Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -47,13 +47,14 @@ from transformers import SiglipImageProcessor, SiglipVisionModel
|
|
47 |
from diffusers_helper.clip_vision import hf_clip_vision_encode
|
48 |
from diffusers_helper.bucket_tools import find_nearest_bucket
|
49 |
|
50 |
-
|
51 |
free_mem_gb = get_cuda_free_memory_gb(gpu)
|
52 |
high_vram = free_mem_gb > 60
|
53 |
|
54 |
print(f'Free VRAM {free_mem_gb} GB')
|
55 |
print(f'High-VRAM Mode: {high_vram}')
|
56 |
|
|
|
57 |
text_encoder = LlamaModel.from_pretrained(
|
58 |
"hunyuanvideo-community/HunyuanVideo",
|
59 |
subfolder='text_encoder',
|
@@ -93,12 +94,14 @@ transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained(
|
|
93 |
torch_dtype=torch.bfloat16
|
94 |
).cpu()
|
95 |
|
|
|
96 |
vae.eval()
|
97 |
text_encoder.eval()
|
98 |
text_encoder_2.eval()
|
99 |
image_encoder.eval()
|
100 |
transformer.eval()
|
101 |
|
|
|
102 |
if not high_vram:
|
103 |
vae.enable_slicing()
|
104 |
vae.enable_tiling()
|
@@ -106,20 +109,22 @@ if not high_vram:
|
|
106 |
transformer.high_quality_fp32_output_for_inference = True
|
107 |
print('transformer.high_quality_fp32_output_for_inference = True')
|
108 |
|
|
|
109 |
transformer.to(dtype=torch.bfloat16)
|
110 |
vae.to(dtype=torch.float16)
|
111 |
image_encoder.to(dtype=torch.float16)
|
112 |
text_encoder.to(dtype=torch.float16)
|
113 |
text_encoder_2.to(dtype=torch.float16)
|
114 |
|
|
|
115 |
vae.requires_grad_(False)
|
116 |
text_encoder.requires_grad_(False)
|
117 |
text_encoder_2.requires_grad_(False)
|
118 |
image_encoder.requires_grad_(False)
|
119 |
transformer.requires_grad_(False)
|
120 |
|
|
|
121 |
if not high_vram:
|
122 |
-
# DynamicSwapInstaller is same as huggingface's enable_sequential_offload but 3x faster
|
123 |
DynamicSwapInstaller.install_model(transformer, device=gpu)
|
124 |
DynamicSwapInstaller.install_model(text_encoder, device=gpu)
|
125 |
else:
|
@@ -140,6 +145,7 @@ examples = [
|
|
140 |
["img_examples/3.png", "The woman dances elegantly among the blossoms, spinning slowly with flowing sleeves and graceful hand movements."]
|
141 |
]
|
142 |
|
|
|
143 |
def generate_examples(input_image, prompt):
|
144 |
t2v=False
|
145 |
n_prompt=""
|
@@ -192,7 +198,8 @@ def generate_examples(input_image, prompt):
|
|
192 |
yield (
|
193 |
gr.update(),
|
194 |
gr.update(visible=True, value=preview),
|
195 |
-
desc,
|
|
|
196 |
gr.update(interactive=False),
|
197 |
gr.update(interactive=True)
|
198 |
)
|
@@ -211,98 +218,69 @@ def generate_examples(input_image, prompt):
|
|
211 |
@torch.no_grad()
|
212 |
def worker(
|
213 |
input_image, prompt, n_prompt, seed,
|
214 |
-
total_second_length, latent_window_size,
|
215 |
-
|
216 |
-
gpu_memory_preservation, use_teacache, mp4_crf
|
217 |
):
|
|
|
218 |
total_latent_sections = (total_second_length * 30) / (latent_window_size * 4)
|
219 |
total_latent_sections = int(max(round(total_latent_sections), 1))
|
220 |
|
221 |
job_id = generate_timestamp()
|
222 |
|
223 |
-
stream.output_queue.push(
|
224 |
-
('progress', (None, '', make_progress_bar_html(0, 'Starting ...')))
|
225 |
-
)
|
226 |
|
227 |
try:
|
228 |
-
#
|
229 |
if not high_vram:
|
230 |
unload_complete_models(
|
231 |
text_encoder, text_encoder_2, image_encoder, vae, transformer
|
232 |
)
|
233 |
|
234 |
# Text encoding
|
235 |
-
stream.output_queue.push(
|
236 |
-
('progress', (None, '', make_progress_bar_html(0, 'Text encoding ...')))
|
237 |
-
)
|
238 |
|
239 |
if not high_vram:
|
240 |
fake_diffusers_current_device(text_encoder, gpu)
|
241 |
load_model_as_complete(text_encoder_2, target_device=gpu)
|
242 |
|
243 |
-
llama_vec, clip_l_pooler = encode_prompt_conds(
|
244 |
-
prompt, text_encoder, text_encoder_2,
|
245 |
-
tokenizer, tokenizer_2
|
246 |
-
)
|
247 |
|
248 |
if cfg == 1:
|
249 |
-
llama_vec_n, clip_l_pooler_n = (
|
250 |
-
torch.zeros_like(llama_vec),
|
251 |
-
torch.zeros_like(clip_l_pooler)
|
252 |
-
)
|
253 |
else:
|
254 |
-
llama_vec_n, clip_l_pooler_n = encode_prompt_conds(
|
255 |
-
n_prompt, text_encoder, text_encoder_2,
|
256 |
-
tokenizer, tokenizer_2
|
257 |
-
)
|
258 |
|
259 |
llama_vec, llama_attention_mask = crop_or_pad_yield_mask(llama_vec, length=512)
|
260 |
llama_vec_n, llama_attention_mask_n = crop_or_pad_yield_mask(llama_vec_n, length=512)
|
261 |
|
262 |
-
#
|
263 |
-
stream.output_queue.push(
|
264 |
-
('progress', (None, '', make_progress_bar_html(0, 'Image processing ...')))
|
265 |
-
)
|
266 |
|
267 |
H, W, C = input_image.shape
|
268 |
height, width = find_nearest_bucket(H, W, resolution=640)
|
269 |
-
input_image_np = resize_and_center_crop(
|
270 |
-
input_image,
|
271 |
-
target_width=width,
|
272 |
-
target_height=height
|
273 |
-
)
|
274 |
|
275 |
-
Image.fromarray(input_image_np).save(
|
276 |
-
os.path.join(outputs_folder, f'{job_id}.png')
|
277 |
-
)
|
278 |
|
279 |
input_image_pt = torch.from_numpy(input_image_np).float() / 127.5 - 1
|
280 |
input_image_pt = input_image_pt.permute(2, 0, 1)[None, :, None]
|
281 |
|
282 |
# VAE encoding
|
283 |
-
stream.output_queue.push(
|
284 |
-
('progress', (None, '', make_progress_bar_html(0, 'VAE encoding ...')))
|
285 |
-
)
|
286 |
|
287 |
if not high_vram:
|
288 |
load_model_as_complete(vae, target_device=gpu)
|
289 |
-
|
290 |
start_latent = vae_encode(input_image_pt, vae)
|
291 |
|
292 |
# CLIP Vision
|
293 |
-
stream.output_queue.push(
|
294 |
-
('progress', (None, '', make_progress_bar_html(0, 'CLIP Vision encoding ...')))
|
295 |
-
)
|
296 |
|
297 |
if not high_vram:
|
298 |
load_model_as_complete(image_encoder, target_device=gpu)
|
299 |
-
|
300 |
-
image_encoder_output = hf_clip_vision_encode(
|
301 |
-
input_image_np, feature_extractor, image_encoder
|
302 |
-
)
|
303 |
image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
|
304 |
|
305 |
-
#
|
306 |
llama_vec = llama_vec.to(transformer.dtype)
|
307 |
llama_vec_n = llama_vec_n.to(transformer.dtype)
|
308 |
clip_l_pooler = clip_l_pooler.to(transformer.dtype)
|
@@ -310,9 +288,7 @@ def worker(
|
|
310 |
image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
|
311 |
|
312 |
# Start sampling
|
313 |
-
stream.output_queue.push(
|
314 |
-
('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...')))
|
315 |
-
)
|
316 |
|
317 |
rnd = torch.Generator("cpu").manual_seed(seed)
|
318 |
|
@@ -322,10 +298,8 @@ def worker(
|
|
322 |
).cpu()
|
323 |
history_pixels = None
|
324 |
|
325 |
-
|
326 |
-
|
327 |
-
dim=2
|
328 |
-
)
|
329 |
total_generated_latent_frames = 1
|
330 |
|
331 |
for section_index in range(total_latent_sections):
|
@@ -351,10 +325,7 @@ def worker(
|
|
351 |
preview = d['denoised']
|
352 |
preview = vae_decode_fake(preview)
|
353 |
preview = (preview * 255.0).detach().cpu().numpy().clip(0, 255).astype(np.uint8)
|
354 |
-
preview = einops.rearrange(
|
355 |
-
preview,
|
356 |
-
'b c t h w -> (b h) (t w) c'
|
357 |
-
)
|
358 |
|
359 |
if stream.input_queue.top() == 'end':
|
360 |
stream.output_queue.push(('end', None))
|
@@ -363,15 +334,12 @@ def worker(
|
|
363 |
current_step = d['i'] + 1
|
364 |
percentage = int(100.0 * current_step / steps)
|
365 |
hint = f'Sampling {current_step}/{steps}'
|
366 |
-
desc = f'
|
367 |
-
stream.output_queue.push(
|
368 |
-
('progress', (preview, desc, make_progress_bar_html(percentage, hint)))
|
369 |
-
)
|
370 |
return
|
371 |
|
372 |
indices = torch.arange(
|
373 |
-
0,
|
374 |
-
sum([1, 16, 2, 1, latent_window_size])
|
375 |
).unsqueeze(0)
|
376 |
(
|
377 |
clean_latent_indices_start,
|
@@ -380,14 +348,13 @@ def worker(
|
|
380 |
clean_latent_1x_indices,
|
381 |
latent_indices
|
382 |
) = indices.split([1, 16, 2, 1, latent_window_size], dim=1)
|
383 |
-
|
384 |
-
|
385 |
-
dim=1
|
386 |
-
)
|
387 |
|
388 |
clean_latents_4x, clean_latents_2x, clean_latents_1x = history_latents[
|
389 |
:, :, -sum([16, 2, 1]):, :, :
|
390 |
].split([16, 2, 1], dim=2)
|
|
|
391 |
clean_latents = torch.cat(
|
392 |
[start_latent.to(history_latents), clean_latents_1x],
|
393 |
dim=2
|
@@ -424,21 +391,13 @@ def worker(
|
|
424 |
)
|
425 |
|
426 |
total_generated_latent_frames += int(generated_latents.shape[2])
|
427 |
-
history_latents = torch.cat(
|
428 |
-
[history_latents, generated_latents.to(history_latents)],
|
429 |
-
dim=2
|
430 |
-
)
|
431 |
|
432 |
if not high_vram:
|
433 |
-
offload_model_from_device_for_memory_preservation(
|
434 |
-
transformer, target_device=gpu,
|
435 |
-
preserved_memory_gb=8
|
436 |
-
)
|
437 |
load_model_as_complete(vae, target_device=gpu)
|
438 |
|
439 |
-
real_history_latents = history_latents[
|
440 |
-
:, :, -total_generated_latent_frames:, :, :
|
441 |
-
]
|
442 |
|
443 |
if history_pixels is None:
|
444 |
history_pixels = vae_decode(real_history_latents, vae).cpu()
|
@@ -456,75 +415,55 @@ def worker(
|
|
456 |
if not high_vram:
|
457 |
unload_complete_models()
|
458 |
|
459 |
-
output_filename = os.path.join(
|
460 |
-
|
461 |
-
)
|
462 |
-
|
463 |
-
save_bcthw_as_mp4(
|
464 |
-
history_pixels, output_filename,
|
465 |
-
fps=30, crf=mp4_crf
|
466 |
-
)
|
467 |
|
468 |
-
print(
|
469 |
-
f'Decoded. Current latent shape {real_history_latents.shape}; pixel shape {history_pixels.shape}'
|
470 |
-
)
|
471 |
|
472 |
stream.output_queue.push(('file', output_filename))
|
473 |
|
474 |
except:
|
475 |
traceback.print_exc()
|
476 |
if not high_vram:
|
477 |
-
unload_complete_models(
|
478 |
-
text_encoder, text_encoder_2, image_encoder, vae, transformer
|
479 |
-
)
|
480 |
|
481 |
stream.output_queue.push(('end', None))
|
482 |
return
|
483 |
|
484 |
def get_duration(
|
485 |
-
input_image, prompt, t2v, n_prompt,
|
486 |
-
total_second_length, latent_window_size,
|
487 |
-
cfg, gs, rs, gpu_memory_preservation,
|
|
|
488 |
):
|
489 |
return total_second_length * 60
|
490 |
|
491 |
@spaces.GPU(duration=get_duration)
|
492 |
def process(
|
493 |
-
input_image, prompt, t2v=False, n_prompt="",
|
494 |
-
|
495 |
-
|
496 |
-
|
497 |
):
|
498 |
global stream
|
499 |
-
|
500 |
if t2v:
|
501 |
default_height, default_width = 640, 640
|
502 |
-
input_image = np.ones(
|
503 |
-
(default_height, default_width, 3),
|
504 |
-
dtype=np.uint8
|
505 |
-
) * 255
|
506 |
print("No input image provided. Using a blank white image.")
|
507 |
else:
|
508 |
-
# ImageEditor에서 받은 composite RGBA를 분리
|
509 |
composite_rgba_uint8 = input_image["composite"]
|
510 |
|
511 |
-
# rgb_uint8: (H,W,3)
|
512 |
rgb_uint8 = composite_rgba_uint8[:, :, :3]
|
513 |
-
# mask_uint8: (H,W)
|
514 |
mask_uint8 = composite_rgba_uint8[:, :, 3]
|
515 |
|
516 |
-
# 흰색 배경
|
517 |
h, w = rgb_uint8.shape[:2]
|
518 |
background_uint8 = np.full((h, w, 3), 255, dtype=np.uint8)
|
519 |
|
520 |
-
# 알파 노멀라이즈
|
521 |
alpha_normalized_float32 = mask_uint8.astype(np.float32) / 255.0
|
522 |
alpha_mask_float32 = np.stack([alpha_normalized_float32]*3, axis=2)
|
523 |
|
524 |
-
|
525 |
-
|
526 |
-
rgb_uint8.astype(np.float32) * alpha_mask_float32 + \
|
527 |
-
background_uint8.astype(np.float32) * (1.0 - alpha_mask_float32)
|
528 |
|
529 |
input_image = np.clip(blended_image_float32, 0, 255).astype(np.uint8)
|
530 |
|
@@ -559,7 +498,8 @@ def process(
|
|
559 |
yield (
|
560 |
gr.update(),
|
561 |
gr.update(visible=True, value=preview),
|
562 |
-
desc,
|
|
|
563 |
gr.update(interactive=False),
|
564 |
gr.update(interactive=True)
|
565 |
)
|
@@ -578,16 +518,16 @@ def process(
|
|
578 |
def end_process():
|
579 |
stream.input_queue.push('end')
|
580 |
|
|
|
581 |
quick_prompts = [
|
582 |
'The girl dances gracefully, with clear movements, full of charm.',
|
583 |
'A character doing some simple body movements.'
|
584 |
]
|
585 |
quick_prompts = [[x] for x in quick_prompts]
|
586 |
|
587 |
-
|
588 |
def make_custom_css():
|
589 |
base_progress_css = make_progress_bar_css()
|
590 |
-
# 아래는 예시로 약간 더 파스텔 톤의 스타일 및 카드형 UI
|
591 |
extra_css = """
|
592 |
body {
|
593 |
background: #fafbfe !important;
|
@@ -595,14 +535,14 @@ def make_custom_css():
|
|
595 |
}
|
596 |
#title-container {
|
597 |
text-align: center;
|
598 |
-
padding:
|
599 |
background: linear-gradient(135deg, #a8c0ff 0%, #fbc2eb 100%);
|
600 |
-
border-radius: 0 0
|
601 |
margin-bottom: 20px;
|
602 |
}
|
603 |
#title-container h1 {
|
604 |
color: white;
|
605 |
-
font-size:
|
606 |
margin: 0;
|
607 |
font-weight: 800;
|
608 |
text-shadow: 1px 2px 2px rgba(0,0,0,0.1);
|
@@ -650,35 +590,30 @@ css = make_custom_css()
|
|
650 |
|
651 |
block = gr.Blocks(css=css).queue()
|
652 |
with block:
|
653 |
-
#
|
654 |
-
with gr.
|
655 |
gr.Markdown("<h1>FramePack I2V</h1>")
|
656 |
|
657 |
-
# 설명 부분
|
658 |
gr.Markdown("""
|
659 |
### Video diffusion, but feels like image diffusion
|
660 |
-
FramePack I2V - a model that predicts future frames from
|
661 |
-
|
662 |
-
***beta FramePack Fill*** - You can also paint over the input image to inpaint the video output.
|
663 |
""")
|
664 |
|
665 |
with gr.Row():
|
666 |
with gr.Column():
|
667 |
input_image = gr.ImageEditor(
|
668 |
type="numpy",
|
669 |
-
label="Image (
|
670 |
height=320,
|
671 |
brush=gr.Brush(colors=["#ffffff"])
|
672 |
)
|
673 |
prompt = gr.Textbox(label="Prompt", value='')
|
|
|
674 |
|
675 |
-
t2v = gr.Checkbox(
|
676 |
-
label="Generate from Text Only (no image)?",
|
677 |
-
value=False
|
678 |
-
)
|
679 |
example_quick_prompts = gr.Dataset(
|
680 |
samples=quick_prompts,
|
681 |
-
label="Quick
|
682 |
samples_per_page=1000,
|
683 |
components=[prompt]
|
684 |
)
|
@@ -695,7 +630,7 @@ with block:
|
|
695 |
end_button = gr.Button(value="Stop Generation", elem_id="stop-button", interactive=False)
|
696 |
|
697 |
total_second_length = gr.Slider(
|
698 |
-
label="Total Video Length (
|
699 |
minimum=1,
|
700 |
maximum=5,
|
701 |
value=2,
|
@@ -707,87 +642,81 @@ with block:
|
|
707 |
use_teacache = gr.Checkbox(
|
708 |
label='Use TeaCache',
|
709 |
value=True,
|
710 |
-
info='Faster speed but
|
711 |
)
|
712 |
n_prompt = gr.Textbox(label="Negative Prompt", value="", visible=False)
|
713 |
seed = gr.Number(label="Seed", value=31337, precision=0)
|
714 |
-
|
715 |
latent_window_size = gr.Slider(
|
716 |
label="Latent Window Size",
|
717 |
-
minimum=1,
|
718 |
-
|
719 |
-
value=9,
|
720 |
-
step=1,
|
721 |
visible=False
|
722 |
)
|
723 |
steps = gr.Slider(
|
724 |
label="Steps",
|
725 |
-
minimum=1,
|
726 |
-
|
727 |
-
|
728 |
-
step=1,
|
729 |
-
info='Not recommended to change significantly.'
|
730 |
)
|
731 |
cfg = gr.Slider(
|
732 |
label="CFG Scale",
|
733 |
-
minimum=1.0,
|
734 |
-
|
735 |
-
value=1.0,
|
736 |
-
step=0.01,
|
737 |
visible=False
|
738 |
)
|
739 |
gs = gr.Slider(
|
740 |
label="Distilled CFG Scale",
|
741 |
-
minimum=1.0,
|
742 |
-
|
743 |
-
|
744 |
-
step=0.01,
|
745 |
-
info='Not recommended to change significantly.'
|
746 |
)
|
747 |
rs = gr.Slider(
|
748 |
label="CFG Re-Scale",
|
749 |
-
minimum=0.0,
|
750 |
-
|
751 |
-
value=0.0,
|
752 |
-
step=0.01,
|
753 |
visible=False
|
754 |
)
|
755 |
gpu_memory_preservation = gr.Slider(
|
756 |
label="GPU Memory Preservation (GB)",
|
757 |
-
minimum=6,
|
758 |
-
|
759 |
-
|
760 |
-
step=0.1,
|
761 |
-
info="Increase if OOM occurs (slower speed)."
|
762 |
)
|
763 |
mp4_crf = gr.Slider(
|
764 |
label="MP4 Compression (CRF)",
|
765 |
-
minimum=0,
|
766 |
-
|
767 |
-
|
768 |
-
step=1,
|
769 |
-
info="Lower is higher quality. 16 is recommended."
|
770 |
)
|
771 |
|
772 |
with gr.Column():
|
773 |
-
preview_image = gr.Image(
|
774 |
-
|
775 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
776 |
progress_desc = gr.Markdown('', elem_classes='no-generating-animation')
|
777 |
progress_bar = gr.HTML('', elem_classes='no-generating-animation')
|
778 |
|
|
|
779 |
gr.HTML("""
|
780 |
<div style="text-align:center; margin-top:20px;">
|
781 |
-
|
782 |
-
|
783 |
</div>
|
784 |
""")
|
785 |
|
786 |
-
# 함수 연결
|
787 |
ips = [
|
788 |
input_image, prompt, t2v, n_prompt, seed,
|
789 |
-
total_second_length, latent_window_size,
|
790 |
-
cfg, gs, rs, gpu_memory_preservation,
|
791 |
use_teacache, mp4_crf
|
792 |
]
|
793 |
start_button.click(
|
@@ -797,7 +726,7 @@ with block:
|
|
797 |
)
|
798 |
end_button.click(fn=end_process)
|
799 |
|
800 |
-
#
|
801 |
# gr.Examples(
|
802 |
# examples=examples,
|
803 |
# inputs=[input_image, prompt],
|
|
|
47 |
from diffusers_helper.clip_vision import hf_clip_vision_encode
|
48 |
from diffusers_helper.bucket_tools import find_nearest_bucket
|
49 |
|
50 |
+
# Check GPU memory
|
51 |
free_mem_gb = get_cuda_free_memory_gb(gpu)
|
52 |
high_vram = free_mem_gb > 60
|
53 |
|
54 |
print(f'Free VRAM {free_mem_gb} GB')
|
55 |
print(f'High-VRAM Mode: {high_vram}')
|
56 |
|
57 |
+
# Load models
|
58 |
text_encoder = LlamaModel.from_pretrained(
|
59 |
"hunyuanvideo-community/HunyuanVideo",
|
60 |
subfolder='text_encoder',
|
|
|
94 |
torch_dtype=torch.bfloat16
|
95 |
).cpu()
|
96 |
|
97 |
+
# Evaluation mode
|
98 |
vae.eval()
|
99 |
text_encoder.eval()
|
100 |
text_encoder_2.eval()
|
101 |
image_encoder.eval()
|
102 |
transformer.eval()
|
103 |
|
104 |
+
# Slicing/Tiling for low VRAM
|
105 |
if not high_vram:
|
106 |
vae.enable_slicing()
|
107 |
vae.enable_tiling()
|
|
|
109 |
transformer.high_quality_fp32_output_for_inference = True
|
110 |
print('transformer.high_quality_fp32_output_for_inference = True')
|
111 |
|
112 |
+
# Move to correct dtype
|
113 |
transformer.to(dtype=torch.bfloat16)
|
114 |
vae.to(dtype=torch.float16)
|
115 |
image_encoder.to(dtype=torch.float16)
|
116 |
text_encoder.to(dtype=torch.float16)
|
117 |
text_encoder_2.to(dtype=torch.float16)
|
118 |
|
119 |
+
# No gradient
|
120 |
vae.requires_grad_(False)
|
121 |
text_encoder.requires_grad_(False)
|
122 |
text_encoder_2.requires_grad_(False)
|
123 |
image_encoder.requires_grad_(False)
|
124 |
transformer.requires_grad_(False)
|
125 |
|
126 |
+
# DynamicSwap if low VRAM
|
127 |
if not high_vram:
|
|
|
128 |
DynamicSwapInstaller.install_model(transformer, device=gpu)
|
129 |
DynamicSwapInstaller.install_model(text_encoder, device=gpu)
|
130 |
else:
|
|
|
145 |
["img_examples/3.png", "The woman dances elegantly among the blossoms, spinning slowly with flowing sleeves and graceful hand movements."]
|
146 |
]
|
147 |
|
148 |
+
# Example generation (optional)
|
149 |
def generate_examples(input_image, prompt):
|
150 |
t2v=False
|
151 |
n_prompt=""
|
|
|
198 |
yield (
|
199 |
gr.update(),
|
200 |
gr.update(visible=True, value=preview),
|
201 |
+
desc,
|
202 |
+
html,
|
203 |
gr.update(interactive=False),
|
204 |
gr.update(interactive=True)
|
205 |
)
|
|
|
218 |
@torch.no_grad()
|
219 |
def worker(
|
220 |
input_image, prompt, n_prompt, seed,
|
221 |
+
total_second_length, latent_window_size, steps,
|
222 |
+
cfg, gs, rs, gpu_memory_preservation, use_teacache, mp4_crf
|
|
|
223 |
):
|
224 |
+
# Calculate total sections
|
225 |
total_latent_sections = (total_second_length * 30) / (latent_window_size * 4)
|
226 |
total_latent_sections = int(max(round(total_latent_sections), 1))
|
227 |
|
228 |
job_id = generate_timestamp()
|
229 |
|
230 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Starting ...'))))
|
|
|
|
|
231 |
|
232 |
try:
|
233 |
+
# Unload if VRAM is low
|
234 |
if not high_vram:
|
235 |
unload_complete_models(
|
236 |
text_encoder, text_encoder_2, image_encoder, vae, transformer
|
237 |
)
|
238 |
|
239 |
# Text encoding
|
240 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Text encoding ...'))))
|
|
|
|
|
241 |
|
242 |
if not high_vram:
|
243 |
fake_diffusers_current_device(text_encoder, gpu)
|
244 |
load_model_as_complete(text_encoder_2, target_device=gpu)
|
245 |
|
246 |
+
llama_vec, clip_l_pooler = encode_prompt_conds(prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
|
|
|
|
|
|
|
247 |
|
248 |
if cfg == 1:
|
249 |
+
llama_vec_n, clip_l_pooler_n = torch.zeros_like(llama_vec), torch.zeros_like(clip_l_pooler)
|
|
|
|
|
|
|
250 |
else:
|
251 |
+
llama_vec_n, clip_l_pooler_n = encode_prompt_conds(n_prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
|
|
|
|
|
|
|
252 |
|
253 |
llama_vec, llama_attention_mask = crop_or_pad_yield_mask(llama_vec, length=512)
|
254 |
llama_vec_n, llama_attention_mask_n = crop_or_pad_yield_mask(llama_vec_n, length=512)
|
255 |
|
256 |
+
# Process image
|
257 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Image processing ...'))))
|
|
|
|
|
258 |
|
259 |
H, W, C = input_image.shape
|
260 |
height, width = find_nearest_bucket(H, W, resolution=640)
|
261 |
+
input_image_np = resize_and_center_crop(input_image, target_width=width, target_height=height)
|
|
|
|
|
|
|
|
|
262 |
|
263 |
+
Image.fromarray(input_image_np).save(os.path.join(outputs_folder, f'{job_id}.png'))
|
|
|
|
|
264 |
|
265 |
input_image_pt = torch.from_numpy(input_image_np).float() / 127.5 - 1
|
266 |
input_image_pt = input_image_pt.permute(2, 0, 1)[None, :, None]
|
267 |
|
268 |
# VAE encoding
|
269 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'VAE encoding ...'))))
|
|
|
|
|
270 |
|
271 |
if not high_vram:
|
272 |
load_model_as_complete(vae, target_device=gpu)
|
|
|
273 |
start_latent = vae_encode(input_image_pt, vae)
|
274 |
|
275 |
# CLIP Vision
|
276 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'CLIP Vision encoding ...'))))
|
|
|
|
|
277 |
|
278 |
if not high_vram:
|
279 |
load_model_as_complete(image_encoder, target_device=gpu)
|
280 |
+
image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
|
|
|
|
|
|
|
281 |
image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
|
282 |
|
283 |
+
# Convert dtype
|
284 |
llama_vec = llama_vec.to(transformer.dtype)
|
285 |
llama_vec_n = llama_vec_n.to(transformer.dtype)
|
286 |
clip_l_pooler = clip_l_pooler.to(transformer.dtype)
|
|
|
288 |
image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
|
289 |
|
290 |
# Start sampling
|
291 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Start sampling ...'))))
|
|
|
|
|
292 |
|
293 |
rnd = torch.Generator("cpu").manual_seed(seed)
|
294 |
|
|
|
298 |
).cpu()
|
299 |
history_pixels = None
|
300 |
|
301 |
+
# Add start_latent
|
302 |
+
history_latents = torch.cat([history_latents, start_latent.to(history_latents)], dim=2)
|
|
|
|
|
303 |
total_generated_latent_frames = 1
|
304 |
|
305 |
for section_index in range(total_latent_sections):
|
|
|
325 |
preview = d['denoised']
|
326 |
preview = vae_decode_fake(preview)
|
327 |
preview = (preview * 255.0).detach().cpu().numpy().clip(0, 255).astype(np.uint8)
|
328 |
+
preview = einops.rearrange(preview, 'b c t h w -> (b h) (t w) c')
|
|
|
|
|
|
|
329 |
|
330 |
if stream.input_queue.top() == 'end':
|
331 |
stream.output_queue.push(('end', None))
|
|
|
334 |
current_step = d['i'] + 1
|
335 |
percentage = int(100.0 * current_step / steps)
|
336 |
hint = f'Sampling {current_step}/{steps}'
|
337 |
+
desc = f'Total generated frames: {int(max(0, total_generated_latent_frames * 4 - 3))}'
|
338 |
+
stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint))))
|
|
|
|
|
339 |
return
|
340 |
|
341 |
indices = torch.arange(
|
342 |
+
0, sum([1, 16, 2, 1, latent_window_size])
|
|
|
343 |
).unsqueeze(0)
|
344 |
(
|
345 |
clean_latent_indices_start,
|
|
|
348 |
clean_latent_1x_indices,
|
349 |
latent_indices
|
350 |
) = indices.split([1, 16, 2, 1, latent_window_size], dim=1)
|
351 |
+
|
352 |
+
clean_latent_indices = torch.cat([clean_latent_indices_start, clean_latent_1x_indices], dim=1)
|
|
|
|
|
353 |
|
354 |
clean_latents_4x, clean_latents_2x, clean_latents_1x = history_latents[
|
355 |
:, :, -sum([16, 2, 1]):, :, :
|
356 |
].split([16, 2, 1], dim=2)
|
357 |
+
|
358 |
clean_latents = torch.cat(
|
359 |
[start_latent.to(history_latents), clean_latents_1x],
|
360 |
dim=2
|
|
|
391 |
)
|
392 |
|
393 |
total_generated_latent_frames += int(generated_latents.shape[2])
|
394 |
+
history_latents = torch.cat([history_latents, generated_latents.to(history_latents)], dim=2)
|
|
|
|
|
|
|
395 |
|
396 |
if not high_vram:
|
397 |
+
offload_model_from_device_for_memory_preservation(transformer, target_device=gpu, preserved_memory_gb=8)
|
|
|
|
|
|
|
398 |
load_model_as_complete(vae, target_device=gpu)
|
399 |
|
400 |
+
real_history_latents = history_latents[:, :, -total_generated_latent_frames:, :, :]
|
|
|
|
|
401 |
|
402 |
if history_pixels is None:
|
403 |
history_pixels = vae_decode(real_history_latents, vae).cpu()
|
|
|
415 |
if not high_vram:
|
416 |
unload_complete_models()
|
417 |
|
418 |
+
output_filename = os.path.join(outputs_folder, f'{job_id}_{total_generated_latent_frames}.mp4')
|
419 |
+
save_bcthw_as_mp4(history_pixels, output_filename, fps=30, crf=mp4_crf)
|
|
|
|
|
|
|
|
|
|
|
|
|
420 |
|
421 |
+
print(f'Decoded. Latent shape {real_history_latents.shape}; pixel shape {history_pixels.shape}')
|
|
|
|
|
422 |
|
423 |
stream.output_queue.push(('file', output_filename))
|
424 |
|
425 |
except:
|
426 |
traceback.print_exc()
|
427 |
if not high_vram:
|
428 |
+
unload_complete_models(text_encoder, text_encoder_2, image_encoder, vae, transformer)
|
|
|
|
|
429 |
|
430 |
stream.output_queue.push(('end', None))
|
431 |
return
|
432 |
|
433 |
def get_duration(
|
434 |
+
input_image, prompt, t2v, n_prompt,
|
435 |
+
seed, total_second_length, latent_window_size,
|
436 |
+
steps, cfg, gs, rs, gpu_memory_preservation,
|
437 |
+
use_teacache, mp4_crf
|
438 |
):
|
439 |
return total_second_length * 60
|
440 |
|
441 |
@spaces.GPU(duration=get_duration)
|
442 |
def process(
|
443 |
+
input_image, prompt, t2v=False, n_prompt="", seed=31337,
|
444 |
+
total_second_length=5, latent_window_size=9, steps=25,
|
445 |
+
cfg=1.0, gs=10.0, rs=0.0, gpu_memory_preservation=6,
|
446 |
+
use_teacache=True, mp4_crf=16
|
447 |
):
|
448 |
global stream
|
|
|
449 |
if t2v:
|
450 |
default_height, default_width = 640, 640
|
451 |
+
input_image = np.ones((default_height, default_width, 3), dtype=np.uint8) * 255
|
|
|
|
|
|
|
452 |
print("No input image provided. Using a blank white image.")
|
453 |
else:
|
|
|
454 |
composite_rgba_uint8 = input_image["composite"]
|
455 |
|
|
|
456 |
rgb_uint8 = composite_rgba_uint8[:, :, :3]
|
|
|
457 |
mask_uint8 = composite_rgba_uint8[:, :, 3]
|
458 |
|
|
|
459 |
h, w = rgb_uint8.shape[:2]
|
460 |
background_uint8 = np.full((h, w, 3), 255, dtype=np.uint8)
|
461 |
|
|
|
462 |
alpha_normalized_float32 = mask_uint8.astype(np.float32) / 255.0
|
463 |
alpha_mask_float32 = np.stack([alpha_normalized_float32]*3, axis=2)
|
464 |
|
465 |
+
blended_image_float32 = rgb_uint8.astype(np.float32) * alpha_mask_float32 + \
|
466 |
+
background_uint8.astype(np.float32) * (1.0 - alpha_mask_float32)
|
|
|
|
|
467 |
|
468 |
input_image = np.clip(blended_image_float32, 0, 255).astype(np.uint8)
|
469 |
|
|
|
498 |
yield (
|
499 |
gr.update(),
|
500 |
gr.update(visible=True, value=preview),
|
501 |
+
desc,
|
502 |
+
html,
|
503 |
gr.update(interactive=False),
|
504 |
gr.update(interactive=True)
|
505 |
)
|
|
|
518 |
def end_process():
|
519 |
stream.input_queue.push('end')
|
520 |
|
521 |
+
|
522 |
quick_prompts = [
|
523 |
'The girl dances gracefully, with clear movements, full of charm.',
|
524 |
'A character doing some simple body movements.'
|
525 |
]
|
526 |
quick_prompts = [[x] for x in quick_prompts]
|
527 |
|
528 |
+
|
529 |
def make_custom_css():
|
530 |
base_progress_css = make_progress_bar_css()
|
|
|
531 |
extra_css = """
|
532 |
body {
|
533 |
background: #fafbfe !important;
|
|
|
535 |
}
|
536 |
#title-container {
|
537 |
text-align: center;
|
538 |
+
padding: 20px 0;
|
539 |
background: linear-gradient(135deg, #a8c0ff 0%, #fbc2eb 100%);
|
540 |
+
border-radius: 0 0 10px 10px;
|
541 |
margin-bottom: 20px;
|
542 |
}
|
543 |
#title-container h1 {
|
544 |
color: white;
|
545 |
+
font-size: 2rem;
|
546 |
margin: 0;
|
547 |
font-weight: 800;
|
548 |
text-shadow: 1px 2px 2px rgba(0,0,0,0.1);
|
|
|
590 |
|
591 |
block = gr.Blocks(css=css).queue()
|
592 |
with block:
|
593 |
+
# Title (use gr.Group instead of gr.Box for older Gradio versions)
|
594 |
+
with gr.Group(elem_id="title-container"):
|
595 |
gr.Markdown("<h1>FramePack I2V</h1>")
|
596 |
|
|
|
597 |
gr.Markdown("""
|
598 |
### Video diffusion, but feels like image diffusion
|
599 |
+
FramePack I2V - a model that predicts future frames from past frames,
|
600 |
+
letting you generate short animations from a single image plus text prompt.
|
|
|
601 |
""")
|
602 |
|
603 |
with gr.Row():
|
604 |
with gr.Column():
|
605 |
input_image = gr.ImageEditor(
|
606 |
type="numpy",
|
607 |
+
label="Image Editor (use Brush for mask)",
|
608 |
height=320,
|
609 |
brush=gr.Brush(colors=["#ffffff"])
|
610 |
)
|
611 |
prompt = gr.Textbox(label="Prompt", value='')
|
612 |
+
t2v = gr.Checkbox(label="Only Text to Video (ignore image)?", value=False)
|
613 |
|
|
|
|
|
|
|
|
|
614 |
example_quick_prompts = gr.Dataset(
|
615 |
samples=quick_prompts,
|
616 |
+
label="Quick Prompts",
|
617 |
samples_per_page=1000,
|
618 |
components=[prompt]
|
619 |
)
|
|
|
630 |
end_button = gr.Button(value="Stop Generation", elem_id="stop-button", interactive=False)
|
631 |
|
632 |
total_second_length = gr.Slider(
|
633 |
+
label="Total Video Length (Seconds)",
|
634 |
minimum=1,
|
635 |
maximum=5,
|
636 |
value=2,
|
|
|
642 |
use_teacache = gr.Checkbox(
|
643 |
label='Use TeaCache',
|
644 |
value=True,
|
645 |
+
info='Faster speed, but may worsen hands/fingers.'
|
646 |
)
|
647 |
n_prompt = gr.Textbox(label="Negative Prompt", value="", visible=False)
|
648 |
seed = gr.Number(label="Seed", value=31337, precision=0)
|
|
|
649 |
latent_window_size = gr.Slider(
|
650 |
label="Latent Window Size",
|
651 |
+
minimum=1, maximum=33,
|
652 |
+
value=9, step=1,
|
|
|
|
|
653 |
visible=False
|
654 |
)
|
655 |
steps = gr.Slider(
|
656 |
label="Steps",
|
657 |
+
minimum=1, maximum=100,
|
658 |
+
value=25, step=1,
|
659 |
+
info='Not recommended to change drastically.'
|
|
|
|
|
660 |
)
|
661 |
cfg = gr.Slider(
|
662 |
label="CFG Scale",
|
663 |
+
minimum=1.0, maximum=32.0,
|
664 |
+
value=1.0, step=0.01,
|
|
|
|
|
665 |
visible=False
|
666 |
)
|
667 |
gs = gr.Slider(
|
668 |
label="Distilled CFG Scale",
|
669 |
+
minimum=1.0, maximum=32.0,
|
670 |
+
value=10.0, step=0.01,
|
671 |
+
info='Not recommended to change drastically.'
|
|
|
|
|
672 |
)
|
673 |
rs = gr.Slider(
|
674 |
label="CFG Re-Scale",
|
675 |
+
minimum=0.0, maximum=1.0,
|
676 |
+
value=0.0, step=0.01,
|
|
|
|
|
677 |
visible=False
|
678 |
)
|
679 |
gpu_memory_preservation = gr.Slider(
|
680 |
label="GPU Memory Preservation (GB)",
|
681 |
+
minimum=6, maximum=128,
|
682 |
+
value=6, step=0.1,
|
683 |
+
info="Increase if OOM occurs, but slower."
|
|
|
|
|
684 |
)
|
685 |
mp4_crf = gr.Slider(
|
686 |
label="MP4 Compression (CRF)",
|
687 |
+
minimum=0, maximum=100,
|
688 |
+
value=16, step=1,
|
689 |
+
info="Lower = better quality. 16 recommended."
|
|
|
|
|
690 |
)
|
691 |
|
692 |
with gr.Column():
|
693 |
+
preview_image = gr.Image(
|
694 |
+
label="Preview Latents",
|
695 |
+
height=200,
|
696 |
+
visible=False
|
697 |
+
)
|
698 |
+
result_video = gr.Video(
|
699 |
+
label="Finished Frames",
|
700 |
+
autoplay=True,
|
701 |
+
show_share_button=False,
|
702 |
+
height=512,
|
703 |
+
loop=True
|
704 |
+
)
|
705 |
progress_desc = gr.Markdown('', elem_classes='no-generating-animation')
|
706 |
progress_bar = gr.HTML('', elem_classes='no-generating-animation')
|
707 |
|
708 |
+
# Extra info
|
709 |
gr.HTML("""
|
710 |
<div style="text-align:center; margin-top:20px;">
|
711 |
+
Share your outputs or get inspired by searching
|
712 |
+
<a href="https://x.com/search?q=framepack&f=live" target="_blank">#framepack</a> on Twitter!
|
713 |
</div>
|
714 |
""")
|
715 |
|
|
|
716 |
ips = [
|
717 |
input_image, prompt, t2v, n_prompt, seed,
|
718 |
+
total_second_length, latent_window_size,
|
719 |
+
steps, cfg, gs, rs, gpu_memory_preservation,
|
720 |
use_teacache, mp4_crf
|
721 |
]
|
722 |
start_button.click(
|
|
|
726 |
)
|
727 |
end_button.click(fn=end_process)
|
728 |
|
729 |
+
# If you want examples, uncomment below:
|
730 |
# gr.Examples(
|
731 |
# examples=examples,
|
732 |
# inputs=[input_image, prompt],
|