#############################################
# from diffusers_helper.hf_login import login
# 필요시 HF 로그인 사용 (주석 해제 후)
#############################################
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 math
import time
# Hugging Face Spaces 환경 인지 확인
IN_HF_SPACE = os.environ.get('SPACE_ID') is not None
# --------- 번역 딕셔너리(영어 고정) ---------
translations = {
"en": {
"title": "FramePack - Image to Video Generation",
"upload_image": "Upload Image",
"prompt": "Prompt",
"quick_prompts": "Quick Prompts",
"start_generation": "Generate",
"stop_generation": "Stop",
"use_teacache": "Use TeaCache",
"teacache_info": "Faster speed, but may result in slightly worse finger and hand generation.",
"negative_prompt": "Negative Prompt",
"seed": "Seed",
# 최대 4초로 UI 표기 수정
"video_length": "Video Length (max 4 seconds)",
"latent_window": "Latent Window Size",
"steps": "Inference Steps",
"steps_info": "Changing this value is not recommended.",
"cfg_scale": "CFG Scale",
"distilled_cfg": "Distilled CFG Scale",
"distilled_cfg_info": "Changing this value is not recommended.",
"cfg_rescale": "CFG Rescale",
"gpu_memory": "GPU Memory Preservation (GB) (larger means slower)",
"gpu_memory_info": "Set this to a larger value if you encounter OOM errors. Larger values cause slower speed.",
"next_latents": "Next Latents",
"generated_video": "Generated Video",
"sampling_note": "Note: The model predicts future frames from past frames. If the start action isn't immediately visible, please wait for more frames.",
"error_message": "Error",
"processing_error": "Processing error",
"network_error": "Network connection is unstable, model download timed out. Please try again later.",
"memory_error": "GPU memory insufficient, please try increasing GPU memory preservation value or reduce video length.",
"model_error": "Failed to load model, possibly due to network issues or high server load. Please try again later.",
"partial_video": "Processing error, but partial video has been generated",
"processing_interrupt": "Processing was interrupted, but partial video has been generated"
}
}
def get_translation(key):
return translations["en"].get(key, key)
#############################################
# diffusers_helper 관련 임포트
#############################################
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 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.utils import (
generate_timestamp,
save_bcthw_as_mp4,
resize_and_center_crop,
crop_or_pad_yield_mask,
soft_append_bcthw
)
from diffusers_helper.bucket_tools import find_nearest_bucket
from diffusers_helper.hunyuan import (
encode_prompt_conds, vae_encode, vae_decode, vae_decode_fake
)
from diffusers_helper.clip_vision import hf_clip_vision_encode
from diffusers_helper.models.hunyuan_video_packed import HunyuanVideoTransformer3DModelPacked
from diffusers_helper.pipelines.k_diffusion_hunyuan import sample_hunyuan
from diffusers import AutoencoderKLHunyuanVideo
from transformers import (
LlamaModel, CLIPTextModel,
LlamaTokenizerFast, CLIPTokenizer,
SiglipVisionModel, SiglipImageProcessor
)
#############################################
# GPU 체크
#############################################
GPU_AVAILABLE = torch.cuda.is_available()
free_mem_gb = 0.0
high_vram = False
if GPU_AVAILABLE:
try:
free_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
high_vram = (free_mem_gb > 60)
except:
pass
print(f"GPU Available: {GPU_AVAILABLE}, free_mem_gb={free_mem_gb}, high_vram={high_vram}")
cpu_fallback_mode = not GPU_AVAILABLE
last_update_time = time.time()
#############################################
# 모델 로드 (전역)
#############################################
text_encoder = None
text_encoder_2 = None
tokenizer = None
tokenizer_2 = None
vae = None
feature_extractor = None
image_encoder = None
transformer = None
# 아래 로직은 질문에 제시된 '두 번째 코드'의 모델 로드 부분을 거의 그대로 사용
def load_global_models():
global text_encoder, text_encoder_2, tokenizer, tokenizer_2
global vae, feature_extractor, image_encoder, transformer
global cpu_fallback_mode
# 이미 로드되었으면 패스
if transformer is not None:
return
# GPU 메모리 정보
device = gpu if GPU_AVAILABLE else cpu
# diffusers_helper.memory.get_cuda_free_memory_gb(gpu)로 더 정확히 구해도 됨
print("Loading models...")
# ======== 실 코드: 두 번째 예시 기준 =========
# (1) 하이브리드 (if high_vram -> GPU로 로드, 아니면 CPU + DynamicSwap)
# 반드시 float16, bfloat16로 로드
text_encoder_local = LlamaModel.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='text_encoder',
torch_dtype=torch.float16
).cpu()
text_encoder_2_local = CLIPTextModel.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='text_encoder_2',
torch_dtype=torch.float16
).cpu()
tokenizer_local = LlamaTokenizerFast.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='tokenizer'
)
tokenizer_2_local = CLIPTokenizer.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='tokenizer_2'
)
vae_local = AutoencoderKLHunyuanVideo.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='vae',
torch_dtype=torch.float16
).cpu()
feature_extractor_local = SiglipImageProcessor.from_pretrained(
"lllyasviel/flux_redux_bfl", subfolder='feature_extractor'
)
image_encoder_local = SiglipVisionModel.from_pretrained(
"lllyasviel/flux_redux_bfl",
subfolder='image_encoder',
torch_dtype=torch.float16
).cpu()
# FramePack_F1_I2V_HY_20250503 (bfloat16)
transformer_local = HunyuanVideoTransformer3DModelPacked.from_pretrained(
'lllyasviel/FramePack_F1_I2V_HY_20250503',
torch_dtype=torch.bfloat16
).cpu()
# eval & dtype
vae_local.eval()
text_encoder_local.eval()
text_encoder_2_local.eval()
image_encoder_local.eval()
transformer_local.eval()
# VAE slicing for low VRAM
if not high_vram:
vae_local.enable_slicing()
vae_local.enable_tiling()
# 오프로드용
transformer_local.high_quality_fp32_output_for_inference = True
transformer_local.to(dtype=torch.bfloat16)
vae_local.to(dtype=torch.float16)
image_encoder_local.to(dtype=torch.float16)
text_encoder_local.to(dtype=torch.float16)
text_encoder_2_local.to(dtype=torch.float16)
# requires_grad_(False)
for m in [vae_local, text_encoder_local, text_encoder_2_local, image_encoder_local, transformer_local]:
m.requires_grad_(False)
# GPU 모드 & VRAM 많으면 전부 GPU
# 그렇지 않으면 DynamicSwap
if GPU_AVAILABLE:
if not high_vram:
DynamicSwapInstaller.install_model(transformer_local, device=gpu)
DynamicSwapInstaller.install_model(text_encoder_local, device=gpu)
else:
text_encoder_local.to(gpu)
text_encoder_2_local.to(gpu)
image_encoder_local.to(gpu)
vae_local.to(gpu)
transformer_local.to(gpu)
else:
cpu_fallback_mode = True
# 글로벌에 할당
print("Model loaded.")
text_encoder = text_encoder_local
text_encoder_2 = text_encoder_2_local
tokenizer = tokenizer_local
tokenizer_2 = tokenizer_2_local
vae = vae_local
feature_extractor = feature_extractor_local
image_encoder = image_encoder_local
transformer = transformer_local
#############################################
# Worker 로직 (두 번째 코드) 그대로
#############################################
stream = AsyncStream()
outputs_folder = './outputs/'
os.makedirs(outputs_folder, exist_ok=True)
@torch.no_grad()
def worker(
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
):
"""
실제 샘플링 로직 (두 번째 코드 기반)
"""
load_global_models() # 모델 로딩
global text_encoder, text_encoder_2, tokenizer, tokenizer_2
global vae, feature_extractor, image_encoder, transformer
global last_update_time
# 최대 4초로 고정
total_second_length = min(total_second_length, 4.0)
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:
# GPU 적을 경우 Unload
if not high_vram and GPU_AVAILABLE:
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 and GPU_AVAILABLE:
fake_diffusers_current_device(text_encoder, gpu)
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 == 1.0:
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_mask = crop_or_pad_yield_mask(llama_vec, length=512)
llama_vec_n, llama_mask_n = crop_or_pad_yield_mask(llama_vec_n, length=512)
# Image processing
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)
if cpu_fallback_mode:
height = min(height, 320)
width = min(width, 320)
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 encode
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'VAE encoding ...'))))
if not high_vram and GPU_AVAILABLE:
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 and GPU_AVAILABLE:
load_model_as_complete(image_encoder, target_device=gpu)
image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
# dtype
llama_vec = llama_vec.to(transformer.dtype)
llama_vec_n = llama_vec_n.to(transformer.dtype)
clip_l_pooler = clip_l_pooler.to(transformer.dtype)
clip_l_pooler_n = clip_l_pooler_n.to(transformer.dtype)
image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
# Start sampling
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...'))))
rnd = torch.Generator("cpu").manual_seed(seed)
# 초기 history latents
history_latents = torch.zeros(size=(1, 16, 16 + 2 + 1, height // 8, width // 8), dtype=torch.float32).cpu()
history_pixels = None
# start_latent 붙이기
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 {section_index+1}/{total_latent_sections}')
if not high_vram and GPU_AVAILABLE:
unload_complete_models()
move_model_to_device_with_memory_preservation(transformer, target_device=gpu, preserved_memory_gb=gpu_memory_preservation)
# teacache
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 stops generation.')
current_step = d['i'] + 1
percentage = int(100.0 * current_step / steps)
hint = f'Sampling {current_step}/{steps}'
desc = f'Section {section_index+1}/{total_latent_sections}'
stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint))))
return
# indices
frames_per_section = latent_window_size * 4 - 3
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[:, :, -19:, :, :].split([16, 2, 1], dim=2)
clean_latents = torch.cat([start_latent.to(history_latents), clean_latents_1x], dim=2)
try:
generated_latents = sample_hunyuan(
transformer=transformer,
sampler='unipc',
width=width,
height=height,
frames=frames_per_section,
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_mask,
prompt_poolers=clip_l_pooler,
negative_prompt_embeds=llama_vec_n,
negative_prompt_embeds_mask=llama_mask_n,
negative_prompt_poolers=clip_l_pooler_n,
device=gpu if GPU_AVAILABLE else cpu,
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
)
except KeyboardInterrupt:
print("User cancelled.")
stream.output_queue.push(('end', None))
return
except Exception as e:
traceback.print_exc()
stream.output_queue.push(('end', None))
return
total_generated_latent_frames += generated_latents.shape[2]
history_latents = torch.cat([history_latents, generated_latents.to(history_latents)], dim=2)
if not high_vram and GPU_AVAILABLE:
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 = frames_per_section
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 and GPU_AVAILABLE:
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=16) # CRF=16
stream.output_queue.push(('file', output_filename))
except:
traceback.print_exc()
if not high_vram and GPU_AVAILABLE:
unload_complete_models(text_encoder, text_encoder_2, image_encoder, vae, transformer)
stream.output_queue.push(('end', None))
return
def end_process():
"""
중단 요청
"""
global stream
stream.input_queue.push('end')
# Gradio에서 이 worker 함수를 비동기로 호출
def process(
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
):
global stream
if input_image is None:
raise ValueError("No input image provided.")
yield None, None, "", "", gr.update(interactive=False), gr.update(interactive=True)
stream = AsyncStream()
async_run(
worker,
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
)
output_filename = None
prev_filename = None
error_message = None
while True:
flag, data = stream.output_queue.next()
if flag == 'file':
output_filename = data
prev_filename = output_filename
yield output_filename, gr.update(), gr.update(), "", gr.update(interactive=False), gr.update(interactive=True)
elif 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)
elif flag == 'error':
error_message = data
print(f"Error: {error_message}")
elif flag == 'end':
if output_filename is None and prev_filename:
output_filename = prev_filename
# 에러가 있었으면 에러 표시
if error_message:
yield (
output_filename, # 마지막 파일 (또는 None)
gr.update(visible=False),
gr.update(),
f"{error_message}
",
gr.update(interactive=True),
gr.update(interactive=False)
)
else:
yield (
output_filename, gr.update(visible=False), gr.update(), "", gr.update(interactive=True), gr.update(interactive=False)
)
break
# UI CSS
def make_custom_css():
base_progress_css = make_progress_bar_css()
pastel_css = """
body {
background: #faf9ff !important;
font-family: "Noto Sans", sans-serif;
}
#app-container {
max-width: 1200px;
margin: 0 auto;
padding: 1rem;
position: relative;
}
#app-container h1 {
color: #5F5AA2;
margin-bottom: 1.2rem;
font-weight: 700;
text-shadow: 1px 1px 2px #bbb;
}
.gr-panel {
background: #ffffffcc;
border: 1px solid #e1dff0;
border-radius: 8px;
padding: 1rem;
box-shadow: 0 1px 3px rgba(0,0,0,0.1);
}
.button-container button {
min-height: 45px;
font-size: 1rem;
font-weight: 600;
border-radius: 6px;
}
.button-container button#start-button {
background-color: #A289E3 !important;
color: #fff !important;
border: 1px solid #a58de2;
}
.button-container button#stop-button {
background-color: #F48A9B !important;
color: #fff !important;
border: 1px solid #f18fa0;
}
.button-container button:hover {
filter: brightness(0.95);
}
.preview-container, .video-container {
border: 1px solid #ded9f2;
border-radius: 8px;
overflow: hidden;
}
.progress-container {
margin-top: 15px;
margin-bottom: 15px;
}
.error-message {
background-color: #FFF5F5;
border: 1px solid #FED7D7;
color: #E53E3E;
padding: 10px;
border-radius: 4px;
margin-top: 10px;
font-weight: 500;
}
@media (max-width: 768px) {
#app-container {
padding: 0.5rem;
}
.mobile-full-width {
flex-direction: column !important;
}
.mobile-full-width > .gr-block {
width: 100% !important;
}
}
"""
return base_progress_css + pastel_css
css = make_custom_css()
# 샘플 프롬프트
quick_prompts = [
["The girl dances gracefully, with clear movements, full of charm."],
["A character doing some simple body movements."]
]
# Gradio UI
block = gr.Blocks(css=css).queue()
with block:
gr.HTML("FramePack - Image to Video Generation
")
with gr.Row(elem_classes="mobile-full-width"):
# 왼쪽
with gr.Column(scale=1, elem_classes="gr-panel"):
input_image = gr.Image(
label=get_translation("upload_image"),
type="numpy",
height=320
)
prompt = gr.Textbox(
label=get_translation("prompt"),
value=''
)
example_quick_prompts = gr.Dataset(
samples=quick_prompts,
label=get_translation("quick_prompts"),
samples_per_page=1000,
components=[prompt]
)
example_quick_prompts.click(
fn=lambda x: x[0],
inputs=[example_quick_prompts],
outputs=prompt,
show_progress=False,
queue=False
)
# 오른쪽
with gr.Column(scale=1, elem_classes="gr-panel"):
with gr.Row(elem_classes="button-container"):
start_button = gr.Button(
value=get_translation("start_generation"),
elem_id="start-button",
variant="primary"
)
stop_button = gr.Button(
value=get_translation("stop_generation"),
elem_id="stop-button",
interactive=False
)
result_video = gr.Video(
label=get_translation("generated_video"),
autoplay=True,
loop=True,
height=320,
elem_classes="video-container"
)
preview_image = gr.Image(
label=get_translation("next_latents"),
visible=False,
height=150,
elem_classes="preview-container"
)
gr.Markdown(get_translation("sampling_note"))
with gr.Group(elem_classes="progress-container"):
progress_desc = gr.Markdown('')
progress_bar = gr.HTML('')
error_message = gr.HTML('', visible=True)
# Advanced
with gr.Accordion("Advanced Settings", open=False, elem_classes="gr-panel"):
use_teacache = gr.Checkbox(
label=get_translation("use_teacache"),
value=True,
info=get_translation("teacache_info")
)
n_prompt = gr.Textbox(label=get_translation("negative_prompt"), value="", visible=False)
seed = gr.Number(
label=get_translation("seed"),
value=31337,
precision=0
)
# 기본 2초, 최대 4초
total_second_length = gr.Slider(
label=get_translation("video_length"),
minimum=1,
maximum=4,
value=2,
step=0.1
)
latent_window_size = gr.Slider(
label=get_translation("latent_window"),
minimum=1,
maximum=33,
value=9,
step=1,
visible=False
)
steps = gr.Slider(
label=get_translation("steps"),
minimum=1,
maximum=100,
value=25,
step=1,
info=get_translation("steps_info")
)
cfg = gr.Slider(
label=get_translation("cfg_scale"),
minimum=1.0,
maximum=32.0,
value=1.0,
step=0.01,
visible=False
)
gs = gr.Slider(
label=get_translation("distilled_cfg"),
minimum=1.0,
maximum=32.0,
value=10.0,
step=0.01,
info=get_translation("distilled_cfg_info")
)
rs = gr.Slider(
label=get_translation("cfg_rescale"),
minimum=0.0,
maximum=1.0,
value=0.0,
step=0.01,
visible=False
)
gpu_memory_preservation = gr.Slider(
label=get_translation("gpu_memory"),
minimum=6,
maximum=128,
value=6,
step=0.1,
info=get_translation("gpu_memory_info")
)
# 버튼 처리
inputs_list = [
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
]
start_button.click(
fn=process,
inputs=inputs_list,
outputs=[result_video, preview_image, progress_desc, progress_bar, start_button, stop_button]
)
stop_button.click(fn=end_process)
block.launch()
#############################################
# from diffusers_helper.hf_login import login
# 필요시 HF 로그인 사용 (주석 해제 후)
#############################################
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 math
import time
# Hugging Face Spaces 환경 인지 확인
IN_HF_SPACE = os.environ.get('SPACE_ID') is not None
# --------- 번역 딕셔너리(영어 고정) ---------
translations = {
"en": {
"title": "FramePack - Image to Video Generation",
"upload_image": "Upload Image",
"prompt": "Prompt",
"quick_prompts": "Quick Prompts",
"start_generation": "Generate",
"stop_generation": "Stop",
"use_teacache": "Use TeaCache",
"teacache_info": "Faster speed, but may result in slightly worse finger and hand generation.",
"negative_prompt": "Negative Prompt",
"seed": "Seed",
# 최대 4초로 UI 표기 수정
"video_length": "Video Length (max 4 seconds)",
"latent_window": "Latent Window Size",
"steps": "Inference Steps",
"steps_info": "Changing this value is not recommended.",
"cfg_scale": "CFG Scale",
"distilled_cfg": "Distilled CFG Scale",
"distilled_cfg_info": "Changing this value is not recommended.",
"cfg_rescale": "CFG Rescale",
"gpu_memory": "GPU Memory Preservation (GB) (larger means slower)",
"gpu_memory_info": "Set this to a larger value if you encounter OOM errors. Larger values cause slower speed.",
"next_latents": "Next Latents",
"generated_video": "Generated Video",
"sampling_note": "Note: The model predicts future frames from past frames. If the start action isn't immediately visible, please wait for more frames.",
"error_message": "Error",
"processing_error": "Processing error",
"network_error": "Network connection is unstable, model download timed out. Please try again later.",
"memory_error": "GPU memory insufficient, please try increasing GPU memory preservation value or reduce video length.",
"model_error": "Failed to load model, possibly due to network issues or high server load. Please try again later.",
"partial_video": "Processing error, but partial video has been generated",
"processing_interrupt": "Processing was interrupted, but partial video has been generated"
}
}
def get_translation(key):
return translations["en"].get(key, key)
#############################################
# diffusers_helper 관련 임포트
#############################################
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 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.utils import (
generate_timestamp,
save_bcthw_as_mp4,
resize_and_center_crop,
crop_or_pad_yield_mask,
soft_append_bcthw
)
from diffusers_helper.bucket_tools import find_nearest_bucket
from diffusers_helper.hunyuan import (
encode_prompt_conds, vae_encode, vae_decode, vae_decode_fake
)
from diffusers_helper.clip_vision import hf_clip_vision_encode
from diffusers_helper.models.hunyuan_video_packed import HunyuanVideoTransformer3DModelPacked
from diffusers_helper.pipelines.k_diffusion_hunyuan import sample_hunyuan
from diffusers import AutoencoderKLHunyuanVideo
from transformers import (
LlamaModel, CLIPTextModel,
LlamaTokenizerFast, CLIPTokenizer,
SiglipVisionModel, SiglipImageProcessor
)
#############################################
# GPU 체크
#############################################
GPU_AVAILABLE = torch.cuda.is_available()
free_mem_gb = 0.0
high_vram = False
if GPU_AVAILABLE:
try:
free_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
high_vram = (free_mem_gb > 60)
except:
pass
print(f"GPU Available: {GPU_AVAILABLE}, free_mem_gb={free_mem_gb}, high_vram={high_vram}")
cpu_fallback_mode = not GPU_AVAILABLE
last_update_time = time.time()
#############################################
# 모델 로드 (전역)
#############################################
text_encoder = None
text_encoder_2 = None
tokenizer = None
tokenizer_2 = None
vae = None
feature_extractor = None
image_encoder = None
transformer = None
# 아래 로직은 질문에 제시된 '두 번째 코드'의 모델 로드 부분을 거의 그대로 사용
def load_global_models():
global text_encoder, text_encoder_2, tokenizer, tokenizer_2
global vae, feature_extractor, image_encoder, transformer
global cpu_fallback_mode
# 이미 로드되었으면 패스
if transformer is not None:
return
# GPU 메모리 정보
device = gpu if GPU_AVAILABLE else cpu
# diffusers_helper.memory.get_cuda_free_memory_gb(gpu)로 더 정확히 구해도 됨
print("Loading models...")
# ======== 실 코드: 두 번째 예시 기준 =========
# (1) 하이브리드 (if high_vram -> GPU로 로드, 아니면 CPU + DynamicSwap)
# 반드시 float16, bfloat16로 로드
text_encoder_local = LlamaModel.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='text_encoder',
torch_dtype=torch.float16
).cpu()
text_encoder_2_local = CLIPTextModel.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='text_encoder_2',
torch_dtype=torch.float16
).cpu()
tokenizer_local = LlamaTokenizerFast.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='tokenizer'
)
tokenizer_2_local = CLIPTokenizer.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='tokenizer_2'
)
vae_local = AutoencoderKLHunyuanVideo.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
subfolder='vae',
torch_dtype=torch.float16
).cpu()
feature_extractor_local = SiglipImageProcessor.from_pretrained(
"lllyasviel/flux_redux_bfl", subfolder='feature_extractor'
)
image_encoder_local = SiglipVisionModel.from_pretrained(
"lllyasviel/flux_redux_bfl",
subfolder='image_encoder',
torch_dtype=torch.float16
).cpu()
# FramePack_F1_I2V_HY_20250503 (bfloat16)
transformer_local = HunyuanVideoTransformer3DModelPacked.from_pretrained(
'lllyasviel/FramePack_F1_I2V_HY_20250503',
torch_dtype=torch.bfloat16
).cpu()
# eval & dtype
vae_local.eval()
text_encoder_local.eval()
text_encoder_2_local.eval()
image_encoder_local.eval()
transformer_local.eval()
# VAE slicing for low VRAM
if not high_vram:
vae_local.enable_slicing()
vae_local.enable_tiling()
# 오프로드용
transformer_local.high_quality_fp32_output_for_inference = True
transformer_local.to(dtype=torch.bfloat16)
vae_local.to(dtype=torch.float16)
image_encoder_local.to(dtype=torch.float16)
text_encoder_local.to(dtype=torch.float16)
text_encoder_2_local.to(dtype=torch.float16)
# requires_grad_(False)
for m in [vae_local, text_encoder_local, text_encoder_2_local, image_encoder_local, transformer_local]:
m.requires_grad_(False)
# GPU 모드 & VRAM 많으면 전부 GPU
# 그렇지 않으면 DynamicSwap
if GPU_AVAILABLE:
if not high_vram:
DynamicSwapInstaller.install_model(transformer_local, device=gpu)
DynamicSwapInstaller.install_model(text_encoder_local, device=gpu)
else:
text_encoder_local.to(gpu)
text_encoder_2_local.to(gpu)
image_encoder_local.to(gpu)
vae_local.to(gpu)
transformer_local.to(gpu)
else:
cpu_fallback_mode = True
# 글로벌에 할당
print("Model loaded.")
text_encoder = text_encoder_local
text_encoder_2 = text_encoder_2_local
tokenizer = tokenizer_local
tokenizer_2 = tokenizer_2_local
vae = vae_local
feature_extractor = feature_extractor_local
image_encoder = image_encoder_local
transformer = transformer_local
#############################################
# Worker 로직 (두 번째 코드) 그대로
#############################################
stream = AsyncStream()
outputs_folder = './outputs/'
os.makedirs(outputs_folder, exist_ok=True)
@torch.no_grad()
def worker(
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
):
"""
실제 샘플링 로직 (두 번째 코드 기반)
"""
load_global_models() # 모델 로딩
global text_encoder, text_encoder_2, tokenizer, tokenizer_2
global vae, feature_extractor, image_encoder, transformer
global last_update_time
# 최대 4초로 고정
total_second_length = min(total_second_length, 4.0)
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:
# GPU 적을 경우 Unload
if not high_vram and GPU_AVAILABLE:
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 and GPU_AVAILABLE:
fake_diffusers_current_device(text_encoder, gpu)
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 == 1.0:
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_mask = crop_or_pad_yield_mask(llama_vec, length=512)
llama_vec_n, llama_mask_n = crop_or_pad_yield_mask(llama_vec_n, length=512)
# Image processing
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)
if cpu_fallback_mode:
height = min(height, 320)
width = min(width, 320)
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 encode
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'VAE encoding ...'))))
if not high_vram and GPU_AVAILABLE:
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 and GPU_AVAILABLE:
load_model_as_complete(image_encoder, target_device=gpu)
image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
# dtype
llama_vec = llama_vec.to(transformer.dtype)
llama_vec_n = llama_vec_n.to(transformer.dtype)
clip_l_pooler = clip_l_pooler.to(transformer.dtype)
clip_l_pooler_n = clip_l_pooler_n.to(transformer.dtype)
image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
# Start sampling
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...'))))
rnd = torch.Generator("cpu").manual_seed(seed)
# 초기 history latents
history_latents = torch.zeros(size=(1, 16, 16 + 2 + 1, height // 8, width // 8), dtype=torch.float32).cpu()
history_pixels = None
# start_latent 붙이기
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 {section_index+1}/{total_latent_sections}')
if not high_vram and GPU_AVAILABLE:
unload_complete_models()
move_model_to_device_with_memory_preservation(transformer, target_device=gpu, preserved_memory_gb=gpu_memory_preservation)
# teacache
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 stops generation.')
current_step = d['i'] + 1
percentage = int(100.0 * current_step / steps)
hint = f'Sampling {current_step}/{steps}'
desc = f'Section {section_index+1}/{total_latent_sections}'
stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint))))
return
# indices
frames_per_section = latent_window_size * 4 - 3
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[:, :, -19:, :, :].split([16, 2, 1], dim=2)
clean_latents = torch.cat([start_latent.to(history_latents), clean_latents_1x], dim=2)
try:
generated_latents = sample_hunyuan(
transformer=transformer,
sampler='unipc',
width=width,
height=height,
frames=frames_per_section,
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_mask,
prompt_poolers=clip_l_pooler,
negative_prompt_embeds=llama_vec_n,
negative_prompt_embeds_mask=llama_mask_n,
negative_prompt_poolers=clip_l_pooler_n,
device=gpu if GPU_AVAILABLE else cpu,
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
)
except KeyboardInterrupt:
print("User cancelled.")
stream.output_queue.push(('end', None))
return
except Exception as e:
traceback.print_exc()
stream.output_queue.push(('end', None))
return
total_generated_latent_frames += generated_latents.shape[2]
history_latents = torch.cat([history_latents, generated_latents.to(history_latents)], dim=2)
if not high_vram and GPU_AVAILABLE:
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 = frames_per_section
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 and GPU_AVAILABLE:
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=16) # CRF=16
stream.output_queue.push(('file', output_filename))
except:
traceback.print_exc()
if not high_vram and GPU_AVAILABLE:
unload_complete_models(text_encoder, text_encoder_2, image_encoder, vae, transformer)
stream.output_queue.push(('end', None))
return
def end_process():
"""
중단 요청
"""
global stream
stream.input_queue.push('end')
# Gradio에서 이 worker 함수를 비동기로 호출
def process(
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
):
global stream
if input_image is None:
raise ValueError("No input image provided.")
yield None, None, "", "", gr.update(interactive=False), gr.update(interactive=True)
stream = AsyncStream()
async_run(
worker,
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
)
output_filename = None
prev_filename = None
error_message = None
while True:
flag, data = stream.output_queue.next()
if flag == 'file':
output_filename = data
prev_filename = output_filename
yield output_filename, gr.update(), gr.update(), "", gr.update(interactive=False), gr.update(interactive=True)
elif 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)
elif flag == 'error':
error_message = data
print(f"Error: {error_message}")
elif flag == 'end':
if output_filename is None and prev_filename:
output_filename = prev_filename
# 에러가 있었으면 에러 표시
if error_message:
yield (
output_filename, # 마지막 파일 (또는 None)
gr.update(visible=False),
gr.update(),
f"{error_message}
",
gr.update(interactive=True),
gr.update(interactive=False)
)
else:
yield (
output_filename, gr.update(visible=False), gr.update(), "", gr.update(interactive=True), gr.update(interactive=False)
)
break
# UI CSS
def make_custom_css():
base_progress_css = make_progress_bar_css()
pastel_css = """
body {
background: #faf9ff !important;
font-family: "Noto Sans", sans-serif;
}
#app-container {
max-width: 1200px;
margin: 0 auto;
padding: 1rem;
position: relative;
}
#app-container h1 {
color: #5F5AA2;
margin-bottom: 1.2rem;
font-weight: 700;
text-shadow: 1px 1px 2px #bbb;
}
.gr-panel {
background: #ffffffcc;
border: 1px solid #e1dff0;
border-radius: 8px;
padding: 1rem;
box-shadow: 0 1px 3px rgba(0,0,0,0.1);
}
.button-container button {
min-height: 45px;
font-size: 1rem;
font-weight: 600;
border-radius: 6px;
}
.button-container button#start-button {
background-color: #A289E3 !important;
color: #fff !important;
border: 1px solid #a58de2;
}
.button-container button#stop-button {
background-color: #F48A9B !important;
color: #fff !important;
border: 1px solid #f18fa0;
}
.button-container button:hover {
filter: brightness(0.95);
}
.preview-container, .video-container {
border: 1px solid #ded9f2;
border-radius: 8px;
overflow: hidden;
}
.progress-container {
margin-top: 15px;
margin-bottom: 15px;
}
.error-message {
background-color: #FFF5F5;
border: 1px solid #FED7D7;
color: #E53E3E;
padding: 10px;
border-radius: 4px;
margin-top: 10px;
font-weight: 500;
}
@media (max-width: 768px) {
#app-container {
padding: 0.5rem;
}
.mobile-full-width {
flex-direction: column !important;
}
.mobile-full-width > .gr-block {
width: 100% !important;
}
}
"""
return base_progress_css + pastel_css
css = make_custom_css()
# 샘플 프롬프트
quick_prompts = [
["The girl dances gracefully, with clear movements, full of charm."],
["A character doing some simple body movements."]
]
# Gradio UI
block = gr.Blocks(css=css).queue()
with block:
gr.HTML("FramePack - Image to Video Generation
")
with gr.Row(elem_classes="mobile-full-width"):
# 왼쪽
with gr.Column(scale=1, elem_classes="gr-panel"):
input_image = gr.Image(
label=get_translation("upload_image"),
type="numpy",
height=320
)
prompt = gr.Textbox(
label=get_translation("prompt"),
value=''
)
example_quick_prompts = gr.Dataset(
samples=quick_prompts,
label=get_translation("quick_prompts"),
samples_per_page=1000,
components=[prompt]
)
example_quick_prompts.click(
fn=lambda x: x[0],
inputs=[example_quick_prompts],
outputs=prompt,
show_progress=False,
queue=False
)
# 오른쪽
with gr.Column(scale=1, elem_classes="gr-panel"):
with gr.Row(elem_classes="button-container"):
start_button = gr.Button(
value=get_translation("start_generation"),
elem_id="start-button",
variant="primary"
)
stop_button = gr.Button(
value=get_translation("stop_generation"),
elem_id="stop-button",
interactive=False
)
result_video = gr.Video(
label=get_translation("generated_video"),
autoplay=True,
loop=True,
height=320,
elem_classes="video-container"
)
preview_image = gr.Image(
label=get_translation("next_latents"),
visible=False,
height=150,
elem_classes="preview-container"
)
gr.Markdown(get_translation("sampling_note"))
with gr.Group(elem_classes="progress-container"):
progress_desc = gr.Markdown('')
progress_bar = gr.HTML('')
error_message = gr.HTML('', visible=True)
# Advanced
with gr.Accordion("Advanced Settings", open=False, elem_classes="gr-panel"):
use_teacache = gr.Checkbox(
label=get_translation("use_teacache"),
value=True,
info=get_translation("teacache_info")
)
n_prompt = gr.Textbox(label=get_translation("negative_prompt"), value="", visible=False)
seed = gr.Number(
label=get_translation("seed"),
value=31337,
precision=0
)
# 기본 2초, 최대 4초
total_second_length = gr.Slider(
label=get_translation("video_length"),
minimum=1,
maximum=4,
value=2,
step=0.1
)
latent_window_size = gr.Slider(
label=get_translation("latent_window"),
minimum=1,
maximum=33,
value=9,
step=1,
visible=False
)
steps = gr.Slider(
label=get_translation("steps"),
minimum=1,
maximum=100,
value=25,
step=1,
info=get_translation("steps_info")
)
cfg = gr.Slider(
label=get_translation("cfg_scale"),
minimum=1.0,
maximum=32.0,
value=1.0,
step=0.01,
visible=False
)
gs = gr.Slider(
label=get_translation("distilled_cfg"),
minimum=1.0,
maximum=32.0,
value=10.0,
step=0.01,
info=get_translation("distilled_cfg_info")
)
rs = gr.Slider(
label=get_translation("cfg_rescale"),
minimum=0.0,
maximum=1.0,
value=0.0,
step=0.01,
visible=False
)
gpu_memory_preservation = gr.Slider(
label=get_translation("gpu_memory"),
minimum=6,
maximum=128,
value=6,
step=0.1,
info=get_translation("gpu_memory_info")
)
# 버튼 처리
inputs_list = [
input_image, prompt, n_prompt, seed,
total_second_length, latent_window_size, steps,
cfg, gs, rs, gpu_memory_preservation, use_teacache
]
start_button.click(
fn=process,
inputs=inputs_list,
outputs=[result_video, preview_image, progress_desc, progress_bar, start_button, stop_button]
)
stop_button.click(fn=end_process)
block.launch()