.eval()
+    
+    if not high_vram:
+        vae.enable_slicing()
+        vae.enable_tiling()
+    
+    transformer.high_quality_fp32_output_for_inference = True
+    print('transformer.high_quality_fp32_output_for_inference = True')
+    
+    transformer.to(dtype=torch.bfloat16)
+    vae.to(dtype=torch.float16)
+    image_encoder.to(dtype=torch.float16)
+    text_encoder.to(dtype=torch.float16)
+    text_encoder_2.to(dtype=torch.float16)
+    
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    text_encoder_2.requires_grad_(False)
+    image_encoder.requires_grad_(False)
+    transformer.requires_grad_(False)
+    
+    if not high_vram:
+        # DynamicSwapInstaller is same as huggingface's enable_sequential_offload but 3x faster
+        DynamicSwapInstaller.install_model(transformer, device=gpu)
+        DynamicSwapInstaller.install_model(text_encoder, device=gpu)
+    else:
+        text_encoder.to(gpu)
+        text_encoder_2.to(gpu)
+        image_encoder.to(gpu)
+        vae.to(gpu)
+        transformer.to(gpu)
+
+stream = AsyncStream()
+
+outputs_folder = './outputs/'
+os.makedirs(outputs_folder, exist_ok=True)
+
+input_image_debug_value = prompt_debug_value = total_second_length_debug_value = None
+
+@spaces.GPU()
+@torch.no_grad()
+def video_encode(video_path, resolution, no_resize, vae, vae_batch_size=16, device="cuda", width=None, height=None):
+    """
+    Encode a video into latent representations using the VAE.
+
+    Args:
+        video_path: Path to the input video file.
+        vae: AutoencoderKLHunyuanVideo model.
+        height, width: Target resolution for resizing frames.
+        vae_batch_size: Number of frames to process per batch.
+        device: Device for computation (e.g., "cuda").
+
+    Returns:
+        start_latent: Latent of the first frame (for compatibility with original code).
+        input_image_np: First frame as numpy array (for CLIP vision encoding).
+        history_latents: Latents of all frames (shape: [1, channels, frames, height//8, width//8]).
+        fps: Frames per second of the input video.
+    """
+    # 20250506 pftq: Normalize video path for Windows compatibility
+    video_path = str(pathlib.Path(video_path).resolve())
+    print(f"Processing video: {video_path}")
+
+    # 20250506 pftq: Check CUDA availability and fallback to CPU if needed
+    if device == "cuda" and not torch.cuda.is_available():
+        print("CUDA is not available, falling back to CPU")
+        device = "cpu"
+
+    try:
+        # 20250506 pftq: Load video and get FPS
+        print("Initializing VideoReader...")
+        vr = decord.VideoReader(video_path)
+        fps = vr.get_avg_fps()  # Get input video FPS
+        num_real_frames = len(vr)
+        print(f"Video loaded: {num_real_frames} frames, FPS: {fps}")
+
+        # Truncate to nearest latent size (multiple of 4)
+        latent_size_factor = 4
+        num_frames = (num_real_frames // latent_size_factor) * latent_size_factor
+        if num_frames != num_real_frames:
+            print(f"Truncating video from {num_real_frames} to {num_frames} frames for latent size compatibility")
+        num_real_frames = num_frames
+
+        # 20250506 pftq: Read frames
+        print("Reading video frames...")
+        frames = vr.get_batch(range(num_real_frames)).asnumpy()  # Shape: (num_real_frames, height, width, channels)
+        print(f"Frames read: {frames.shape}")
+
+        # 20250506 pftq: Get native video resolution
+        native_height, native_width = frames.shape[1], frames.shape[2]
+        print(f"Native video resolution: {native_width}x{native_height}")
+
+        # 20250506 pftq: Use native resolution if height/width not specified, otherwise use provided values
+        target_height = native_height if height is None else height
+        target_width = native_width if width is None else width
+
+        # 20250506 pftq: Adjust to nearest bucket for model compatibility
+        if not no_resize:
+            target_height, target_width = find_nearest_bucket(target_height, target_width, resolution=resolution)
+            print(f"Adjusted resolution: {target_width}x{target_height}")
+        else:
+            print(f"Using native resolution without resizing: {target_width}x{target_height}")
+
+        # 20250506 pftq: Preprocess frames to match original image processing
+        processed_frames = []
+        for i, frame in enumerate(frames):
+            #print(f"Preprocessing frame {i+1}/{num_frames}")
+            frame_np = resize_and_center_crop(frame, target_width=target_width, target_height=target_height)
+            processed_frames.append(frame_np)
+        processed_frames = np.stack(processed_frames)  # Shape: (num_real_frames, height, width, channels)
+        print(f"Frames preprocessed: {processed_frames.shape}")
+
+        # 20250506 pftq: Save first frame for CLIP vision encoding
+        input_image_np = processed_frames[0]
+
+        # 20250506 pftq: Convert to tensor and normalize to [-1, 1]
+        print("Converting frames to tensor...")
+        frames_pt = torch.from_numpy(processed_frames).float() / 127.5 - 1
+        frames_pt = frames_pt.permute(0, 3, 1, 2)  # Shape: (num_real_frames, channels, height, width)
+        frames_pt = frames_pt.unsqueeze(0)  # Shape: (1, num_real_frames, channels, height, width)
+        frames_pt = frames_pt.permute(0, 2, 1, 3, 4)  # Shape: (1, channels, num_real_frames, height, width)
+        print(f"Tensor shape: {frames_pt.shape}")
+
+        # 20250507 pftq: Save pixel frames for use in worker
+        input_video_pixels = frames_pt.cpu()
+
+        # 20250506 pftq: Move to device
+        print(f"Moving tensor to device: {device}")
+        frames_pt = frames_pt.to(device)
+        print("Tensor moved to device")
+
+        # 20250506 pftq: Move VAE to device
+        print(f"Moving VAE to device: {device}")
+        vae.to(device)
+        print("VAE moved to device")
+
+        # 20250506 pftq: Encode frames in batches
+        print(f"Encoding input video frames in VAE batch size {vae_batch_size} (reduce if memory issues here or if forcing video resolution)")
+        latents = []
+        vae.eval()
+        with torch.no_grad():
+            for i in tqdm(range(0, frames_pt.shape[2], vae_batch_size), desc="Encoding video frames", mininterval=0.1):
+                #print(f"Encoding batch {i//vae_batch_size + 1}: frames {i} to {min(i + vae_batch_size, frames_pt.shape[2])}")
+                batch = frames_pt[:, :, i:i + vae_batch_size]  # Shape: (1, channels, batch_size, height, width)
+                try:
+                    # 20250506 pftq: Log GPU memory before encoding
+                    if device == "cuda":
+                        free_mem = torch.cuda.memory_allocated() / 1024**3
+                        #print(f"GPU memory before encoding: {free_mem:.2f} GB")
+                    batch_latent = vae_encode(batch, vae)
+                    # 20250506 pftq: Synchronize CUDA to catch issues
+                    if device == "cuda":
+                        torch.cuda.synchronize()
+                        #print(f"GPU memory after encoding: {torch.cuda.memory_allocated() / 1024**3:.2f} GB")
+                    latents.append(batch_latent)
+                    #print(f"Batch encoded, latent shape: {batch_latent.shape}")
+                except RuntimeError as e:
+                    print(f"Error during VAE encoding: {str(e)}")
+                    if device == "cuda" and "out of memory" in str(e).lower():
+                        print("CUDA out of memory, try reducing vae_batch_size or using CPU")
+                    raise
+
+        # 20250506 pftq: Concatenate latents
+        print("Concatenating latents...")
+        history_latents = torch.cat(latents, dim=2)  # Shape: (1, channels, frames, height//8, width//8)
+        print(f"History latents shape: {history_latents.shape}")
+
+        # 20250506 pftq: Get first frame's latent
+        start_latent = history_latents[:, :, :1]  # Shape: (1, channels, 1, height//8, width//8)
+        print(f"Start latent shape: {start_latent.shape}")
+
+        # 20250506 pftq: Move VAE back to CPU to free GPU memory
+        if device == "cuda":
+            vae.to(cpu)
+            torch.cuda.empty_cache()
+            print("VAE moved back to CPU, CUDA cache cleared")
+
+        return start_latent, input_image_np, history_latents, fps, target_height, target_width, input_video_pixels
+
+    except Exception as e:
+        print(f"Error in video_encode: {str(e)}")
+        raise
+
+# 20250508 pftq: for saving prompt to mp4 metadata comments
+def set_mp4_comments_imageio_ffmpeg(input_file, comments):
+    try:
+        # Get the path to the bundled FFmpeg binary from imageio-ffmpeg
+        ffmpeg_path = imageio_ffmpeg.get_ffmpeg_exe()
+
+        # Check if input file exists
+        if not os.path.exists(input_file):
+            print(f"Error: Input file {input_file} does not exist")
+            return False
+
+        # Create a temporary file path
+        temp_file = tempfile.NamedTemporaryFile(suffix='.mp4', delete=False).name
+
+        # FFmpeg command using the bundled binary
+        command = [
+            ffmpeg_path,                   # Use imageio-ffmpeg's FFmpeg
+            '-i', input_file,              # input file
+            '-metadata', f'comment={comments}',  # set comment metadata
+            '-c:v', 'copy',                # copy video stream without re-encoding
+            '-c:a', 'copy',                # copy audio stream without re-encoding
+            '-y',                          # overwrite output file if it exists
+            temp_file                      # temporary output file
+        ]
+
+        # Run the FFmpeg command
+        result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+
+        if result.returncode == 0:
+            # Replace the original file with the modified one
+            shutil.move(temp_file, input_file)
+            print(f"Successfully added comments to {input_file}")
+            return True
+        else:
+            # Clean up temp file if FFmpeg fails
+            if os.path.exists(temp_file):
+                os.remove(temp_file)
+            print(f"Error: FFmpeg failed with message:\n{result.stderr}")
+            return False
+
+    except Exception as e:
+        # Clean up temp file in case of other errors
+        if 'temp_file' in locals() and os.path.exists(temp_file):
+            os.remove(temp_file)
+        print(f"Error saving prompt to video metadata, ffmpeg may be required: "+str(e))
+        return False
+
+@torch.no_grad()
+def worker(input_image, prompts, n_prompt, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf):
+    def encode_prompt(prompt, n_prompt):
+        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)
+
+        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)
+        return [llama_vec, clip_l_pooler, llama_vec_n, clip_l_pooler_n, llama_attention_mask, llama_attention_mask_n]
+
+    total_latent_sections = (total_second_length * 30) / (latent_window_size * 4)
+    total_latent_sections = int(max(round(total_latent_sections), 1))
+
+    job_id = generate_timestamp()
+
+    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)
+
+        prompt_parameters = []
+        
+        for prompt_part in prompts:
+            prompt_parameters.append(encode_prompt(prompt_part, n_prompt))
+
+        # Processing input image
+
+        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Image processing ...'))))
+
+        H, W, C = input_image.shape
+        height, width = find_nearest_bucket(H, W, resolution=640)
+        input_image_np = resize_and_center_crop(input_image, target_width=width, target_height=height)
+
+        Image.fromarray(input_image_np).save(os.path.join(outputs_folder, f'{job_id}.png'))
+
+        input_image_pt = torch.from_numpy(input_image_np).float() / 127.5 - 1
+        input_image_pt = input_image_pt.permute(2, 0, 1)[None, :, None]
+
+        # VAE encoding
+
+        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'VAE encoding ...'))))
+
+        if not high_vram:
+            load_model_as_complete(vae, target_device=gpu)
+
+        start_latent = vae_encode(input_image_pt, vae)
+
+        # 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
+
+        image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
+
+        # Sampling
+
+        stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...'))))
+
+        rnd = torch.Generator("cpu").manual_seed(seed)
+
+        history_latents = torch.zeros(size=(1, 16, 16 + 2 + 1, height // 8, width // 8), dtype=torch.float32).cpu()
+        history_pixels = None
+
+        history_latents = torch.cat([history_latents, start_latent.to(history_latents)], dim=2)
+        total_generated_latent_frames = 1
+
+        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 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(0)
+
+            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 generated frames: {int(max(0, total_generated_latent_frames * 4 - 3))}, Video length: {max(0, (total_generated_latent_frames * 4 - 3) / 30) :.2f} seconds (FPS-30). The video is being extended now ...'
+                stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint))))
+                return
+
+            indices = torch.arange(0, sum([1, 16, 2, 1, latent_window_size])).unsqueeze(0)
+            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)
+
+            clean_latents_4x, clean_latents_2x, clean_latents_1x = history_latents[:, :, -sum([16, 2, 1]):, :, :].split([16, 2, 1], dim=2)
+            clean_latents = torch.cat([start_latent.to(history_latents), clean_latents_1x], dim=2)
+
+            generated_latents = sample_hunyuan(
+                transformer=transformer,
+                sampler='unipc',
+                width=width,
+                height=height,
+                frames=latent_window_size * 4 - 3,
+                real_guidance_scale=cfg,
+                distilled_guidance_scale=gs,
+                guidance_rescale=rs,
+                # shift=3.0,
+                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 = 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()
+
+            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))
+    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(input_image, prompt, t2v, n_prompt, randomize_seed, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf):
+    global total_second_length_debug_value
+
+    if total_second_length_debug_value is not None:
+        return min(total_second_length_debug_value * 60, 600)
+    return total_second_length * 60
+
+
+@spaces.GPU(duration=get_duration)
+def process(input_image, prompt, 
+            t2v=False, 
+            n_prompt="",
+            randomize_seed=True,
+            seed=31337, 
+            total_second_length=5, 
+            latent_window_size=9, 
+            steps=25, 
+            cfg=1.0, 
+            gs=10.0, 
+            rs=0.0, 
+            gpu_memory_preservation=6, 
+            use_teacache=True,
+            mp4_crf=16
+           ):
+    global stream, input_image_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_image_debug_value is not None or prompt_debug_value is not None or total_second_length_debug_value is not None:
+        print("Debug mode")
+        input_image = input_image_debug_value
+        prompt = prompt_debug_value
+        total_second_length = total_second_length_debug_value
+        input_image_debug_value = prompt_debug_value = total_second_length_debug_value = None
+
+    if randomize_seed:
+        seed = random.randint(0, np.iinfo(np.int32).max)
+
+    prompts = prompt.split(";")
+    
+    # assert input_image is not None, 'No input image!'
+    if t2v:
+        default_height, default_width = 640, 640
+        input_image = np.ones((default_height, default_width, 3), dtype=np.uint8) * 255
+        print("No input image provided. Using a blank white image.")
+    
+    yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
+
+    stream = AsyncStream()
+
+    async_run(worker, input_image, prompts, n_prompt, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf)
+
+    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)
+
+        if flag == 'end':
+            yield output_filename, gr.update(visible=False), gr.update(), '', gr.update(interactive=True), gr.update(interactive=False)
+            break
+
+
+def end_process():
+    stream.input_queue.push('end')
+
+
+css = make_progress_bar_css()
+block = gr.Blocks(css=css).queue()
+with block:
+    if torch.cuda.device_count() == 0:
+        with gr.Row():
+            gr.HTML("""
+    <p style="background-color: red;"><big><big><big><b>⚠️To use FramePack, <a href="https://huggingface.co/spaces/Fabrice-TIERCELIN/SUPIR?duplicate=true">duplicate this space</a> and set a GPU with 30 GB VRAM.</b>
+    
+    You can't use FramePack directly here because this space runs on a CPU, which is not enough for FramePack. Please provide <a href="https://huggingface.co/spaces/Fabrice-TIERCELIN/SUPIR/discussions/new">feedback</a> if you have issues.
+    </big></big></big></p>
+    """)
+    # 20250506 pftq: Updated title to reflect video input functionality
+    gr.Markdown('# Framepack F1 with Image Input or with Video Input (Video Extension)')
+    gr.Markdown(f"""### Video diffusion, but feels like image diffusion
+*FramePack F1 - a FramePack model that only predicts future frames from history frames*
+### *beta* FramePack Fill 🖋️- draw a mask over the input image to inpaint the video output
+adapted from the officical code repo [FramePack](https://github.com/lllyasviel/FramePack) by [lllyasviel](lllyasviel/FramePack_F1_I2V_HY_20250503) and [FramePack Studio](https://github.com/colinurbs/FramePack-Studio) 🙌🏻
+    """)
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(sources='upload', type="numpy", label="Image", height=320)
+            prompt = gr.Textbox(label="Prompt", value='')
+            t2v = gr.Checkbox(label="do text-to-video", value=False)
+
+            with gr.Row():
+                start_button = gr.Button(value="Start Generation", variant="primary")
+                end_button = gr.Button(value="End Generation", variant="stop", interactive=False)
+
+            total_second_length = gr.Slider(label="Generated Video Length (Seconds)", minimum=1, maximum=5, value=2, step=0.1)
+            with gr.Accordion("Advanced settings", open=False):
+                use_teacache = gr.Checkbox(label='Use TeaCache', value=True, info='Faster speed, but often makes hands and fingers slightly worse.')
+    
+                n_prompt = gr.Textbox(label="Negative Prompt", value="Missing arm, unrealistic position, blurred, blurry")  # Not used
+                randomize_seed = gr.Checkbox(label='Randomize seed', value=True, info='If checked, the seed is always different')
+                seed = gr.Slider(label="Seed", minimum=0, maximum=np.iinfo(np.int32).max, step=1, randomize=True)
+    
+                
+                latent_window_size = gr.Slider(label="Latent Window Size", minimum=1, maximum=33, value=9, step=1)  # Should not change
+                steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=25, step=1, info='Changing this value is not recommended.')
+    
+                cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=1.0, step=0.01)  # Should not change
+                gs = gr.Slider(label="Distilled CFG Scale", minimum=1.0, maximum=32.0, value=10.0, step=0.01, info='Changing this value is not recommended; 3=blurred motions& & unsharped; 10 focus motion')
+                rs = gr.Slider(label="CFG Re-Scale", minimum=0.0, maximum=1.0, value=0.0, step=0.01)  # Should not change
+    
+                gpu_memory_preservation = gr.Slider(label="GPU Inference Preserved Memory (GB) (larger means slower)", minimum=6, maximum=128, value=6, step=0.1, info="Set this number to a larger value if you encounter OOM. Larger value causes slower speed.")
+    
+                mp4_crf = gr.Slider(label="MP4 Compression", minimum=0, maximum=100, value=16, step=1, info="Lower means better quality. 0 is uncompressed. Change to 16 if you get black outputs. ")
+
+            with gr.Accordion("Debug", open=False):
+                input_image_debug = gr.Image(type="numpy", label="Image Debug", height=320)
+                prompt_debug = gr.Textbox(label="Prompt Debug", value='')
+                total_second_length_debug = gr.Slider(label="Additional Video Length to Generate (Seconds) Debug", minimum=1, maximum=120, value=5, step=0.1)
+
+        with gr.Column():
+            preview_image = gr.Image(label="Next Latents", height=200, visible=False)
+            result_video = gr.Video(label="Finished Frames", autoplay=True, show_share_button=False, height=512, loop=True)
+            progress_desc = gr.Markdown('', elem_classes='no-generating-animation')
+            progress_bar = gr.HTML('', elem_classes='no-generating-animation')
+
+    ips = [input_image, prompt, t2v, n_prompt, randomize_seed, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf]
+    start_button.click(fn=process, inputs=ips, outputs=[result_video, preview_image, progress_desc, progress_bar, start_button, end_button])
+    end_button.click(fn=end_process)
+
+    with gr.Row(visible=False):
+        gr.Examples(
+        examples = [
+                [
+                    "./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",
+                    False, # t2v
+                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
+                    True, # randomize_seed
+                    42, # seed
+                    1, # total_second_length
+                    9, # latent_window_size
+                    25, # steps
+                    1.0, # cfg
+                    10.0, # gs
+                    0.0, # rs
+                    6, # gpu_memory_preservation
+                    True, # use_teacache
+                    16 # mp4_crf
+                ],
+                [
+                    "./img_examples/Example1.png", # input_image
+                    "A dolphin emerges from the water, photorealistic, realistic, intricate details, 8k, insanely detailed",
+                    False, # t2v
+                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
+                    True, # randomize_seed
+                    42, # seed
+                    1, # total_second_length
+                    9, # latent_window_size
+                    25, # steps
+                    1.0, # cfg
+                    10.0, # gs
+                    0.0, # rs
+                    6, # gpu_memory_preservation
+                    True, # use_teacache
+                    16 # mp4_crf
+                ],
+                [
+                    "./img_examples/Example1.png", # input_image
+                    "We are sinking, photorealistic, realistic, intricate details, 8k, insanely detailed",
+                    False, # t2v
+                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
+                    True, # randomize_seed
+                    42, # seed
+                    1, # total_second_length
+                    9, # latent_window_size
+                    25, # steps
+                    1.0, # cfg
+                    10.0, # gs
+                    0.0, # rs
+                    6, # gpu_memory_preservation
+                    True, # use_teacache
+                    16 # mp4_crf
+                ],
+                [
+                    "./img_examples/Example1.png", # input_image
+                    "A boat is passing, photorealistic, realistic, intricate details, 8k, insanely detailed",
+                    False, # t2v
+                    "Missing arm, unrealistic position, blurred, blurry", # n_prompt
+                    True, # randomize_seed
+                    42, # seed
+                    1, # total_second_length
+                    9, # latent_window_size
+                    25, # steps
+                    1.0, # cfg
+                    10.0, # gs
+                    0.0, # rs
+                    6, # gpu_memory_preservation
+                    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 = True,
+    )
+
+    gr.Markdown('## Guide')
+    gr.Markdown("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.")
+
+    
+    def handle_field_debug_change(input_image_debug_data, prompt_debug_data, total_second_length_debug_data):
+        global input_image_debug_value, prompt_debug_value, total_second_length_debug_value
+        input_image_debug_value = input_image_debug_data
+        prompt_debug_value = prompt_debug_data
+        total_second_length_debug_value = total_second_length_debug_data
+        return []
+    
+    input_image_debug.upload(
+        fn=handle_field_debug_change,
+        inputs=[input_image_debug, prompt_debug, total_second_length_debug],
+        outputs=[]
+    )
+    
+    prompt_debug.change(
+        fn=handle_field_debug_change,
+        inputs=[input_image_debug, prompt_debug, total_second_length_debug],
+        outputs=[]
+    )
+    
+    total_second_length_debug.change(
+        fn=handle_field_debug_change,
+        inputs=[input_image_debug, prompt_debug, total_second_length_debug],
+        outputs=[]
+    )
+
+block.launch(mcp_server=False, ssr_mode=False)