Upload 3 files
Browse files- README.md +5 -5
- app.py +34 -95
- requirements.txt +1 -6
README.md
CHANGED
|
@@ -1,14 +1,14 @@
|
|
| 1 |
---
|
| 2 |
-
title: FramePack F1
|
| 3 |
-
emoji:
|
| 4 |
colorFrom: pink
|
| 5 |
colorTo: gray
|
| 6 |
sdk: gradio
|
| 7 |
-
sdk_version: 5.
|
| 8 |
app_file: app.py
|
| 9 |
-
pinned:
|
| 10 |
license: apache-2.0
|
| 11 |
short_description: fast video generation from images & text
|
| 12 |
---
|
| 13 |
-
|
| 14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
| 1 |
---
|
| 2 |
+
title: FramePack F1
|
| 3 |
+
emoji: 📹⚡️
|
| 4 |
colorFrom: pink
|
| 5 |
colorTo: gray
|
| 6 |
sdk: gradio
|
| 7 |
+
sdk_version: 5.30.0
|
| 8 |
app_file: app.py
|
| 9 |
+
pinned: false
|
| 10 |
license: apache-2.0
|
| 11 |
short_description: fast video generation from images & text
|
| 12 |
---
|
| 13 |
+
paper: arxiv:2504.12626
|
| 14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
CHANGED
|
@@ -29,7 +29,7 @@ from diffusers_helper.bucket_tools import find_nearest_bucket
|
|
| 29 |
|
| 30 |
|
| 31 |
free_mem_gb = get_cuda_free_memory_gb(gpu)
|
| 32 |
-
high_vram = free_mem_gb >
|
| 33 |
|
| 34 |
print(f'Free VRAM {free_mem_gb} GB')
|
| 35 |
print(f'High-VRAM Mode: {high_vram}')
|
|
@@ -43,9 +43,6 @@ vae = AutoencoderKLHunyuanVideo.from_pretrained("hunyuanvideo-community/HunyuanV
|
|
| 43 |
feature_extractor = SiglipImageProcessor.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='feature_extractor')
|
| 44 |
image_encoder = SiglipVisionModel.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='image_encoder', torch_dtype=torch.float16).cpu()
|
| 45 |
|
| 46 |
-
# quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
|
| 47 |
-
# transformer = HunyuanVideoTransformer3DModelPacked.from_single_file("https://huggingface.co/sirolim/FramePack_F1_I2V_FP8/resolve/main/FramePack_F1_I2V_HY_fp8_e4m3fn.safetensors", torch_dtype=torch.bfloat16)
|
| 48 |
-
# transformer = HunyuanVideoTransformer3DModelPacked.from_single_file('sirolim/FramePack_F1_I2V_FP8', "FramePack_F1_I2V_HY_fp8_e4m3fn.safetensors", use_safetensors=True, torch_dtype=torch.bfloat16).cpu()
|
| 49 |
transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained('lllyasviel/FramePack_F1_I2V_HY_20250503', torch_dtype=torch.bfloat16).cpu()
|
| 50 |
|
| 51 |
vae.eval()
|
|
@@ -95,10 +92,6 @@ examples = [
|
|
| 95 |
["img_examples/3.png", "The woman dances elegantly among the blossoms, spinning slowly with flowing sleeves and graceful hand movements."],
|
| 96 |
]
|
| 97 |
|
| 98 |
-
input_image_debug_value = None
|
| 99 |
-
prompt_debug_value = None
|
| 100 |
-
total_second_length_debug_value = None
|
| 101 |
-
|
| 102 |
def generate_examples(input_image, prompt):
|
| 103 |
|
| 104 |
t2v=False
|
|
@@ -345,10 +338,6 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
| 345 |
return
|
| 346 |
|
| 347 |
def get_duration(input_image, prompt, t2v, n_prompt, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf):
|
| 348 |
-
global total_second_length_debug_value
|
| 349 |
-
|
| 350 |
-
if total_second_length_debug_value is not None:
|
| 351 |
-
return total_second_length_debug_value * 60
|
| 352 |
return total_second_length * 60
|
| 353 |
|
| 354 |
@spaces.GPU(duration=get_duration)
|
|
@@ -366,25 +355,38 @@ def process(input_image, prompt,
|
|
| 366 |
use_teacache=True,
|
| 367 |
mp4_crf=16
|
| 368 |
):
|
| 369 |
-
global stream
|
| 370 |
-
|
| 371 |
-
if input_image_debug_value is not None:
|
| 372 |
-
input_image = input_image_debug_value
|
| 373 |
-
input_image_debug_value = None
|
| 374 |
-
|
| 375 |
-
if prompt_debug_value is not None:
|
| 376 |
-
prompt = prompt_debug_value
|
| 377 |
-
prompt_debug_value = None
|
| 378 |
-
|
| 379 |
-
if total_second_length_debug_value is not None:
|
| 380 |
-
total_second_length = total_second_length_debug_value
|
| 381 |
-
total_second_length_debug_value = None
|
| 382 |
|
| 383 |
# assert input_image is not None, 'No input image!'
|
| 384 |
if t2v:
|
| 385 |
default_height, default_width = 640, 640
|
| 386 |
input_image = np.ones((default_height, default_width, 3), dtype=np.uint8) * 255
|
| 387 |
print("No input image provided. Using a blank white image.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 388 |
|
| 389 |
yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
|
| 390 |
|
|
@@ -424,12 +426,15 @@ quick_prompts = [[x] for x in quick_prompts]
|
|
| 424 |
css = make_progress_bar_css()
|
| 425 |
block = gr.Blocks(css=css).queue()
|
| 426 |
with block:
|
| 427 |
-
gr.Markdown('# FramePack
|
| 428 |
-
gr.Markdown(f"""###
|
|
|
|
|
|
|
|
|
|
| 429 |
""")
|
| 430 |
with gr.Row():
|
| 431 |
with gr.Column():
|
| 432 |
-
input_image = gr.
|
| 433 |
prompt = gr.Textbox(label="Prompt", value='')
|
| 434 |
t2v = gr.Checkbox(label="do text-to-video", value=False)
|
| 435 |
example_quick_prompts = gr.Dataset(samples=quick_prompts, label='Quick List', samples_per_page=1000, components=[prompt])
|
|
@@ -459,11 +464,6 @@ with block:
|
|
| 459 |
|
| 460 |
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. ")
|
| 461 |
|
| 462 |
-
with gr.Accordion("Debug", open=False):
|
| 463 |
-
input_image_debug = gr.Image(type="numpy", label="Image Debug", height=320)
|
| 464 |
-
prompt_debug = gr.Textbox(label="Prompt Debug", value='')
|
| 465 |
-
total_second_length_debug = gr.Slider(label="Additional Video Length to Generate (Seconds) Debug", minimum=1, maximum=120, value=5, step=0.1)
|
| 466 |
-
|
| 467 |
with gr.Column():
|
| 468 |
preview_image = gr.Image(label="Next Latents", height=200, visible=False)
|
| 469 |
result_video = gr.Video(label="Finished Frames", autoplay=True, show_share_button=False, height=512, loop=True)
|
|
@@ -484,66 +484,5 @@ with block:
|
|
| 484 |
# cache_examples=True
|
| 485 |
# )
|
| 486 |
|
| 487 |
-
with gr.Row(visible=False):
|
| 488 |
-
gr.Examples(
|
| 489 |
-
examples = [
|
| 490 |
-
[
|
| 491 |
-
"./img_examples/Example1.png", # input_image
|
| 492 |
-
"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",
|
| 493 |
-
False, # t2v
|
| 494 |
-
"", # n_prompt
|
| 495 |
-
42, # seed
|
| 496 |
-
1, # total_second_length
|
| 497 |
-
9, # latent_window_size
|
| 498 |
-
25, # steps
|
| 499 |
-
1.0, # cfg
|
| 500 |
-
3.0, # gs
|
| 501 |
-
0.0, # rs
|
| 502 |
-
6, # gpu_memory_preservation
|
| 503 |
-
True, # use_teacache
|
| 504 |
-
16 # mp4_crf
|
| 505 |
-
],
|
| 506 |
-
],
|
| 507 |
-
run_on_click = True,
|
| 508 |
-
fn = process,
|
| 509 |
-
inputs = ips,
|
| 510 |
-
outputs = [result_video, preview_image, progress_desc, progress_bar, start_button, end_button],
|
| 511 |
-
cache_examples = True,
|
| 512 |
-
)
|
| 513 |
-
|
| 514 |
-
|
| 515 |
-
def handle_input_image_debug_upload(input):
|
| 516 |
-
global input_image_debug_value
|
| 517 |
-
input_image_debug_value = input
|
| 518 |
-
return []
|
| 519 |
-
|
| 520 |
-
def handle_prompt_debug_change(input):
|
| 521 |
-
global prompt_debug_value
|
| 522 |
-
prompt_debug_value = input
|
| 523 |
-
return []
|
| 524 |
-
|
| 525 |
-
def handle_total_second_length_debug_change(input):
|
| 526 |
-
global total_second_length_debug_value
|
| 527 |
-
total_second_length_debug_value = input
|
| 528 |
-
return []
|
| 529 |
-
|
| 530 |
-
input_image_debug.upload(
|
| 531 |
-
fn=handle_input_image_debug_upload,
|
| 532 |
-
inputs=[input_image_debug],
|
| 533 |
-
outputs=[]
|
| 534 |
-
)
|
| 535 |
-
|
| 536 |
-
prompt_debug.change(
|
| 537 |
-
fn=handle_prompt_debug_change,
|
| 538 |
-
inputs=[prompt_debug],
|
| 539 |
-
outputs=[]
|
| 540 |
-
)
|
| 541 |
-
|
| 542 |
-
total_second_length_debug.change(
|
| 543 |
-
fn=handle_total_second_length_debug_change,
|
| 544 |
-
inputs=[total_second_length_debug],
|
| 545 |
-
outputs=[]
|
| 546 |
-
)
|
| 547 |
-
|
| 548 |
|
| 549 |
-
block.launch(
|
|
|
|
| 29 |
|
| 30 |
|
| 31 |
free_mem_gb = get_cuda_free_memory_gb(gpu)
|
| 32 |
+
high_vram = free_mem_gb > 60
|
| 33 |
|
| 34 |
print(f'Free VRAM {free_mem_gb} GB')
|
| 35 |
print(f'High-VRAM Mode: {high_vram}')
|
|
|
|
| 43 |
feature_extractor = SiglipImageProcessor.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='feature_extractor')
|
| 44 |
image_encoder = SiglipVisionModel.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='image_encoder', torch_dtype=torch.float16).cpu()
|
| 45 |
|
|
|
|
|
|
|
|
|
|
| 46 |
transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained('lllyasviel/FramePack_F1_I2V_HY_20250503', torch_dtype=torch.bfloat16).cpu()
|
| 47 |
|
| 48 |
vae.eval()
|
|
|
|
| 92 |
["img_examples/3.png", "The woman dances elegantly among the blossoms, spinning slowly with flowing sleeves and graceful hand movements."],
|
| 93 |
]
|
| 94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
def generate_examples(input_image, prompt):
|
| 96 |
|
| 97 |
t2v=False
|
|
|
|
| 338 |
return
|
| 339 |
|
| 340 |
def get_duration(input_image, prompt, t2v, n_prompt, seed, total_second_length, latent_window_size, steps, cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf):
|
|
|
|
|
|
|
|
|
|
|
|
|
| 341 |
return total_second_length * 60
|
| 342 |
|
| 343 |
@spaces.GPU(duration=get_duration)
|
|
|
|
| 355 |
use_teacache=True,
|
| 356 |
mp4_crf=16
|
| 357 |
):
|
| 358 |
+
global stream
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
|
| 360 |
# assert input_image is not None, 'No input image!'
|
| 361 |
if t2v:
|
| 362 |
default_height, default_width = 640, 640
|
| 363 |
input_image = np.ones((default_height, default_width, 3), dtype=np.uint8) * 255
|
| 364 |
print("No input image provided. Using a blank white image.")
|
| 365 |
+
else:
|
| 366 |
+
composite_rgba_uint8 = input_image["composite"]
|
| 367 |
+
|
| 368 |
+
# rgb_uint8 will be (H, W, 3), dtype uint8
|
| 369 |
+
rgb_uint8 = composite_rgba_uint8[:, :, :3]
|
| 370 |
+
# mask_uint8 will be (H, W), dtype uint8
|
| 371 |
+
mask_uint8 = composite_rgba_uint8[:, :, 3]
|
| 372 |
+
|
| 373 |
+
# Create background
|
| 374 |
+
h, w = rgb_uint8.shape[:2]
|
| 375 |
+
# White background, (H, W, 3), dtype uint8
|
| 376 |
+
background_uint8 = np.full((h, w, 3), 255, dtype=np.uint8)
|
| 377 |
+
|
| 378 |
+
# Normalize mask to range [0.0, 1.0].
|
| 379 |
+
alpha_normalized_float32 = mask_uint8.astype(np.float32) / 255.0
|
| 380 |
+
|
| 381 |
+
# Expand alpha to 3 channels to match RGB images for broadcasting.
|
| 382 |
+
# alpha_mask_float32 will have shape (H, W, 3)
|
| 383 |
+
alpha_mask_float32 = np.stack([alpha_normalized_float32] * 3, axis=2)
|
| 384 |
+
|
| 385 |
+
# alpha blending
|
| 386 |
+
blended_image_float32 = rgb_uint8.astype(np.float32) * alpha_mask_float32 + \
|
| 387 |
+
background_uint8.astype(np.float32) * (1.0 - alpha_mask_float32)
|
| 388 |
+
|
| 389 |
+
input_image = np.clip(blended_image_float32, 0, 255).astype(np.uint8)
|
| 390 |
|
| 391 |
yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
|
| 392 |
|
|
|
|
| 426 |
css = make_progress_bar_css()
|
| 427 |
block = gr.Blocks(css=css).queue()
|
| 428 |
with block:
|
| 429 |
+
gr.Markdown('# FramePack-F1')
|
| 430 |
+
gr.Markdown(f"""### Video diffusion, but feels like image diffusion
|
| 431 |
+
*FramePack F1 - a FramePack model that only predicts future frames from history frames*
|
| 432 |
+
### *beta* FramePack Fill 🖋️- draw a mask over the input image to inpaint the video output
|
| 433 |
+
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) 🙌🏻
|
| 434 |
""")
|
| 435 |
with gr.Row():
|
| 436 |
with gr.Column():
|
| 437 |
+
input_image = gr.ImageEditor(type="numpy", label="Image", height=320, brush=gr.Brush(colors=["#ffffff"]))
|
| 438 |
prompt = gr.Textbox(label="Prompt", value='')
|
| 439 |
t2v = gr.Checkbox(label="do text-to-video", value=False)
|
| 440 |
example_quick_prompts = gr.Dataset(samples=quick_prompts, label='Quick List', samples_per_page=1000, components=[prompt])
|
|
|
|
| 464 |
|
| 465 |
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. ")
|
| 466 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 467 |
with gr.Column():
|
| 468 |
preview_image = gr.Image(label="Next Latents", height=200, visible=False)
|
| 469 |
result_video = gr.Video(label="Finished Frames", autoplay=True, show_share_button=False, height=512, loop=True)
|
|
|
|
| 484 |
# cache_examples=True
|
| 485 |
# )
|
| 486 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 487 |
|
| 488 |
+
block.launch(share=True, mcp_server=True)
|
requirements.txt
CHANGED
|
@@ -15,9 +15,4 @@ einops
|
|
| 15 |
opencv-contrib-python
|
| 16 |
safetensors
|
| 17 |
huggingface_hub
|
| 18 |
-
spaces
|
| 19 |
-
decord
|
| 20 |
-
imageio_ffmpeg
|
| 21 |
-
sageattention
|
| 22 |
-
xformers
|
| 23 |
-
bitsandbytes
|
|
|
|
| 15 |
opencv-contrib-python
|
| 16 |
safetensors
|
| 17 |
huggingface_hub
|
| 18 |
+
spaces
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|