Spaces:
Running
on
Zero
Running
on
Zero
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 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 | |
import spaces | |
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 | |
parser = argparse.ArgumentParser() | |
parser.add_argument('--share', action='store_true') | |
parser.add_argument("--server", type=str, default='0.0.0.0') | |
parser.add_argument("--port", type=int, required=False) | |
parser.add_argument("--inbrowser", action='store_true') | |
args = parser.parse_args() | |
print(args) | |
free_mem_gb = get_cuda_free_memory_gb(gpu) | |
high_vram = free_mem_gb > 60 | |
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/FramePackI2V_HY', 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_video_debug_value = prompt_debug_value = total_second_length_debug_value = None | |
# 20250506 pftq: Added function to encode input video frames into latents | |
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] | |
end_of_input_video_image_np = processed_frames[-1] | |
# 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) | |
end_of_input_video_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, end_of_input_video_latent, end_of_input_video_image_np | |
except Exception as e: | |
print(f"Error in video_encode: {str(e)}") | |
raise | |
# 20250507 pftq: New function to encode a single image (end frame) | |
def image_encode(image_np, target_width, target_height, vae, image_encoder, feature_extractor, device="cuda"): | |
""" | |
Encode a single image into a latent and compute its CLIP vision embedding. | |
Args: | |
image_np: Input image as numpy array. | |
target_width, target_height: Exact resolution to resize the image to (matches start frame). | |
vae: AutoencoderKLHunyuanVideo model. | |
image_encoder: SiglipVisionModel for CLIP vision encoding. | |
feature_extractor: SiglipImageProcessor for preprocessing. | |
device: Device for computation (e.g., "cuda"). | |
Returns: | |
latent: Latent representation of the image (shape: [1, channels, 1, height//8, width//8]). | |
clip_embedding: CLIP vision embedding of the image. | |
processed_image_np: Processed image as numpy array (after resizing). | |
""" | |
# 20250507 pftq: Process end frame with exact start frame dimensions | |
print("Processing end frame...") | |
try: | |
print(f"Using exact start frame resolution for end frame: {target_width}x{target_height}") | |
# Resize and preprocess image to match start frame | |
processed_image_np = resize_and_center_crop(image_np, target_width=target_width, target_height=target_height) | |
# Convert to tensor and normalize | |
image_pt = torch.from_numpy(processed_image_np).float() / 127.5 - 1 | |
image_pt = image_pt.permute(2, 0, 1).unsqueeze(0).unsqueeze(2) # Shape: [1, channels, 1, height, width] | |
image_pt = image_pt.to(device) | |
# Move VAE to device | |
vae.to(device) | |
# Encode to latent | |
latent = vae_encode(image_pt, vae) | |
print(f"image_encode vae output shape: {latent.shape}") | |
# Move image encoder to device | |
image_encoder.to(device) | |
# Compute CLIP vision embedding | |
clip_embedding = hf_clip_vision_encode(processed_image_np, feature_extractor, image_encoder).last_hidden_state | |
# Move models back to CPU and clear cache | |
if device == "cuda": | |
vae.to(cpu) | |
image_encoder.to(cpu) | |
torch.cuda.empty_cache() | |
print("VAE and image encoder moved back to CPU, CUDA cache cleared") | |
print(f"End latent shape: {latent.shape}") | |
return latent, clip_embedding, processed_image_np | |
except Exception as e: | |
print(f"Error in image_encode: {str(e)}") | |
raise | |
# 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 | |
# 20250506 pftq: Modified worker to accept video input, and clean frame count | |
def worker(input_video, end_frame, end_frame_weight, 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 | |
start_latent, input_image_np, video_latents, fps, height, width, input_video_pixels, end_of_input_video_latent, end_of_input_video_image_np = 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 | |
start_embedding = image_encoder_last_hidden_state | |
end_of_input_video_output = hf_clip_vision_encode(end_of_input_video_image_np, feature_extractor, image_encoder) | |
end_of_input_video_last_hidden_state = end_of_input_video_output.last_hidden_state | |
end_of_input_video_embedding = end_of_input_video_last_hidden_state | |
# 20250507 pftq: Process end frame if provided | |
end_latent = None | |
end_clip_embedding = None | |
if end_frame is not None: | |
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'End frame encoding ...')))) | |
end_latent, end_clip_embedding, _ = image_encode( | |
end_frame, target_width=width, target_height=height, vae=vae, | |
image_encoder=image_encoder, feature_extractor=feature_extractor, device=gpu | |
) | |
# 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) | |
end_of_input_video_embedding = end_of_input_video_embedding.to(transformer.dtype) | |
# 20250509 pftq: Restored original placement of total_latent_sections after video_encode | |
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 = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")+f"_framepack-videoinput-endframe_{width}-{total_second_length}sec_seed-{seed}_steps-{steps}_distilled-{gs}_cfg-{cfg}" | |
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...')))) | |
rnd = torch.Generator("cpu").manual_seed(seed) | |
history_latents = video_latents.cpu() | |
history_pixels = None | |
total_generated_latent_frames = 0 | |
previous_video = None | |
# 20250509 Generate backwards with end frame for better end frame anchoring | |
if total_latent_sections > 4: | |
latent_paddings = [3] + [2] * (total_latent_sections - 3) + [1, 0] | |
else: | |
latent_paddings = list(reversed(range(total_latent_sections))) | |
for section_index, latent_padding in enumerate(latent_paddings): | |
is_start_of_video = latent_padding == 0 | |
is_end_of_video = latent_padding == latent_paddings[0] | |
latent_padding_size = latent_padding * latent_window_size | |
if stream.input_queue.top() == 'end': | |
stream.output_queue.push(('end', None)) | |
return | |
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): | |
try: | |
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}. Generating part {total_latent_sections - section_index} of {total_latent_sections} backward...' | |
stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint)))) | |
except ConnectionResetError as e: | |
print(f"Suppressed ConnectionResetError in callback: {e}") | |
return | |
# 20250509 pftq: Dynamic frame allocation like original num_clean_frames, fix split error | |
available_frames = video_latents.shape[2] if is_start_of_video else history_latents.shape[2] | |
if is_start_of_video: | |
effective_clean_frames = 1 # avoid jumpcuts from input video | |
else: | |
effective_clean_frames = max(0, num_clean_frames - 1) if num_clean_frames > 1 else 1 | |
clean_latent_pre_frames = effective_clean_frames | |
num_2x_frames = min(2, max(1, available_frames - clean_latent_pre_frames - 1)) if available_frames > clean_latent_pre_frames + 1 else 1 | |
num_4x_frames = min(16, max(1, available_frames - clean_latent_pre_frames - num_2x_frames)) if available_frames > clean_latent_pre_frames + num_2x_frames else 1 | |
total_context_frames = num_2x_frames + num_4x_frames | |
total_context_frames = min(total_context_frames, available_frames - clean_latent_pre_frames) | |
# 20250511 pftq: Dynamically adjust post_frames based on clean_latents_post | |
post_frames = 1 if is_end_of_video and end_latent is not None else effective_clean_frames # 20250511 pftq: Single frame for end_latent, otherwise padding causes still image | |
indices = torch.arange(0, clean_latent_pre_frames + latent_padding_size + latent_window_size + post_frames + num_2x_frames + num_4x_frames).unsqueeze(0) | |
clean_latent_indices_pre, blank_indices, latent_indices, clean_latent_indices_post, clean_latent_2x_indices, clean_latent_4x_indices = indices.split( | |
[clean_latent_pre_frames, latent_padding_size, latent_window_size, post_frames, num_2x_frames, num_4x_frames], dim=1 | |
) | |
clean_latent_indices = torch.cat([clean_latent_indices_pre, clean_latent_indices_post], dim=1) | |
# 20250509 pftq: Split context frames dynamically for 2x and 4x only | |
context_frames = history_latents[:, :, -(total_context_frames + clean_latent_pre_frames):-clean_latent_pre_frames, :, :] if total_context_frames > 0 else history_latents[:, :, :1, :, :] | |
split_sizes = [num_4x_frames, num_2x_frames] | |
split_sizes = [s for s in split_sizes if s > 0] | |
if split_sizes and context_frames.shape[2] >= sum(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[:, :, :1, :, :] | |
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[:, :, :1, :, :] | |
else: | |
clean_latents_4x = clean_latents_2x = history_latents[:, :, :1, :, :] | |
clean_latents_pre = video_latents[:, :, -min(effective_clean_frames, video_latents.shape[2]):].to(history_latents) # smoother motion but jumpcuts if end frame is too different, must change clean_latent_pre_frames to effective_clean_frames also | |
clean_latents_post = history_latents[:, :, :min(effective_clean_frames, history_latents.shape[2]), :, :] # smoother motion, must change post_frames to effective_clean_frames also | |
if is_end_of_video: | |
clean_latents_post = torch.zeros_like(end_of_input_video_latent).to(history_latents) | |
# 20250509 pftq: handle end frame if available | |
if end_latent is not None: | |
#current_end_frame_weight = end_frame_weight * (latent_padding / latent_paddings[0]) | |
#current_end_frame_weight = current_end_frame_weight * 0.5 + 0.5 | |
current_end_frame_weight = end_frame_weight # changing this over time introduces discontinuity | |
# 20250511 pftq: Removed end frame weight adjustment as it has no effect | |
image_encoder_last_hidden_state = (1 - current_end_frame_weight) * end_of_input_video_embedding + end_clip_embedding * current_end_frame_weight | |
image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype) | |
# 20250511 pftq: Use end_latent only | |
if is_end_of_video: | |
clean_latents_post = end_latent.to(history_latents)[:, :, :1, :, :] # Ensure single frame | |
# 20250511 pftq: Pad clean_latents_pre to match clean_latent_pre_frames if needed | |
if clean_latents_pre.shape[2] < clean_latent_pre_frames: | |
clean_latents_pre = clean_latents_pre.repeat(1, 1, clean_latent_pre_frames // clean_latents_pre.shape[2], 1, 1) | |
# 20250511 pftq: Pad clean_latents_post to match post_frames if needed | |
if clean_latents_post.shape[2] < post_frames: | |
clean_latents_post = clean_latents_post.repeat(1, 1, post_frames // clean_latents_post.shape[2], 1, 1) | |
clean_latents = torch.cat([clean_latents_pre, clean_latents_post], dim=2) | |
max_frames = min(latent_window_size * 4 - 3, history_latents.shape[2] * 4) | |
print(f"Generating video {idx+1} of {batch} with seed {seed}, part {total_latent_sections - section_index} of {total_latent_sections} backward") | |
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, | |
) | |
if is_start_of_video: | |
generated_latents = torch.cat([video_latents[:, :, -1:].to(generated_latents), generated_latents], dim=2) | |
total_generated_latent_frames += int(generated_latents.shape[2]) | |
history_latents = torch.cat([generated_latents.to(history_latents), 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 + 1) if is_start_of_video else (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(current_pixels, history_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=fps, crf=mp4_crf) | |
print(f"Latest video saved: {output_filename}") | |
set_mp4_comments_imageio_ffmpeg(output_filename, f"Prompt: {prompt} | Negative Prompt: {n_prompt}") | |
print(f"Prompt saved to mp4 metadata comments: {output_filename}") | |
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)) | |
if is_start_of_video: | |
break | |
history_pixels = torch.cat([input_video_pixels, history_pixels], dim=2) | |
#overlapped_frames = latent_window_size * 4 - 3 | |
#history_pixels = soft_append_bcthw(input_video_pixels, history_pixels, overlapped_frames) | |
output_filename = os.path.join(outputs_folder, f'{job_id}_final.mp4') | |
save_bcthw_as_mp4(history_pixels, output_filename, fps=fps, crf=mp4_crf) | |
print(f"Final video with input blend saved: {output_filename}") | |
set_mp4_comments_imageio_ffmpeg(output_filename, f"Prompt: {prompt} | Negative Prompt: {n_prompt}") | |
print(f"Prompt saved to mp4 metadata comments: {output_filename}") | |
stream.output_queue.push(('file', output_filename)) | |
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 | |
# 20250506 pftq: Modified process to pass clean frame count, etc | |
def get_duration( | |
input_video, end_frame, end_frame_weight, 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 * 2, 600) | |
return total_second_length * 60 * 2 | |
def process( | |
input_video, end_frame, end_frame_weight, 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 or prompt_debug_value is not None or total_second_length_debug_value is not None: | |
input_video = input_video_debug_value | |
prompt = prompt_debug_value | |
total_second_length = total_second_length_debug_value | |
input_video_debug_value = prompt_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, input_video, end_frame, end_frame_weight, 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') | |
quick_prompts = [ | |
'The girl dances gracefully, with clear movements, full of charm.', | |
'A character doing some simple body movements.', | |
] | |
quick_prompts = [[x] for x in quick_prompts] | |
css = make_progress_bar_css() | |
block = gr.Blocks(css=css).queue( | |
max_size=10 # 20250507 pftq: Limit queue size | |
) | |
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 with Video Input (Video Extension) + End Frame') | |
with gr.Row(): | |
with gr.Column(): | |
# 20250506 pftq: Changed to Video input from Image | |
with gr.Row(): | |
input_video = gr.Video(sources='upload', label="Input Video", height=320) | |
with gr.Column(): | |
# 20250507 pftq: Added end_frame + weight | |
end_frame = gr.Image(sources='upload', type="numpy", label="End Frame (Optional) - Reduce context frames if very different from input video or if it is jumpcutting/slowing to still image.", height=320) | |
end_frame_weight = gr.Slider(label="End Frame Weight", minimum=0.0, maximum=1.0, value=1.0, step=0.01, info='Reduce to treat more as a reference image; no effect') | |
prompt = gr.Textbox(label="Prompt", value='') | |
with gr.Row(): | |
start_button = gr.Button(value="Start Generation", variant="primary") | |
end_button = gr.Button(value="End Generation", variant="stop", interactive=False) | |
with gr.Accordion("Advanced settings", open=False): | |
with gr.Row(): | |
use_teacache = gr.Checkbox(label='Use TeaCache', value=True, info='Faster speed, but often makes hands and fingers slightly worse.') | |
no_resize = gr.Checkbox(label='Force Original Video Resolution (No Resizing)', value=False, info='Might run out of VRAM (720p requires > 24GB VRAM).') | |
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) | |
batch = gr.Slider(label="Batch Size (Number of Videos)", minimum=1, maximum=1000, value=1, step=1, info='Generate multiple videos each with a different seed.') | |
resolution = gr.Number(label="Resolution (max width or height)", value=640, precision=0) | |
total_second_length = gr.Slider(label="Additional Video Length to Generate (Seconds)", minimum=1, maximum=120, value=5, step=0.1) | |
# 20250506 pftq: Reduced default distilled guidance scale to improve adherence to input video | |
gs = gr.Slider(label="Distilled CFG Scale", minimum=1.0, maximum=32.0, value=10.0, step=0.01, info='Prompt adherence at the cost of less details from the input video, but to a lesser extent than Context Frames.') | |
cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=1.0, step=0.01, info='Use instead of Distilled for more detail/control + Negative Prompt (make sure Distilled=1). Doubles render time.') # Should not change | |
rs = gr.Slider(label="CFG Re-Scale", minimum=0.0, maximum=1.0, value=0.0, step=0.01) # Should not change | |
n_prompt = gr.Textbox(label="Negative Prompt", value="Missing arm, unrealistic position, blurred, blurry", info='Requires using normal CFG (undistilled) instead of Distilled (set Distilled=1 and CFG > 1).') | |
steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=25, step=1, info='Expensive. Increase for more quality, especially if using high non-distilled CFG.') | |
# 20250506 pftq: Renamed slider to Number of Context Frames and updated description | |
num_clean_frames = gr.Slider(label="Number of Context Frames (Adherence to Video)", minimum=2, maximum=10, value=5, step=1, info="Expensive. Retain more video details. Reduce if memory issues or motion too restricted (jumpcut, ignoring prompt, still).") | |
default_vae = 32 | |
if high_vram: | |
default_vae = 128 | |
elif free_mem_gb>=20: | |
default_vae = 64 | |
vae_batch = gr.Slider(label="VAE Batch Size for Input Video", minimum=4, maximum=256, value=default_vae, step=4, info="Expensive. Increase for better quality frames during fast motion. Reduce if running out of memory") | |
latent_window_size = gr.Slider(label="Latent Window Size", minimum=9, maximum=49, value=9, step=1, info='Expensive. Generate more frames at a time (larger chunks). Less degradation but higher VRAM cost.') | |
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_video_debug = gr.Video(sources='upload', label="Input Video 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') | |
with gr.Row(visible=False): | |
gr.Examples( | |
examples = [ | |
[ | |
"./img_examples/Example1.mp4", # input_video | |
None, # end_frame | |
0.0, # end_frame_weight | |
"View of the sea as far as the eye can see, from the seaside, a piece of land is barely visible on the horizon at the middle, the sky is radiant, reflections of the sun in the water, photorealistic, realistic, intricate details, 8k, insanely detailed", | |
"Missing arm, unrealistic position, blurred, blurry", # n_prompt | |
True, # randomize_seed | |
42, # seed | |
1, # batch | |
640, # resolution | |
1, # total_second_length | |
9, # latent_window_size | |
25, # steps | |
1.0, # cfg | |
10.0, # gs | |
0.0, # rs | |
6, # gpu_memory_preservation | |
True, # use_teacache | |
False, # no_resize | |
16, # mp4_crf | |
5, # num_clean_frames | |
default_vae | |
], | |
], | |
run_on_click = True, | |
fn = process, | |
inputs = [input_video, end_frame, end_frame_weight, 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], | |
outputs = [result_video, preview_image, progress_desc, progress_bar, start_button, end_button], | |
cache_examples = True, | |
) | |
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> | |
""") | |
# 20250506 pftq: Updated inputs to include num_clean_frames | |
ips = [input_video, end_frame, end_frame_weight, 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] | |
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) | |
def handle_field_debug_change(input_video_debug_data, prompt_debug_data, total_second_length_debug_data): | |
global input_video_debug_value, prompt_debug_value, total_second_length_debug_value | |
input_video_debug_value = input_video_debug_data | |
prompt_debug_value = prompt_debug_data | |
total_second_length_debug_value = total_second_length_debug_data | |
return [] | |
input_video_debug.upload( | |
fn=handle_field_debug_change, | |
inputs=[input_video_debug, prompt_debug, total_second_length_debug], | |
outputs=[] | |
) | |
prompt_debug.change( | |
fn=handle_field_debug_change, | |
inputs=[input_video_debug, prompt_debug, total_second_length_debug], | |
outputs=[] | |
) | |
total_second_length_debug.change( | |
fn=handle_field_debug_change, | |
inputs=[input_video_debug, prompt_debug, total_second_length_debug], | |
outputs=[] | |
) | |
block.launch(share=True) |