Update app.py
Browse files
app.py
CHANGED
@@ -18,153 +18,227 @@ import safetensors.torch as sf
|
|
18 |
import numpy as np
|
19 |
import math
|
20 |
|
21 |
-
# Hugging Face Space
|
22 |
IN_HF_SPACE = os.environ.get('SPACE_ID') is not None
|
23 |
|
24 |
-
# GPU
|
25 |
GPU_AVAILABLE = False
|
26 |
GPU_INITIALIZED = False
|
27 |
last_update_time = time.time()
|
28 |
|
29 |
-
#
|
30 |
if IN_HF_SPACE:
|
31 |
try:
|
32 |
import spaces
|
33 |
-
print("Hugging Face Space
|
34 |
|
35 |
-
# GPU利用可能性をチェック
|
36 |
try:
|
37 |
GPU_AVAILABLE = torch.cuda.is_available()
|
38 |
-
print(f"GPU
|
39 |
if GPU_AVAILABLE:
|
40 |
-
print(f"GPU
|
41 |
-
print(f"GPU
|
42 |
|
43 |
-
#
|
44 |
-
test_tensor = torch.zeros(1, device='cuda')
|
45 |
-
test_tensor = test_tensor + 1
|
46 |
del test_tensor
|
47 |
-
print("GPU
|
48 |
else:
|
49 |
-
print("
|
50 |
except Exception as e:
|
51 |
GPU_AVAILABLE = False
|
52 |
-
print(f"GPU
|
53 |
-
print("CPU
|
54 |
except ImportError:
|
55 |
-
print("spaces
|
56 |
GPU_AVAILABLE = torch.cuda.is_available()
|
57 |
|
58 |
from PIL import Image
|
59 |
from diffusers import AutoencoderKLHunyuanVideo
|
60 |
-
from transformers import
|
61 |
-
|
62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
from diffusers_helper.models.hunyuan_video_packed import HunyuanVideoTransformer3DModelPacked
|
64 |
from diffusers_helper.pipelines.k_diffusion_hunyuan import sample_hunyuan
|
65 |
-
from diffusers_helper.memory import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
from diffusers_helper.thread_utils import AsyncStream, async_run
|
67 |
from diffusers_helper.gradio.progress_bar import make_progress_bar_css, make_progress_bar_html
|
68 |
-
from transformers import SiglipImageProcessor, SiglipVisionModel
|
69 |
from diffusers_helper.clip_vision import hf_clip_vision_encode
|
70 |
-
from diffusers_helper.bucket_tools import find_nearest_bucket
|
71 |
|
72 |
outputs_folder = './outputs/'
|
73 |
os.makedirs(outputs_folder, exist_ok=True)
|
74 |
|
75 |
-
#
|
76 |
if not IN_HF_SPACE:
|
77 |
-
# 非Spaces環境でのみCUDAメモリを取得
|
78 |
try:
|
79 |
if torch.cuda.is_available():
|
80 |
free_mem_gb = get_cuda_free_memory_gb(gpu)
|
81 |
-
print(f'
|
82 |
else:
|
83 |
-
free_mem_gb = 6.0
|
84 |
-
print("CUDA
|
85 |
except Exception as e:
|
86 |
-
free_mem_gb = 6.0
|
87 |
-
print(f"CUDA
|
88 |
-
|
89 |
high_vram = free_mem_gb > 60
|
90 |
-
print(f'
|
91 |
else:
|
92 |
-
|
93 |
-
print("Spaces環境でデフォルトのメモリ設定を使用します")
|
94 |
try:
|
95 |
if GPU_AVAILABLE:
|
96 |
-
free_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9 * 0.9
|
97 |
-
high_vram = free_mem_gb > 10
|
98 |
else:
|
99 |
-
free_mem_gb = 6.0
|
100 |
high_vram = False
|
101 |
except Exception as e:
|
102 |
-
print(f"GPU
|
103 |
-
free_mem_gb = 6.0
|
104 |
high_vram = False
|
105 |
-
|
106 |
-
print(f'GPUメモリ: {free_mem_gb:.2f} GB, 高VRAMモード: {high_vram}')
|
107 |
|
108 |
-
# modelsグローバル変数でモデル参照を保存
|
109 |
models = {}
|
110 |
-
cpu_fallback_mode = not GPU_AVAILABLE
|
|
|
111 |
|
112 |
-
# モデルロード関数を使用
|
113 |
def load_models():
|
|
|
|
|
|
|
114 |
global models, cpu_fallback_mode, GPU_INITIALIZED
|
115 |
|
116 |
if GPU_INITIALIZED:
|
117 |
-
print("
|
118 |
return models
|
119 |
|
120 |
-
print("
|
121 |
-
|
122 |
try:
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
# メモリ節約のために精度を下げる
|
128 |
dtype = torch.float16 if GPU_AVAILABLE else torch.float32
|
129 |
transformer_dtype = torch.bfloat16 if GPU_AVAILABLE else torch.float32
|
130 |
-
|
131 |
-
print(f"使用デバイス: {device}, モデル精度: {dtype}, Transformer精度: {transformer_dtype}")
|
132 |
-
|
133 |
-
# モデルを読み込み
|
134 |
-
try:
|
135 |
-
text_encoder = LlamaModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder', torch_dtype=dtype).to(model_device)
|
136 |
-
text_encoder_2 = CLIPTextModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder_2', torch_dtype=dtype).to(model_device)
|
137 |
-
tokenizer = LlamaTokenizerFast.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer')
|
138 |
-
tokenizer_2 = CLIPTokenizer.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer_2')
|
139 |
-
vae = AutoencoderKLHunyuanVideo.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='vae', torch_dtype=dtype).to(model_device)
|
140 |
|
141 |
-
|
142 |
-
image_encoder = SiglipVisionModel.from_pretrained("lllyasviel/flux_redux_bfl", subfolder='image_encoder', torch_dtype=dtype).to(model_device)
|
143 |
|
144 |
-
|
145 |
-
|
146 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
147 |
except Exception as e:
|
148 |
-
print(f"
|
149 |
-
print("
|
150 |
-
|
151 |
-
# 精度を下げて再試行
|
152 |
dtype = torch.float32
|
153 |
transformer_dtype = torch.float32
|
154 |
cpu_fallback_mode = True
|
155 |
-
|
156 |
-
text_encoder = LlamaModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder', torch_dtype=dtype).to('cpu')
|
157 |
-
text_encoder_2 = CLIPTextModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='text_encoder_2', torch_dtype=dtype).to('cpu')
|
158 |
-
tokenizer = LlamaTokenizerFast.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer')
|
159 |
-
tokenizer_2 = CLIPTokenizer.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='tokenizer_2')
|
160 |
-
vae = AutoencoderKLHunyuanVideo.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder='vae', torch_dtype=dtype).to('cpu')
|
161 |
|
162 |
-
|
163 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
164 |
|
165 |
-
transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained(
|
166 |
-
|
167 |
-
|
|
|
|
|
|
|
168 |
|
169 |
vae.eval()
|
170 |
text_encoder.eval()
|
@@ -177,9 +251,8 @@ def load_models():
|
|
177 |
vae.enable_tiling()
|
178 |
|
179 |
transformer.high_quality_fp32_output_for_inference = True
|
180 |
-
print(
|
181 |
|
182 |
-
# モデル精度を設定
|
183 |
if not cpu_fallback_mode:
|
184 |
transformer.to(dtype=transformer_dtype)
|
185 |
vae.to(dtype=dtype)
|
@@ -196,7 +269,6 @@ def load_models():
|
|
196 |
if torch.cuda.is_available() and not cpu_fallback_mode:
|
197 |
try:
|
198 |
if not high_vram:
|
199 |
-
# DynamicSwapInstallerはhuggingfaceのenable_sequential_offloadと同じですが3倍高速です
|
200 |
DynamicSwapInstaller.install_model(transformer, device=device)
|
201 |
DynamicSwapInstaller.install_model(text_encoder, device=device)
|
202 |
else:
|
@@ -205,14 +277,13 @@ def load_models():
|
|
205 |
image_encoder.to(device)
|
206 |
vae.to(device)
|
207 |
transformer.to(device)
|
208 |
-
print(f"
|
209 |
except Exception as e:
|
210 |
-
print(f"
|
211 |
-
print("CPU
|
212 |
cpu_fallback_mode = True
|
213 |
-
|
214 |
-
|
215 |
-
models = {
|
216 |
'text_encoder': text_encoder,
|
217 |
'text_encoder_2': text_encoder_2,
|
218 |
'tokenizer': tokenizer,
|
@@ -224,196 +295,168 @@ def load_models():
|
|
224 |
}
|
225 |
|
226 |
GPU_INITIALIZED = True
|
227 |
-
|
|
|
228 |
return models
|
229 |
except Exception as e:
|
230 |
-
print(f"
|
231 |
traceback.print_exc()
|
232 |
-
|
233 |
-
# より詳細なエラー情報を記録
|
234 |
-
error_info = {
|
235 |
-
"error": str(e),
|
236 |
-
"traceback": traceback.format_exc(),
|
237 |
-
"cuda_available": torch.cuda.is_available(),
|
238 |
-
"device": "cpu" if cpu_fallback_mode else "cuda",
|
239 |
-
}
|
240 |
-
|
241 |
-
# トラブルシューティングのためにエラー情報をファイルに保存
|
242 |
-
try:
|
243 |
-
with open(os.path.join(outputs_folder, "error_log.txt"), "w") as f:
|
244 |
-
f.write(str(error_info))
|
245 |
-
except:
|
246 |
-
pass
|
247 |
-
|
248 |
-
# アプリが引き続き実行を試みることができるよう空の辞書を返す
|
249 |
cpu_fallback_mode = True
|
250 |
return {}
|
251 |
|
252 |
|
253 |
-
#
|
254 |
if IN_HF_SPACE and 'spaces' in globals() and GPU_AVAILABLE:
|
255 |
try:
|
256 |
@spaces.GPU
|
257 |
def initialize_models():
|
258 |
-
"""@spaces.GPU装飾子内でモデルを初期化"""
|
259 |
global GPU_INITIALIZED
|
260 |
try:
|
261 |
result = load_models()
|
262 |
GPU_INITIALIZED = True
|
263 |
return result
|
264 |
except Exception as e:
|
265 |
-
print(f"spaces.GPU
|
266 |
-
traceback.print_exc()
|
267 |
global cpu_fallback_mode
|
268 |
cpu_fallback_mode = True
|
269 |
-
# 装飾子を使わずに再試行
|
270 |
return load_models()
|
271 |
except Exception as e:
|
272 |
-
print(f"spaces.GPU
|
273 |
-
# 装飾子がエラーの場合、非装飾子版を直接使用
|
274 |
def initialize_models():
|
275 |
return load_models()
|
|
|
|
|
|
|
276 |
|
277 |
|
278 |
-
# 以下の関数内部でモデルの取得を遅延させる
|
279 |
def get_models():
|
280 |
-
"""
|
281 |
-
global models
|
282 |
-
|
283 |
-
|
284 |
model_loading_key = "__model_loading__"
|
285 |
-
|
286 |
if not models:
|
287 |
-
# モデルが読み込み中かチェック
|
288 |
if model_loading_key in globals():
|
289 |
-
print("
|
290 |
-
# モデル読み込み完了を待機
|
291 |
import time
|
292 |
-
|
293 |
-
while not models and model_loading_key in globals():
|
294 |
time.sleep(0.5)
|
295 |
-
|
296 |
-
|
297 |
-
print("モデル読み込み待機がタイムアウトしました")
|
298 |
break
|
299 |
-
|
300 |
if models:
|
301 |
return models
|
302 |
-
|
303 |
try:
|
304 |
-
# 読み込みフラグを設定
|
305 |
globals()[model_loading_key] = True
|
306 |
-
|
307 |
if IN_HF_SPACE and 'spaces' in globals() and GPU_AVAILABLE and not cpu_fallback_mode:
|
308 |
try:
|
309 |
-
print("@spaces.GPU
|
310 |
-
|
|
|
311 |
except Exception as e:
|
312 |
-
print(f"GPU
|
313 |
-
|
314 |
-
models
|
315 |
else:
|
316 |
-
|
317 |
-
models
|
318 |
except Exception as e:
|
319 |
-
print(f"
|
320 |
-
|
321 |
-
# 空の辞書を確保
|
322 |
-
models = {}
|
323 |
finally:
|
324 |
-
# 成功か失敗にかかわらず、読み込みフラグを削除
|
325 |
if model_loading_key in globals():
|
326 |
del globals()[model_loading_key]
|
327 |
-
|
328 |
return models
|
329 |
|
330 |
|
331 |
-
#
|
332 |
PREDEFINED_RESOLUTIONS = [
|
333 |
(416, 960), (448, 864), (480, 832), (512, 768), (544, 704),
|
334 |
(576, 672), (608, 640), (640, 608), (672, 576), (704, 544),
|
335 |
(768, 512), (832, 480), (864, 448), (960, 416)
|
336 |
]
|
337 |
|
338 |
-
# 最も近いアスペクト比を見つける関数
|
339 |
def find_closest_aspect_ratio(width, height, target_resolutions):
|
340 |
"""
|
341 |
-
|
342 |
-
|
343 |
-
引数:
|
344 |
-
width: 元の画像の幅
|
345 |
-
height: 元の画像の高さ
|
346 |
-
target_resolutions: 目標解像度のリスト(幅, 高さ)のタプル
|
347 |
-
|
348 |
-
戻り値:
|
349 |
-
tuple: 最も近いアスペクト比の (target_width, target_height)
|
350 |
"""
|
351 |
original_aspect = width / height
|
352 |
-
|
353 |
-
# 各目標解像度に対してアスペクト比の差を計算
|
354 |
min_diff = float('inf')
|
355 |
closest_resolution = None
|
356 |
|
357 |
-
for
|
358 |
-
target_aspect =
|
359 |
diff = abs(original_aspect - target_aspect)
|
360 |
-
|
361 |
if diff < min_diff:
|
362 |
min_diff = diff
|
363 |
-
closest_resolution = (
|
364 |
-
|
365 |
return closest_resolution
|
366 |
|
367 |
|
368 |
stream = AsyncStream()
|
369 |
|
370 |
-
|
371 |
@torch.no_grad()
|
372 |
-
def worker(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
373 |
global last_update_time
|
374 |
last_update_time = time.time()
|
375 |
-
|
376 |
-
#
|
377 |
total_second_length = min(total_second_length, 3.0)
|
378 |
|
379 |
-
# モデルを取得
|
380 |
try:
|
381 |
-
|
382 |
-
if not
|
383 |
-
|
384 |
-
print(
|
385 |
-
stream.output_queue.push(('error',
|
386 |
stream.output_queue.push(('end', None))
|
387 |
return
|
388 |
|
389 |
-
text_encoder =
|
390 |
-
text_encoder_2 =
|
391 |
-
tokenizer =
|
392 |
-
tokenizer_2 =
|
393 |
-
vae =
|
394 |
-
feature_extractor =
|
395 |
-
image_encoder =
|
396 |
-
transformer =
|
397 |
except Exception as e:
|
398 |
-
|
399 |
-
print(
|
400 |
traceback.print_exc()
|
401 |
-
stream.output_queue.push(('error',
|
402 |
stream.output_queue.push(('end', None))
|
403 |
return
|
404 |
-
|
405 |
-
|
406 |
-
device
|
407 |
-
|
408 |
-
|
409 |
-
# CPUモードに合わせてパラメータを調整
|
410 |
if cpu_fallback_mode:
|
411 |
-
print("CPU
|
412 |
-
# CPU処理を高速化するために処理サイズを小さくする
|
413 |
latent_window_size = min(latent_window_size, 5)
|
414 |
-
steps = min(steps, 15)
|
415 |
-
total_second_length = min(total_second_length, 2.0)
|
416 |
-
|
417 |
total_latent_sections = (total_second_length * 30) / (latent_window_size * 4)
|
418 |
total_latent_sections = int(max(round(total_latent_sections), 1))
|
419 |
|
@@ -423,22 +466,20 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
423 |
history_latents = None
|
424 |
total_generated_latent_frames = 0
|
425 |
|
426 |
-
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, '
|
427 |
|
428 |
try:
|
429 |
-
# GPUをクリーン
|
430 |
if not high_vram and not cpu_fallback_mode:
|
431 |
try:
|
432 |
unload_complete_models(
|
433 |
text_encoder, text_encoder_2, image_encoder, vae, transformer
|
434 |
)
|
435 |
except Exception as e:
|
436 |
-
print(f"
|
437 |
-
# 処理を中断せずに続行
|
438 |
|
439 |
-
#
|
440 |
last_update_time = time.time()
|
441 |
-
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, '
|
442 |
|
443 |
try:
|
444 |
if not high_vram and not cpu_fallback_mode:
|
@@ -446,7 +487,6 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
446 |
load_model_as_complete(text_encoder_2, target_device=device)
|
447 |
|
448 |
llama_vec, clip_l_pooler = encode_prompt_conds(prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
|
449 |
-
|
450 |
if cfg == 1:
|
451 |
llama_vec_n, clip_l_pooler_n = torch.zeros_like(llama_vec), torch.zeros_like(clip_l_pooler)
|
452 |
else:
|
@@ -455,85 +495,72 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
455 |
llama_vec, llama_attention_mask = crop_or_pad_yield_mask(llama_vec, length=512)
|
456 |
llama_vec_n, llama_attention_mask_n = crop_or_pad_yield_mask(llama_vec_n, length=512)
|
457 |
except Exception as e:
|
458 |
-
|
459 |
-
print(
|
460 |
traceback.print_exc()
|
461 |
-
stream.output_queue.push(('error',
|
462 |
stream.output_queue.push(('end', None))
|
463 |
return
|
464 |
|
465 |
-
#
|
466 |
try:
|
467 |
H, W, C = input_image.shape
|
|
|
468 |
|
469 |
-
#
|
470 |
-
target_width, target_height = find_closest_aspect_ratio(W, H, PREDEFINED_RESOLUTIONS)
|
471 |
-
|
472 |
-
# height, width変数も保持する(元のコードとの互換性のため)
|
473 |
-
width = target_width
|
474 |
-
height = target_height
|
475 |
-
|
476 |
-
# CPUモードの場合、処理サイズを小さくする
|
477 |
if cpu_fallback_mode:
|
478 |
-
scale_factor = min(320 /
|
479 |
-
|
480 |
-
|
481 |
-
# 縮小後の値も更新
|
482 |
-
height = target_height
|
483 |
-
width = target_width
|
484 |
-
|
485 |
-
print(f'元の画像サイズ: {W}x{H}, リサイズ先: {target_width}x{target_height}')
|
486 |
|
487 |
-
|
488 |
-
input_image_np = resize_and_center_crop(input_image, target_width=
|
489 |
Image.fromarray(input_image_np).save(os.path.join(outputs_folder, f'{job_id}.png'))
|
490 |
|
491 |
input_image_pt = torch.from_numpy(input_image_np).float() / 127.5 - 1
|
492 |
input_image_pt = input_image_pt.permute(2, 0, 1)[None, :, None]
|
493 |
except Exception as e:
|
494 |
-
|
495 |
-
print(
|
496 |
traceback.print_exc()
|
497 |
-
stream.output_queue.push(('error',
|
498 |
stream.output_queue.push(('end', None))
|
499 |
return
|
500 |
|
501 |
-
# VAE
|
502 |
last_update_time = time.time()
|
503 |
-
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'VAE
|
504 |
|
505 |
try:
|
506 |
if not high_vram and not cpu_fallback_mode:
|
507 |
load_model_as_complete(vae, target_device=device)
|
508 |
-
|
509 |
start_latent = vae_encode(input_image_pt, vae)
|
510 |
except Exception as e:
|
511 |
-
|
512 |
-
print(
|
513 |
traceback.print_exc()
|
514 |
-
stream.output_queue.push(('error',
|
515 |
stream.output_queue.push(('end', None))
|
516 |
return
|
517 |
|
518 |
# CLIP Vision
|
519 |
last_update_time = time.time()
|
520 |
-
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'CLIP Vision
|
521 |
|
522 |
try:
|
523 |
if not high_vram and not cpu_fallback_mode:
|
524 |
load_model_as_complete(image_encoder, target_device=device)
|
525 |
-
|
526 |
image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
|
527 |
image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
|
528 |
except Exception as e:
|
529 |
-
|
530 |
-
print(
|
531 |
traceback.print_exc()
|
532 |
-
stream.output_queue.push(('error',
|
533 |
stream.output_queue.push(('end', None))
|
534 |
return
|
535 |
|
536 |
-
#
|
537 |
try:
|
538 |
llama_vec = llama_vec.to(transformer.dtype)
|
539 |
llama_vec_n = llama_vec_n.to(transformer.dtype)
|
@@ -541,75 +568,76 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
541 |
clip_l_pooler_n = clip_l_pooler_n.to(transformer.dtype)
|
542 |
image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
|
543 |
except Exception as e:
|
544 |
-
|
545 |
-
print(
|
546 |
traceback.print_exc()
|
547 |
-
stream.output_queue.push(('error',
|
548 |
stream.output_queue.push(('end', None))
|
549 |
return
|
550 |
|
551 |
-
#
|
552 |
last_update_time = time.time()
|
553 |
-
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, '
|
554 |
|
555 |
rnd = torch.Generator("cpu").manual_seed(seed)
|
556 |
num_frames = latent_window_size * 4 - 3
|
557 |
|
558 |
try:
|
559 |
-
history_latents = torch.zeros(
|
|
|
|
|
|
|
560 |
history_pixels = None
|
561 |
total_generated_latent_frames = 0
|
562 |
except Exception as e:
|
563 |
-
|
564 |
-
print(
|
565 |
traceback.print_exc()
|
566 |
-
stream.output_queue.push(('error',
|
567 |
stream.output_queue.push(('end', None))
|
568 |
return
|
569 |
|
570 |
-
latent_paddings = reversed(range(total_latent_sections))
|
571 |
-
|
572 |
if total_latent_sections > 4:
|
573 |
-
|
574 |
-
# total_latent_sections > 4の場合、展開するよりもいくつかの項目を複製する方が
|
575 |
-
# 良い結果になるようです
|
576 |
-
# 比較するために、latent_paddings = list(reversed(range(total_latent_sections)))を
|
577 |
-
# 使用して下記のトリックを削除することもできます
|
578 |
-
latent_paddings = [3] + [2] * (total_latent_sections - 3) + [1, 0]
|
579 |
|
580 |
for latent_padding in latent_paddings:
|
581 |
last_update_time = time.time()
|
582 |
-
is_last_section = latent_padding == 0
|
583 |
latent_padding_size = latent_padding * latent_window_size
|
584 |
|
585 |
if stream.input_queue.top() == 'end':
|
586 |
-
# 終了時に現在の動画を保存することを確認
|
587 |
if history_pixels is not None and total_generated_latent_frames > 0:
|
588 |
try:
|
589 |
-
|
590 |
-
save_bcthw_as_mp4(history_pixels,
|
591 |
-
stream.output_queue.push(('file',
|
592 |
except Exception as e:
|
593 |
-
print(f"
|
594 |
-
|
595 |
stream.output_queue.push(('end', None))
|
596 |
return
|
597 |
|
598 |
-
print(f'latent_padding_size = {latent_padding_size}, is_last_section
|
599 |
|
600 |
try:
|
601 |
indices = torch.arange(0, sum([1, latent_padding_size, latent_window_size, 1, 2, 16])).unsqueeze(0)
|
602 |
-
|
603 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
604 |
|
605 |
clean_latents_pre = start_latent.to(history_latents)
|
606 |
-
|
607 |
-
clean_latents = torch.cat([clean_latents_pre,
|
608 |
except Exception as e:
|
609 |
-
|
610 |
-
print(
|
611 |
traceback.print_exc()
|
612 |
-
# 完全に終了せずに次のイテレーションを試みる
|
613 |
if last_output_filename:
|
614 |
stream.output_queue.push(('file', last_output_filename))
|
615 |
continue
|
@@ -617,17 +645,17 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
617 |
if not high_vram and not cpu_fallback_mode:
|
618 |
try:
|
619 |
unload_complete_models()
|
620 |
-
move_model_to_device_with_memory_preservation(
|
|
|
|
|
621 |
except Exception as e:
|
622 |
-
print(f"transformer
|
623 |
-
# パフォーマンスに影響する可能性はありますが、終了する必要はないので続行
|
624 |
|
625 |
if use_teacache and not cpu_fallback_mode:
|
626 |
try:
|
627 |
transformer.initialize_teacache(enable_teacache=True, num_steps=steps)
|
628 |
except Exception as e:
|
629 |
-
print(f"
|
630 |
-
# teacacheを無効にして続行
|
631 |
transformer.initialize_teacache(enable_teacache=False)
|
632 |
else:
|
633 |
transformer.initialize_teacache(enable_teacache=False)
|
@@ -635,65 +663,39 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
635 |
def callback(d):
|
636 |
global last_update_time
|
637 |
last_update_time = time.time()
|
638 |
-
|
639 |
try:
|
640 |
-
|
641 |
-
|
642 |
-
|
643 |
-
|
644 |
-
|
645 |
-
|
646 |
-
|
647 |
-
|
648 |
-
|
649 |
-
|
650 |
-
|
651 |
-
|
652 |
-
|
653 |
-
|
654 |
-
|
655 |
-
raise KeyboardInterrupt('ユーザーによるタスク停止')
|
656 |
-
except Exception as e:
|
657 |
-
print(f"【デバッグ】コールバック関数: キュー先頭信号チェック中にエラー: {e}")
|
658 |
-
|
659 |
-
preview = d['denoised']
|
660 |
-
preview = vae_decode_fake(preview)
|
661 |
-
|
662 |
-
preview = (preview * 255.0).detach().cpu().numpy().clip(0, 255).astype(np.uint8)
|
663 |
-
preview = einops.rearrange(preview, 'b c t h w -> (b h) (t w) c')
|
664 |
-
|
665 |
-
current_step = d['i'] + 1
|
666 |
-
percentage = int(100.0 * current_step / steps)
|
667 |
-
hint = f'サンプリング中 {current_step}/{steps}'
|
668 |
-
desc = f'総生成フレーム数: {int(max(0, total_generated_latent_frames * 4 - 3))}, 動画長: {max(0, (total_generated_latent_frames * 4 - 3) / 30) :.2f} 秒 (FPS-30). 動画を現在拡張中...'
|
669 |
-
stream.output_queue.push(('progress', (preview, desc, make_progress_bar_html(percentage, hint))))
|
670 |
-
except KeyboardInterrupt as e:
|
671 |
-
# 中断例外をキャッチして再スローし、サンプリング関数に伝播されるようにする
|
672 |
-
print(f"【デバッグ】コールバック関数: KeyboardInterruptをキャッチ: {e}")
|
673 |
-
print("【デバッグ】コールバック関数: 中断例外を再スロー、サンプリング関数に伝播")
|
674 |
raise
|
675 |
-
except Exception as
|
676 |
-
print(f"
|
677 |
-
# サンプリングプロセスを中断しない
|
678 |
-
print(f"【デバッグ】コールバック関数: ステップ {d['i']} 完了")
|
679 |
return
|
680 |
|
681 |
try:
|
682 |
-
|
683 |
-
print(f"サンプリング開始、デバイス: {device}, データ型: {transformer.dtype}, TeaCache使用: {use_teacache and not cpu_fallback_mode}")
|
684 |
-
|
685 |
try:
|
686 |
-
print("【デバッグ】sample_hunyuanサンプリングプロセス開始")
|
687 |
generated_latents = sample_hunyuan(
|
688 |
transformer=transformer,
|
689 |
sampler='unipc',
|
690 |
-
width=
|
691 |
-
height=
|
692 |
frames=num_frames,
|
693 |
real_guidance_scale=cfg,
|
694 |
distilled_guidance_scale=gs,
|
695 |
guidance_rescale=rs,
|
696 |
-
# shift=3.0,
|
697 |
num_inference_steps=steps,
|
698 |
generator=rnd,
|
699 |
prompt_embeds=llama_vec,
|
@@ -708,181 +710,119 @@ def worker(input_image, prompt, n_prompt, seed, total_second_length, latent_wind
|
|
708 |
latent_indices=latent_indices,
|
709 |
clean_latents=clean_latents,
|
710 |
clean_latent_indices=clean_latent_indices,
|
711 |
-
clean_latents_2x=
|
712 |
-
clean_latent_2x_indices=
|
713 |
-
clean_latents_4x=
|
714 |
-
clean_latent_4x_indices=
|
715 |
-
callback=callback
|
716 |
)
|
717 |
-
|
718 |
-
print(f"【デバッグ】サンプリング完了、所要時間: {time.time() - sampling_start_time:.2f}秒")
|
719 |
except KeyboardInterrupt as e:
|
720 |
-
|
721 |
-
print(f"【デバッグ】KeyboardInterruptをキャッチ: {e}")
|
722 |
-
print("【デバッグ】ユーザーによるサンプリングプロセス中断、中断ロジック処理中")
|
723 |
-
|
724 |
-
# 既に生成された動画がある場合、最後に生成された動画を返す
|
725 |
if last_output_filename:
|
726 |
-
print(f"【デバッグ】部分的に生成された動画あり: {last_output_filename}、この動画を返します")
|
727 |
stream.output_queue.push(('file', last_output_filename))
|
728 |
-
|
729 |
else:
|
730 |
-
|
731 |
-
|
732 |
-
|
733 |
-
print(f"【デバッグ】エラーメッセージを送信: {error_msg}")
|
734 |
-
stream.output_queue.push(('error', error_msg))
|
735 |
-
print("【デバッグ】end信号を送信")
|
736 |
stream.output_queue.push(('end', None))
|
737 |
-
print("【デバッグ】中断処理完了、リターン")
|
738 |
return
|
739 |
except Exception as e:
|
740 |
-
print(f"
|
741 |
traceback.print_exc()
|
742 |
-
|
743 |
-
# 既に生成された動画がある場合、最後に生成された動画を返す
|
744 |
if last_output_filename:
|
745 |
stream.output_queue.push(('file', last_output_filename))
|
746 |
-
|
747 |
-
|
748 |
-
error_msg = f"サンプリングプロセス中にエラーが発生しましたが、部分的に生成された動画を返します: {e}"
|
749 |
-
stream.output_queue.push(('error', error_msg))
|
750 |
else:
|
751 |
-
|
752 |
-
|
753 |
-
stream.output_queue.push(('error', error_msg))
|
754 |
-
|
755 |
stream.output_queue.push(('end', None))
|
756 |
return
|
757 |
|
758 |
try:
|
759 |
if is_last_section:
|
760 |
generated_latents = torch.cat([start_latent.to(generated_latents), generated_latents], dim=2)
|
761 |
-
|
762 |
total_generated_latent_frames += int(generated_latents.shape[2])
|
763 |
history_latents = torch.cat([generated_latents.to(history_latents), history_latents], dim=2)
|
764 |
except Exception as e:
|
765 |
-
|
766 |
-
print(
|
767 |
traceback.print_exc()
|
768 |
-
|
769 |
if last_output_filename:
|
770 |
stream.output_queue.push(('file', last_output_filename))
|
771 |
-
stream.output_queue.push(('error',
|
772 |
stream.output_queue.push(('end', None))
|
773 |
return
|
774 |
|
775 |
if not high_vram and not cpu_fallback_mode:
|
776 |
try:
|
777 |
-
offload_model_from_device_for_memory_preservation(
|
|
|
|
|
778 |
load_model_as_complete(vae, target_device=device)
|
779 |
except Exception as e:
|
780 |
-
print(f"
|
781 |
-
# 続行
|
782 |
|
783 |
try:
|
784 |
-
real_history_latents = history_latents[:, :, :total_generated_latent_frames
|
785 |
except Exception as e:
|
786 |
-
|
787 |
-
print(
|
788 |
-
|
789 |
if last_output_filename:
|
790 |
stream.output_queue.push(('file', last_output_filename))
|
791 |
continue
|
792 |
|
793 |
try:
|
794 |
-
vae_start_time = time.time()
|
795 |
-
print(f"VAEデコード開始、潜在変数形状: {real_history_latents.shape}")
|
796 |
-
|
797 |
if history_pixels is None:
|
798 |
history_pixels = vae_decode(real_history_latents, vae).cpu()
|
799 |
else:
|
800 |
section_latent_frames = (latent_window_size * 2 + 1) if is_last_section else (latent_window_size * 2)
|
801 |
overlapped_frames = latent_window_size * 4 - 3
|
802 |
-
|
803 |
current_pixels = vae_decode(real_history_latents[:, :, :section_latent_frames], vae).cpu()
|
804 |
history_pixels = soft_append_bcthw(current_pixels, history_pixels, overlapped_frames)
|
805 |
|
806 |
-
print(f"VAEデコード完了、所要時間: {time.time() - vae_start_time:.2f}秒")
|
807 |
-
|
808 |
-
if not high_vram and not cpu_fallback_mode:
|
809 |
-
try:
|
810 |
-
unload_complete_models()
|
811 |
-
except Exception as e:
|
812 |
-
print(f"モデルのアンロード中にエラーが発生しました: {e}")
|
813 |
-
|
814 |
output_filename = os.path.join(outputs_folder, f'{job_id}_{total_generated_latent_frames}.mp4')
|
815 |
-
|
816 |
-
save_start_time = time.time()
|
817 |
save_bcthw_as_mp4(history_pixels, output_filename, fps=30, crf=18)
|
818 |
-
print(f"動画保存完了、所要時間: {time.time() - save_start_time:.2f}秒")
|
819 |
-
|
820 |
-
print(f'デコード完了。現在の潜在変数形状 {real_history_latents.shape}; ピクセル形状 {history_pixels.shape}')
|
821 |
-
|
822 |
last_output_filename = output_filename
|
823 |
stream.output_queue.push(('file', output_filename))
|
824 |
except Exception as e:
|
825 |
-
print(f"
|
826 |
traceback.print_exc()
|
827 |
-
|
828 |
-
# 既に生成された動画がある場合、最後に生成された動画を返す
|
829 |
if last_output_filename:
|
830 |
stream.output_queue.push(('file', last_output_filename))
|
831 |
-
|
832 |
-
|
833 |
-
error_msg = f"動画のデコードまたは保存中にエラーが発生しました: {e}"
|
834 |
-
stream.output_queue.push(('error', error_msg))
|
835 |
-
|
836 |
-
# 次のイテレーションを試みる
|
837 |
continue
|
838 |
|
839 |
if is_last_section:
|
840 |
break
|
841 |
except Exception as e:
|
842 |
-
print(f"
|
843 |
-
print(f"【デバッグ】エラー詳細:")
|
844 |
traceback.print_exc()
|
845 |
-
|
846 |
-
# 中断型例外かチェック
|
847 |
-
if isinstance(e, KeyboardInterrupt):
|
848 |
-
print("【デバッグ】外部KeyboardInterrupt例外を検出")
|
849 |
-
|
850 |
if not high_vram and not cpu_fallback_mode:
|
851 |
try:
|
852 |
-
print("【デバッグ】リソース解放のためモデルをアンロード")
|
853 |
unload_complete_models(
|
854 |
text_encoder, text_encoder_2, image_encoder, vae, transformer
|
855 |
)
|
856 |
-
|
857 |
-
|
858 |
-
print(f"【デバッグ】モデルのアンロード中にエラー: {unload_error}")
|
859 |
-
pass
|
860 |
-
|
861 |
-
# 既に生成された動画がある場合、最後に生成された動画を返す
|
862 |
if last_output_filename:
|
863 |
-
print(f"【デバッグ】外部例外処理: 生成済み部分動画を返す {last_output_filename}")
|
864 |
stream.output_queue.push(('file', last_output_filename))
|
865 |
-
|
866 |
-
|
867 |
-
|
868 |
-
# エラーメッセージを返す
|
869 |
-
error_msg = f"処理中にエラーが発生しました: {e}"
|
870 |
-
print(f"【デバッグ】外部例外処理: エラーメッセージを送信: {error_msg}")
|
871 |
-
stream.output_queue.push(('error', error_msg))
|
872 |
|
873 |
-
|
874 |
-
print("【デバッグ】ワーカー関数終了、end信号を送信")
|
875 |
stream.output_queue.push(('end', None))
|
876 |
-
return
|
877 |
|
878 |
|
879 |
-
#
|
880 |
if IN_HF_SPACE and 'spaces' in globals():
|
881 |
@spaces.GPU
|
882 |
def process_with_gpu(input_image, prompt, n_prompt, seed, total_second_length, use_teacache):
|
883 |
global stream
|
884 |
-
assert input_image is not None,
|
885 |
|
|
|
886 |
latent_window_size = 9
|
887 |
steps = 25
|
888 |
cfg = 1.0
|
@@ -890,79 +830,60 @@ if IN_HF_SPACE and 'spaces' in globals():
|
|
890 |
rs = 0.0
|
891 |
gpu_memory_preservation = 6
|
892 |
|
893 |
-
|
894 |
-
# UI状態の初期化
|
895 |
yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
|
896 |
-
|
897 |
try:
|
898 |
stream = AsyncStream()
|
899 |
-
|
900 |
-
|
901 |
-
|
|
|
|
|
|
|
902 |
|
903 |
output_filename = None
|
904 |
prev_output_filename = None
|
905 |
error_message = None
|
906 |
|
907 |
-
# ワーカーの出力を継続的にチェック
|
908 |
while True:
|
909 |
try:
|
910 |
flag, data = stream.output_queue.next()
|
911 |
-
|
912 |
if flag == 'file':
|
913 |
output_filename = data
|
914 |
prev_output_filename = output_filename
|
915 |
-
# ファイル成功時にエラー表示をクリア
|
916 |
yield output_filename, gr.update(), gr.update(), '', gr.update(interactive=False), gr.update(interactive=True)
|
917 |
-
|
918 |
-
if flag == 'progress':
|
919 |
preview, desc, html = data
|
920 |
-
# 進捗更新時にエラーメッセージを変更せず、停止ボタンがインタラクティブであることを確認
|
921 |
yield gr.update(), gr.update(visible=True, value=preview), desc, html, gr.update(interactive=False), gr.update(interactive=True)
|
922 |
-
|
923 |
-
if flag == 'error':
|
924 |
error_message = data
|
925 |
-
print(f"
|
926 |
-
|
927 |
-
|
928 |
-
if flag == 'end':
|
929 |
-
# 最後の動画ファイルがある場合、確実に返す
|
930 |
if output_filename is None and prev_output_filename is not None:
|
931 |
output_filename = prev_output_filename
|
932 |
-
|
933 |
-
# エラーメッセージがある場合、わかりやすいエラー表示を作成
|
934 |
if error_message:
|
935 |
yield output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
936 |
else:
|
937 |
-
# 成功時にエラー表示をしない
|
938 |
yield output_filename, gr.update(visible=False), gr.update(), '', gr.update(interactive=True), gr.update(interactive=False)
|
939 |
break
|
940 |
except Exception as e:
|
941 |
-
print(f"
|
942 |
-
|
943 |
-
|
944 |
-
if current_time - last_update_time > 60: # 60秒間更新がない場合、処理がフリーズした可能性
|
945 |
-
print(f"処理がフリーズした可能性があります。{current_time - last_update_time:.1f}秒間更新がありません")
|
946 |
-
|
947 |
-
# 部分的に生成された動画がある場合、それを返す
|
948 |
if prev_output_filename:
|
949 |
yield prev_output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
950 |
else:
|
951 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
952 |
break
|
953 |
-
|
954 |
except Exception as e:
|
955 |
-
print(f"
|
956 |
traceback.print_exc()
|
957 |
-
error_msg = str(e)
|
958 |
-
|
959 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
960 |
|
961 |
process = process_with_gpu
|
962 |
else:
|
963 |
def process(input_image, prompt, n_prompt, seed, total_second_length, use_teacache):
|
964 |
global stream
|
965 |
-
assert input_image is not None,
|
966 |
|
967 |
latent_window_size = 9
|
968 |
steps = 25
|
@@ -971,373 +892,252 @@ else:
|
|
971 |
rs = 0.0
|
972 |
gpu_memory_preservation = 6
|
973 |
|
974 |
-
# UI状態の初期化
|
975 |
yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
|
976 |
-
|
977 |
try:
|
978 |
stream = AsyncStream()
|
979 |
-
|
980 |
-
|
981 |
-
|
|
|
|
|
|
|
982 |
|
983 |
output_filename = None
|
984 |
prev_output_filename = None
|
985 |
error_message = None
|
986 |
|
987 |
-
# ワーカーの出力を継続的にチェック
|
988 |
while True:
|
989 |
try:
|
990 |
flag, data = stream.output_queue.next()
|
991 |
-
|
992 |
if flag == 'file':
|
993 |
output_filename = data
|
994 |
prev_output_filename = output_filename
|
995 |
-
# ファイル成功時にエラー表示をクリア
|
996 |
yield output_filename, gr.update(), gr.update(), '', gr.update(interactive=False), gr.update(interactive=True)
|
997 |
-
|
998 |
-
if flag == 'progress':
|
999 |
preview, desc, html = data
|
1000 |
-
# 進捗更新時にエラーメッセージを変更せず、停止ボタンがインタラクティブであることを確認
|
1001 |
yield gr.update(), gr.update(visible=True, value=preview), desc, html, gr.update(interactive=False), gr.update(interactive=True)
|
1002 |
-
|
1003 |
-
if flag == 'error':
|
1004 |
error_message = data
|
1005 |
-
print(f"
|
1006 |
-
|
1007 |
-
|
1008 |
-
if flag == 'end':
|
1009 |
-
# 最後の動画ファイルがある場合、確実に返す
|
1010 |
if output_filename is None and prev_output_filename is not None:
|
1011 |
output_filename = prev_output_filename
|
1012 |
-
|
1013 |
-
# エラーメッセージがある場合、わかりやすいエラー表示を作成
|
1014 |
if error_message:
|
1015 |
yield output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
1016 |
else:
|
1017 |
-
# 成功時にエラー表示をしない
|
1018 |
yield output_filename, gr.update(visible=False), gr.update(), '', gr.update(interactive=True), gr.update(interactive=False)
|
1019 |
break
|
1020 |
except Exception as e:
|
1021 |
-
print(f"
|
1022 |
-
|
1023 |
-
|
1024 |
-
if current_time - last_update_time > 60: # 60秒間更新がない場合、処理がフリーズした可能性
|
1025 |
-
print(f"処理がフリーズした可能性があります。{current_time - last_update_time:.1f}秒間更新がありません")
|
1026 |
-
|
1027 |
-
# 部分的に生成された動画がある場合、それを返す
|
1028 |
if prev_output_filename:
|
1029 |
yield prev_output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
1030 |
else:
|
1031 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
1032 |
break
|
1033 |
-
|
1034 |
except Exception as e:
|
1035 |
-
print(f"
|
1036 |
traceback.print_exc()
|
1037 |
-
error_msg = str(e)
|
1038 |
-
|
1039 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
1040 |
-
|
1041 |
|
1042 |
def end_process():
|
1043 |
-
"""
|
1044 |
-
|
1045 |
-
|
|
|
|
|
1046 |
if 'stream' in globals() and stream is not None:
|
1047 |
-
# 送信前にキューの状態を確認
|
1048 |
try:
|
1049 |
current_top = stream.input_queue.top()
|
1050 |
-
print(f"
|
1051 |
except Exception as e:
|
1052 |
-
print(f"
|
1053 |
-
|
1054 |
-
# end信号を送信
|
1055 |
try:
|
1056 |
stream.input_queue.push('end')
|
1057 |
-
print("
|
1058 |
-
|
1059 |
-
# 信号が正常に送信されたか確認
|
1060 |
-
try:
|
1061 |
-
current_top_after = stream.input_queue.top()
|
1062 |
-
print(f"【デバッグ】送信後のキュー先頭信号: {current_top_after}")
|
1063 |
-
except Exception as e:
|
1064 |
-
print(f"【デバッグ】送信後のキュー状態確認中にエラー: {e}")
|
1065 |
-
|
1066 |
except Exception as e:
|
1067 |
-
print(f"
|
1068 |
else:
|
1069 |
-
print("
|
1070 |
return None
|
1071 |
|
1072 |
|
1073 |
quick_prompts = [
|
1074 |
-
|
1075 |
]
|
1076 |
-
quick_prompts = [[x] for x in quick_prompts]
|
1077 |
|
1078 |
-
|
1079 |
-
# カスタムCSSを作成し、レスポンシブレイアウトのサポートを追加
|
1080 |
def make_custom_css():
|
1081 |
-
|
1082 |
-
|
1083 |
-
|
1084 |
-
|
|
|
|
|
1085 |
#app-container {
|
1086 |
-
max-width:
|
1087 |
margin: 0 auto;
|
|
|
|
|
1088 |
}
|
1089 |
-
|
1090 |
-
|
1091 |
-
/* ページタイトルのスタイル */
|
1092 |
h1 {
|
1093 |
font-size: 2rem;
|
1094 |
text-align: center;
|
1095 |
margin-bottom: 1rem;
|
|
|
|
|
1096 |
}
|
1097 |
-
|
1098 |
-
/* ボタンのスタイル */
|
1099 |
.start-btn, .stop-btn {
|
1100 |
min-height: 45px;
|
1101 |
font-size: 1rem;
|
|
|
1102 |
}
|
1103 |
-
|
1104 |
-
|
1105 |
-
|
1106 |
-
h1 {
|
1107 |
-
font-size: 1.5rem;
|
1108 |
-
margin-bottom: 0.5rem;
|
1109 |
-
}
|
1110 |
-
|
1111 |
-
/* 単一カラムレイアウト */
|
1112 |
-
.mobile-full-width {
|
1113 |
-
flex-direction: column !important;
|
1114 |
-
}
|
1115 |
-
|
1116 |
-
.mobile-full-width > .gr-block {
|
1117 |
-
min-width: 100% !important;
|
1118 |
-
flex-grow: 1;
|
1119 |
-
}
|
1120 |
-
|
1121 |
-
/* 動画サイズの調整 */
|
1122 |
-
.video-container {
|
1123 |
-
height: auto !important;
|
1124 |
-
}
|
1125 |
-
|
1126 |
-
/* ボタンサイズの調整 */
|
1127 |
-
.button-container button {
|
1128 |
-
min-height: 50px;
|
1129 |
-
font-size: 1rem;
|
1130 |
-
touch-action: manipulation;
|
1131 |
-
}
|
1132 |
-
|
1133 |
-
/* スライダーの調整 */
|
1134 |
-
.slider-container input[type="range"] {
|
1135 |
-
height: 30px;
|
1136 |
-
}
|
1137 |
-
}
|
1138 |
-
|
1139 |
-
/* タブレットデバイスのスタイル */
|
1140 |
-
@media (min-width: 769px) and (max-width: 1024px) {
|
1141 |
-
.tablet-adjust {
|
1142 |
-
width: 48% !important;
|
1143 |
-
}
|
1144 |
}
|
1145 |
-
|
1146 |
-
|
1147 |
-
|
1148 |
-
.dark-mode-text {
|
1149 |
-
color: #f0f0f0;
|
1150 |
-
}
|
1151 |
-
|
1152 |
-
.dark-mode-bg {
|
1153 |
-
background-color: #2a2a2a;
|
1154 |
-
}
|
1155 |
}
|
1156 |
-
|
1157 |
-
|
1158 |
-
button, input, select, textarea {
|
1159 |
-
font-size: 16px; /* iOSでの拡大を防止 */
|
1160 |
-
}
|
1161 |
-
|
1162 |
-
/* タッチ操作の最適化 */
|
1163 |
-
button, .interactive-element {
|
1164 |
-
min-height: 44px;
|
1165 |
-
min-width: 44px;
|
1166 |
}
|
1167 |
-
|
1168 |
-
|
1169 |
-
|
1170 |
-
|
1171 |
-
background-color: #000;
|
1172 |
}
|
1173 |
-
|
1174 |
-
/* プログレスバーのス���イル強化 */
|
1175 |
.progress-container {
|
1176 |
-
margin-top:
|
1177 |
-
margin-bottom:
|
1178 |
}
|
1179 |
-
|
1180 |
-
/* エラーメッセージのスタイル */
|
1181 |
-
#error-message {
|
1182 |
-
color: #ff4444;
|
1183 |
-
font-weight: bold;
|
1184 |
-
padding: 10px;
|
1185 |
-
border-radius: 4px;
|
1186 |
-
margin-top: 10px;
|
1187 |
-
}
|
1188 |
-
|
1189 |
-
/* エラーコンテナの正しい表示 */
|
1190 |
.error-message {
|
1191 |
-
background-color:
|
|
|
|
|
1192 |
padding: 10px;
|
1193 |
border-radius: 4px;
|
1194 |
margin-top: 10px;
|
1195 |
-
border: 1px solid #ffcccc;
|
1196 |
-
}
|
1197 |
-
|
1198 |
-
/* 多言語エラーメッセージの処理 */
|
1199 |
-
.error-msg-en, .error-msg-ja {
|
1200 |
-
font-weight: bold;
|
1201 |
}
|
1202 |
-
|
1203 |
-
/* エラーアイコン */
|
1204 |
.error-icon {
|
1205 |
-
color: #
|
1206 |
-
font-size: 18px;
|
1207 |
margin-right: 8px;
|
1208 |
}
|
1209 |
-
|
1210 |
-
|
1211 |
-
|
1212 |
-
|
1213 |
-
border:
|
1214 |
-
|
1215 |
-
margin: 0;
|
1216 |
}
|
1217 |
-
|
1218 |
-
|
1219 |
-
|
1220 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
1221 |
}
|
1222 |
"""
|
1223 |
-
|
1224 |
-
# CSSを結合
|
1225 |
-
combined_css = progress_bar_css + responsive_css
|
1226 |
-
return combined_css
|
1227 |
-
|
1228 |
|
1229 |
css = make_custom_css()
|
|
|
1230 |
block = gr.Blocks(css=css).queue()
|
1231 |
with block:
|
1232 |
-
gr.HTML("<h1
|
1233 |
|
1234 |
-
# mobile-full-widthクラスを持つレスポンシブ行を使用
|
1235 |
with gr.Row(elem_classes="mobile-full-width"):
|
1236 |
-
with gr.Column(scale=1
|
1237 |
-
# 二言語ラベルを追加 - 画像アップロード
|
1238 |
input_image = gr.Image(
|
1239 |
-
sources='upload',
|
1240 |
-
type="numpy",
|
1241 |
-
label="
|
1242 |
-
elem_id="input-image",
|
1243 |
height=320
|
1244 |
)
|
1245 |
-
|
1246 |
prompt = gr.Textbox(
|
1247 |
-
label="
|
1248 |
-
value='The camera smoothly orbits around the center of the scene
|
1249 |
-
elem_id="prompt-input"
|
1250 |
)
|
1251 |
-
|
1252 |
example_quick_prompts = gr.Dataset(
|
1253 |
-
samples=quick_prompts,
|
1254 |
-
label=
|
1255 |
-
samples_per_page=1000,
|
1256 |
components=[prompt]
|
1257 |
)
|
1258 |
-
example_quick_prompts.click(
|
|
|
|
|
|
|
|
|
|
|
|
|
1259 |
|
1260 |
-
# スタイルと二言語ラベルを追加したボタン
|
1261 |
with gr.Row(elem_classes="button-container"):
|
1262 |
start_button = gr.Button(
|
1263 |
-
value="
|
1264 |
-
elem_classes="start-btn",
|
1265 |
-
elem_id="start-button",
|
1266 |
variant="primary"
|
1267 |
)
|
1268 |
-
|
1269 |
end_button = gr.Button(
|
1270 |
-
value="
|
1271 |
-
elem_classes="stop-btn",
|
1272 |
-
elem_id="stop-button",
|
1273 |
interactive=False
|
1274 |
)
|
1275 |
|
1276 |
-
|
1277 |
-
|
1278 |
-
|
1279 |
-
|
1280 |
-
|
1281 |
-
|
1282 |
-
|
1283 |
-
|
1284 |
-
|
1285 |
-
|
1286 |
-
|
1287 |
-
label="シード値 / Seed",
|
1288 |
-
value=31337,
|
1289 |
-
precision=0
|
1290 |
-
)
|
1291 |
|
1292 |
-
|
1293 |
-
with gr.Group(elem_classes="slider-container"):
|
1294 |
-
total_second_length = gr.Slider(
|
1295 |
-
label="動画の長さ(最大3秒) / Video Length (max 3 seconds)",
|
1296 |
-
minimum=0.5,
|
1297 |
-
maximum=3,
|
1298 |
-
value=1,
|
1299 |
-
step=0.1
|
1300 |
-
)
|
1301 |
-
|
1302 |
-
# 右側のプレビューと結果カラム
|
1303 |
-
with gr.Column(scale=1, elem_classes="mobile-full-width"):
|
1304 |
-
# プレビュー画像
|
1305 |
preview_image = gr.Image(
|
1306 |
-
label="
|
1307 |
-
height=200,
|
1308 |
visible=False,
|
1309 |
elem_classes="preview-container"
|
1310 |
)
|
1311 |
-
|
1312 |
-
# 動画結果コンテナ
|
1313 |
result_video = gr.Video(
|
1314 |
-
label="
|
1315 |
-
autoplay=True,
|
1316 |
-
show_share_button=True, # 共有ボタンを追加
|
1317 |
-
height=512,
|
1318 |
loop=True,
|
1319 |
-
|
1320 |
-
|
|
|
1321 |
)
|
1322 |
-
|
1323 |
-
|
1324 |
-
|
|
|
|
|
1325 |
|
1326 |
-
# 進捗インジケーター
|
1327 |
with gr.Group(elem_classes="progress-container"):
|
1328 |
-
progress_desc = gr.Markdown(''
|
1329 |
-
progress_bar = gr.HTML(''
|
1330 |
-
|
1331 |
-
# エラーメッセージエリア - カスタムエラーメッセージ形式をサポートするHTMLコンポーネントを使用
|
1332 |
error_message = gr.HTML('', elem_id='error-message', visible=True)
|
1333 |
|
1334 |
-
#
|
1335 |
ips = [input_image, prompt, n_prompt, seed, total_second_length, use_teacache]
|
1336 |
-
|
1337 |
-
|
1338 |
-
|
|
|
|
|
1339 |
end_button.click(fn=end_process)
|
1340 |
|
1341 |
-
|
1342 |
-
block.launch()
|
1343 |
-
|
|
|
18 |
import numpy as np
|
19 |
import math
|
20 |
|
21 |
+
# Check if running in Hugging Face Space
|
22 |
IN_HF_SPACE = os.environ.get('SPACE_ID') is not None
|
23 |
|
24 |
+
# Track GPU availability
|
25 |
GPU_AVAILABLE = False
|
26 |
GPU_INITIALIZED = False
|
27 |
last_update_time = time.time()
|
28 |
|
29 |
+
# If running in a HF Space, import spaces
|
30 |
if IN_HF_SPACE:
|
31 |
try:
|
32 |
import spaces
|
33 |
+
print("Running inside a Hugging Face Space, 'spaces' module imported.")
|
34 |
|
|
|
35 |
try:
|
36 |
GPU_AVAILABLE = torch.cuda.is_available()
|
37 |
+
print(f"GPU available: {GPU_AVAILABLE}")
|
38 |
if GPU_AVAILABLE:
|
39 |
+
print(f"GPU device name: {torch.cuda.get_device_name(0)}")
|
40 |
+
print(f"GPU memory: {torch.cuda.get_device_properties(0).total_memory / 1e9} GB")
|
41 |
|
42 |
+
# Small GPU operation test
|
43 |
+
test_tensor = torch.zeros(1, device='cuda') + 1
|
|
|
44 |
del test_tensor
|
45 |
+
print("GPU test operation succeeded.")
|
46 |
else:
|
47 |
+
print("Warning: CUDA says it's available, but no GPU device was detected.")
|
48 |
except Exception as e:
|
49 |
GPU_AVAILABLE = False
|
50 |
+
print(f"Error checking GPU: {e}")
|
51 |
+
print("Falling back to CPU mode.")
|
52 |
except ImportError:
|
53 |
+
print("Could not import 'spaces' module. Possibly not in a HF Space.")
|
54 |
GPU_AVAILABLE = torch.cuda.is_available()
|
55 |
|
56 |
from PIL import Image
|
57 |
from diffusers import AutoencoderKLHunyuanVideo
|
58 |
+
from transformers import (
|
59 |
+
LlamaModel,
|
60 |
+
CLIPTextModel,
|
61 |
+
LlamaTokenizerFast,
|
62 |
+
CLIPTokenizer,
|
63 |
+
SiglipImageProcessor,
|
64 |
+
SiglipVisionModel
|
65 |
+
)
|
66 |
+
from diffusers_helper.hunyuan import (
|
67 |
+
encode_prompt_conds,
|
68 |
+
vae_decode,
|
69 |
+
vae_encode,
|
70 |
+
vae_decode_fake
|
71 |
+
)
|
72 |
+
from diffusers_helper.utils import (
|
73 |
+
save_bcthw_as_mp4,
|
74 |
+
crop_or_pad_yield_mask,
|
75 |
+
soft_append_bcthw,
|
76 |
+
resize_and_center_crop,
|
77 |
+
generate_timestamp
|
78 |
+
)
|
79 |
from diffusers_helper.models.hunyuan_video_packed import HunyuanVideoTransformer3DModelPacked
|
80 |
from diffusers_helper.pipelines.k_diffusion_hunyuan import sample_hunyuan
|
81 |
+
from diffusers_helper.memory import (
|
82 |
+
cpu,
|
83 |
+
gpu,
|
84 |
+
get_cuda_free_memory_gb,
|
85 |
+
move_model_to_device_with_memory_preservation,
|
86 |
+
offload_model_from_device_for_memory_preservation,
|
87 |
+
fake_diffusers_current_device,
|
88 |
+
DynamicSwapInstaller,
|
89 |
+
unload_complete_models,
|
90 |
+
load_model_as_complete
|
91 |
+
)
|
92 |
from diffusers_helper.thread_utils import AsyncStream, async_run
|
93 |
from diffusers_helper.gradio.progress_bar import make_progress_bar_css, make_progress_bar_html
|
|
|
94 |
from diffusers_helper.clip_vision import hf_clip_vision_encode
|
|
|
95 |
|
96 |
outputs_folder = './outputs/'
|
97 |
os.makedirs(outputs_folder, exist_ok=True)
|
98 |
|
99 |
+
# Manage GPU memory if not in HF Space
|
100 |
if not IN_HF_SPACE:
|
|
|
101 |
try:
|
102 |
if torch.cuda.is_available():
|
103 |
free_mem_gb = get_cuda_free_memory_gb(gpu)
|
104 |
+
print(f'Free VRAM: {free_mem_gb} GB')
|
105 |
else:
|
106 |
+
free_mem_gb = 6.0
|
107 |
+
print("CUDA not available, using default memory setting.")
|
108 |
except Exception as e:
|
109 |
+
free_mem_gb = 6.0
|
110 |
+
print(f"Error getting CUDA memory: {e}, using default=6GB")
|
|
|
111 |
high_vram = free_mem_gb > 60
|
112 |
+
print(f'High-VRAM mode: {high_vram}')
|
113 |
else:
|
114 |
+
print("Using default memory settings in a HF Space.")
|
|
|
115 |
try:
|
116 |
if GPU_AVAILABLE:
|
117 |
+
free_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9 * 0.9
|
118 |
+
high_vram = free_mem_gb > 10
|
119 |
else:
|
120 |
+
free_mem_gb = 6.0
|
121 |
high_vram = False
|
122 |
except Exception as e:
|
123 |
+
print(f"Error retrieving GPU memory: {e}")
|
124 |
+
free_mem_gb = 6.0
|
125 |
high_vram = False
|
126 |
+
print(f'GPU mem: {free_mem_gb:.2f} GB, high_vram={high_vram}')
|
|
|
127 |
|
|
|
128 |
models = {}
|
129 |
+
cpu_fallback_mode = not GPU_AVAILABLE
|
130 |
+
|
131 |
|
|
|
132 |
def load_models():
|
133 |
+
"""
|
134 |
+
Load the entire pipeline models (VAE, text encoders, image encoder, transformer).
|
135 |
+
"""
|
136 |
global models, cpu_fallback_mode, GPU_INITIALIZED
|
137 |
|
138 |
if GPU_INITIALIZED:
|
139 |
+
print("Models are already loaded. Skipping duplicate loading.")
|
140 |
return models
|
141 |
|
142 |
+
print("Starting model load...")
|
143 |
+
|
144 |
try:
|
145 |
+
device = 'cuda' if (GPU_AVAILABLE and not cpu_fallback_mode) else 'cpu'
|
146 |
+
model_device = 'cpu'
|
147 |
+
|
|
|
|
|
148 |
dtype = torch.float16 if GPU_AVAILABLE else torch.float32
|
149 |
transformer_dtype = torch.bfloat16 if GPU_AVAILABLE else torch.float32
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
|
151 |
+
print(f"Device: {device}, VAE/encoders dtype={dtype}, transformer dtype={transformer_dtype}")
|
|
|
152 |
|
153 |
+
try:
|
154 |
+
text_encoder = LlamaModel.from_pretrained(
|
155 |
+
"hunyuanvideo-community/HunyuanVideo",
|
156 |
+
subfolder='text_encoder',
|
157 |
+
torch_dtype=dtype
|
158 |
+
).to(model_device)
|
159 |
+
text_encoder_2 = CLIPTextModel.from_pretrained(
|
160 |
+
"hunyuanvideo-community/HunyuanVideo",
|
161 |
+
subfolder='text_encoder_2',
|
162 |
+
torch_dtype=dtype
|
163 |
+
).to(model_device)
|
164 |
+
tokenizer = LlamaTokenizerFast.from_pretrained(
|
165 |
+
"hunyuanvideo-community/HunyuanVideo",
|
166 |
+
subfolder='tokenizer'
|
167 |
+
)
|
168 |
+
tokenizer_2 = CLIPTokenizer.from_pretrained(
|
169 |
+
"hunyuanvideo-community/HunyuanVideo",
|
170 |
+
subfolder='tokenizer_2'
|
171 |
+
)
|
172 |
+
vae = AutoencoderKLHunyuanVideo.from_pretrained(
|
173 |
+
"hunyuanvideo-community/HunyuanVideo",
|
174 |
+
subfolder='vae',
|
175 |
+
torch_dtype=dtype
|
176 |
+
).to(model_device)
|
177 |
+
|
178 |
+
feature_extractor = SiglipImageProcessor.from_pretrained(
|
179 |
+
"lllyasviel/flux_redux_bfl",
|
180 |
+
subfolder='feature_extractor'
|
181 |
+
)
|
182 |
+
image_encoder = SiglipVisionModel.from_pretrained(
|
183 |
+
"lllyasviel/flux_redux_bfl",
|
184 |
+
subfolder='image_encoder',
|
185 |
+
torch_dtype=dtype
|
186 |
+
).to(model_device)
|
187 |
+
|
188 |
+
# Use a custom rotating-landscape model (for example)
|
189 |
+
transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained(
|
190 |
+
"tori29umai/FramePackI2V_HY_rotate_landscape",
|
191 |
+
torch_dtype=transformer_dtype
|
192 |
+
).to(model_device)
|
193 |
+
|
194 |
+
print("All models loaded successfully.")
|
195 |
except Exception as e:
|
196 |
+
print(f"Error loading models: {e}")
|
197 |
+
print("Retry with float32 on CPU.")
|
|
|
|
|
198 |
dtype = torch.float32
|
199 |
transformer_dtype = torch.float32
|
200 |
cpu_fallback_mode = True
|
|
|
|
|
|
|
|
|
|
|
|
|
201 |
|
202 |
+
text_encoder = LlamaModel.from_pretrained(
|
203 |
+
"hunyuanvideo-community/HunyuanVideo",
|
204 |
+
subfolder='text_encoder',
|
205 |
+
torch_dtype=dtype
|
206 |
+
).to('cpu')
|
207 |
+
text_encoder_2 = CLIPTextModel.from_pretrained(
|
208 |
+
"hunyuanvideo-community/HunyuanVideo",
|
209 |
+
subfolder='text_encoder_2',
|
210 |
+
torch_dtype=dtype
|
211 |
+
).to('cpu')
|
212 |
+
tokenizer = LlamaTokenizerFast.from_pretrained(
|
213 |
+
"hunyuanvideo-community/HunyuanVideo",
|
214 |
+
subfolder='tokenizer'
|
215 |
+
)
|
216 |
+
tokenizer_2 = CLIPTokenizer.from_pretrained(
|
217 |
+
"hunyuanvideo-community/HunyuanVideo",
|
218 |
+
subfolder='tokenizer_2'
|
219 |
+
)
|
220 |
+
vae = AutoencoderKLHunyuanVideo.from_pretrained(
|
221 |
+
"hunyuanvideo-community/HunyuanVideo",
|
222 |
+
subfolder='vae',
|
223 |
+
torch_dtype=dtype
|
224 |
+
).to('cpu')
|
225 |
+
|
226 |
+
feature_extractor = SiglipImageProcessor.from_pretrained(
|
227 |
+
"lllyasviel/flux_redux_bfl",
|
228 |
+
subfolder='feature_extractor'
|
229 |
+
)
|
230 |
+
image_encoder = SiglipVisionModel.from_pretrained(
|
231 |
+
"lllyasviel/flux_redux_bfl",
|
232 |
+
subfolder='image_encoder',
|
233 |
+
torch_dtype=dtype
|
234 |
+
).to('cpu')
|
235 |
|
236 |
+
transformer = HunyuanVideoTransformer3DModelPacked.from_pretrained(
|
237 |
+
"tori29umai/FramePackI2V_HY_rotate_landscape",
|
238 |
+
torch_dtype=transformer_dtype
|
239 |
+
).to('cpu')
|
240 |
+
|
241 |
+
print("Models loaded in CPU-only fallback mode.")
|
242 |
|
243 |
vae.eval()
|
244 |
text_encoder.eval()
|
|
|
251 |
vae.enable_tiling()
|
252 |
|
253 |
transformer.high_quality_fp32_output_for_inference = True
|
254 |
+
print("transformer.high_quality_fp32_output_for_inference = True")
|
255 |
|
|
|
256 |
if not cpu_fallback_mode:
|
257 |
transformer.to(dtype=transformer_dtype)
|
258 |
vae.to(dtype=dtype)
|
|
|
269 |
if torch.cuda.is_available() and not cpu_fallback_mode:
|
270 |
try:
|
271 |
if not high_vram:
|
|
|
272 |
DynamicSwapInstaller.install_model(transformer, device=device)
|
273 |
DynamicSwapInstaller.install_model(text_encoder, device=device)
|
274 |
else:
|
|
|
277 |
image_encoder.to(device)
|
278 |
vae.to(device)
|
279 |
transformer.to(device)
|
280 |
+
print(f"Successfully moved models to {device}")
|
281 |
except Exception as e:
|
282 |
+
print(f"Error moving models to {device}: {e}")
|
283 |
+
print("Falling back to CPU.")
|
284 |
cpu_fallback_mode = True
|
285 |
+
|
286 |
+
models_local = {
|
|
|
287 |
'text_encoder': text_encoder,
|
288 |
'text_encoder_2': text_encoder_2,
|
289 |
'tokenizer': tokenizer,
|
|
|
295 |
}
|
296 |
|
297 |
GPU_INITIALIZED = True
|
298 |
+
models.update(models_local)
|
299 |
+
print(f"Model load complete. Mode: {'CPU' if cpu_fallback_mode else 'GPU'}")
|
300 |
return models
|
301 |
except Exception as e:
|
302 |
+
print(f"Unexpected error in load_models(): {e}")
|
303 |
traceback.print_exc()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
304 |
cpu_fallback_mode = True
|
305 |
return {}
|
306 |
|
307 |
|
308 |
+
# Use GPU decorator if in HF Space
|
309 |
if IN_HF_SPACE and 'spaces' in globals() and GPU_AVAILABLE:
|
310 |
try:
|
311 |
@spaces.GPU
|
312 |
def initialize_models():
|
|
|
313 |
global GPU_INITIALIZED
|
314 |
try:
|
315 |
result = load_models()
|
316 |
GPU_INITIALIZED = True
|
317 |
return result
|
318 |
except Exception as e:
|
319 |
+
print(f"Error in @spaces.GPU model init: {e}")
|
|
|
320 |
global cpu_fallback_mode
|
321 |
cpu_fallback_mode = True
|
|
|
322 |
return load_models()
|
323 |
except Exception as e:
|
324 |
+
print(f"Error creating spaces.GPU decorator: {e}")
|
|
|
325 |
def initialize_models():
|
326 |
return load_models()
|
327 |
+
else:
|
328 |
+
def initialize_models():
|
329 |
+
return load_models()
|
330 |
|
331 |
|
|
|
332 |
def get_models():
|
333 |
+
"""
|
334 |
+
Retrieve the global models or load them if not yet loaded.
|
335 |
+
"""
|
336 |
+
global models
|
337 |
model_loading_key = "__model_loading__"
|
338 |
+
|
339 |
if not models:
|
|
|
340 |
if model_loading_key in globals():
|
341 |
+
print("Models are loading. Please wait.")
|
|
|
342 |
import time
|
343 |
+
start_time = time.time()
|
344 |
+
while (not models) and (model_loading_key in globals()):
|
345 |
time.sleep(0.5)
|
346 |
+
if time.time() - start_time > 60:
|
347 |
+
print("Timed out waiting for model load.")
|
|
|
348 |
break
|
|
|
349 |
if models:
|
350 |
return models
|
|
|
351 |
try:
|
|
|
352 |
globals()[model_loading_key] = True
|
|
|
353 |
if IN_HF_SPACE and 'spaces' in globals() and GPU_AVAILABLE and not cpu_fallback_mode:
|
354 |
try:
|
355 |
+
print("Loading models via @spaces.GPU")
|
356 |
+
models_local = initialize_models()
|
357 |
+
models.update(models_local)
|
358 |
except Exception as e:
|
359 |
+
print(f"GPU decorator load error: {e}, fallback to direct load.")
|
360 |
+
models_local = load_models()
|
361 |
+
models.update(models_local)
|
362 |
else:
|
363 |
+
models_local = load_models()
|
364 |
+
models.update(models_local)
|
365 |
except Exception as e:
|
366 |
+
print(f"Unexpected error while loading models: {e}")
|
367 |
+
models.clear()
|
|
|
|
|
368 |
finally:
|
|
|
369 |
if model_loading_key in globals():
|
370 |
del globals()[model_loading_key]
|
|
|
371 |
return models
|
372 |
|
373 |
|
374 |
+
# Predefined resolutions for a rotating-landscape scenario
|
375 |
PREDEFINED_RESOLUTIONS = [
|
376 |
(416, 960), (448, 864), (480, 832), (512, 768), (544, 704),
|
377 |
(576, 672), (608, 640), (640, 608), (672, 576), (704, 544),
|
378 |
(768, 512), (832, 480), (864, 448), (960, 416)
|
379 |
]
|
380 |
|
|
|
381 |
def find_closest_aspect_ratio(width, height, target_resolutions):
|
382 |
"""
|
383 |
+
Find the resolution in 'target_resolutions' whose aspect ratio
|
384 |
+
is closest to the original image aspect ratio (width/height).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
385 |
"""
|
386 |
original_aspect = width / height
|
|
|
|
|
387 |
min_diff = float('inf')
|
388 |
closest_resolution = None
|
389 |
|
390 |
+
for tw, th in target_resolutions:
|
391 |
+
target_aspect = tw / th
|
392 |
diff = abs(original_aspect - target_aspect)
|
|
|
393 |
if diff < min_diff:
|
394 |
min_diff = diff
|
395 |
+
closest_resolution = (tw, th)
|
|
|
396 |
return closest_resolution
|
397 |
|
398 |
|
399 |
stream = AsyncStream()
|
400 |
|
|
|
401 |
@torch.no_grad()
|
402 |
+
def worker(
|
403 |
+
input_image,
|
404 |
+
prompt,
|
405 |
+
n_prompt,
|
406 |
+
seed,
|
407 |
+
total_second_length,
|
408 |
+
latent_window_size,
|
409 |
+
steps,
|
410 |
+
cfg,
|
411 |
+
gs,
|
412 |
+
rs,
|
413 |
+
gpu_memory_preservation,
|
414 |
+
use_teacache
|
415 |
+
):
|
416 |
+
"""
|
417 |
+
Background worker that performs the actual generation.
|
418 |
+
"""
|
419 |
global last_update_time
|
420 |
last_update_time = time.time()
|
421 |
+
|
422 |
+
# For demonstration, limit max length to 3 seconds
|
423 |
total_second_length = min(total_second_length, 3.0)
|
424 |
|
|
|
425 |
try:
|
426 |
+
models_local = get_models()
|
427 |
+
if not models_local:
|
428 |
+
err_msg = "Failed to load models. Check logs for details."
|
429 |
+
print(err_msg)
|
430 |
+
stream.output_queue.push(('error', err_msg))
|
431 |
stream.output_queue.push(('end', None))
|
432 |
return
|
433 |
|
434 |
+
text_encoder = models_local['text_encoder']
|
435 |
+
text_encoder_2 = models_local['text_encoder_2']
|
436 |
+
tokenizer = models_local['tokenizer']
|
437 |
+
tokenizer_2 = models_local['tokenizer_2']
|
438 |
+
vae = models_local['vae']
|
439 |
+
feature_extractor = models_local['feature_extractor']
|
440 |
+
image_encoder = models_local['image_encoder']
|
441 |
+
transformer = models_local['transformer']
|
442 |
except Exception as e:
|
443 |
+
err = f"Error retrieving models: {e}"
|
444 |
+
print(err)
|
445 |
traceback.print_exc()
|
446 |
+
stream.output_queue.push(('error', err))
|
447 |
stream.output_queue.push(('end', None))
|
448 |
return
|
449 |
+
|
450 |
+
device = 'cuda' if (GPU_AVAILABLE and not cpu_fallback_mode) else 'cpu'
|
451 |
+
print(f"Inference device: {device}")
|
452 |
+
|
453 |
+
# Adjust parameters if in CPU fallback
|
|
|
454 |
if cpu_fallback_mode:
|
455 |
+
print("CPU fallback mode: using smaller parameters for performance.")
|
|
|
456 |
latent_window_size = min(latent_window_size, 5)
|
457 |
+
steps = min(steps, 15)
|
458 |
+
total_second_length = min(total_second_length, 2.0)
|
459 |
+
|
460 |
total_latent_sections = (total_second_length * 30) / (latent_window_size * 4)
|
461 |
total_latent_sections = int(max(round(total_latent_sections), 1))
|
462 |
|
|
|
466 |
history_latents = None
|
467 |
total_generated_latent_frames = 0
|
468 |
|
469 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Starting ...'))))
|
470 |
|
471 |
try:
|
|
|
472 |
if not high_vram and not cpu_fallback_mode:
|
473 |
try:
|
474 |
unload_complete_models(
|
475 |
text_encoder, text_encoder_2, image_encoder, vae, transformer
|
476 |
)
|
477 |
except Exception as e:
|
478 |
+
print(f"Error unloading models: {e}")
|
|
|
479 |
|
480 |
+
# Text encode
|
481 |
last_update_time = time.time()
|
482 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Encoding text ...'))))
|
483 |
|
484 |
try:
|
485 |
if not high_vram and not cpu_fallback_mode:
|
|
|
487 |
load_model_as_complete(text_encoder_2, target_device=device)
|
488 |
|
489 |
llama_vec, clip_l_pooler = encode_prompt_conds(prompt, text_encoder, text_encoder_2, tokenizer, tokenizer_2)
|
|
|
490 |
if cfg == 1:
|
491 |
llama_vec_n, clip_l_pooler_n = torch.zeros_like(llama_vec), torch.zeros_like(clip_l_pooler)
|
492 |
else:
|
|
|
495 |
llama_vec, llama_attention_mask = crop_or_pad_yield_mask(llama_vec, length=512)
|
496 |
llama_vec_n, llama_attention_mask_n = crop_or_pad_yield_mask(llama_vec_n, length=512)
|
497 |
except Exception as e:
|
498 |
+
err = f"Text encoding error: {e}"
|
499 |
+
print(err)
|
500 |
traceback.print_exc()
|
501 |
+
stream.output_queue.push(('error', err))
|
502 |
stream.output_queue.push(('end', None))
|
503 |
return
|
504 |
|
505 |
+
# Process input image
|
506 |
try:
|
507 |
H, W, C = input_image.shape
|
508 |
+
target_w, target_h = find_closest_aspect_ratio(W, H, PREDEFINED_RESOLUTIONS)
|
509 |
|
510 |
+
# If CPU fallback, scale down
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
511 |
if cpu_fallback_mode:
|
512 |
+
scale_factor = min(320 / target_h, 320 / target_w)
|
513 |
+
target_h = int(target_h * scale_factor)
|
514 |
+
target_w = int(target_w * scale_factor)
|
|
|
|
|
|
|
|
|
|
|
515 |
|
516 |
+
print(f"Original image: {W}x{H}, resizing to: {target_w}x{target_h}")
|
517 |
+
input_image_np = resize_and_center_crop(input_image, target_width=target_w, target_height=target_h)
|
518 |
Image.fromarray(input_image_np).save(os.path.join(outputs_folder, f'{job_id}.png'))
|
519 |
|
520 |
input_image_pt = torch.from_numpy(input_image_np).float() / 127.5 - 1
|
521 |
input_image_pt = input_image_pt.permute(2, 0, 1)[None, :, None]
|
522 |
except Exception as e:
|
523 |
+
err = f"Image processing error: {e}"
|
524 |
+
print(err)
|
525 |
traceback.print_exc()
|
526 |
+
stream.output_queue.push(('error', err))
|
527 |
stream.output_queue.push(('end', None))
|
528 |
return
|
529 |
|
530 |
+
# VAE encode
|
531 |
last_update_time = time.time()
|
532 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'VAE encoding ...'))))
|
533 |
|
534 |
try:
|
535 |
if not high_vram and not cpu_fallback_mode:
|
536 |
load_model_as_complete(vae, target_device=device)
|
|
|
537 |
start_latent = vae_encode(input_image_pt, vae)
|
538 |
except Exception as e:
|
539 |
+
err = f"VAE encode error: {e}"
|
540 |
+
print(err)
|
541 |
traceback.print_exc()
|
542 |
+
stream.output_queue.push(('error', err))
|
543 |
stream.output_queue.push(('end', None))
|
544 |
return
|
545 |
|
546 |
# CLIP Vision
|
547 |
last_update_time = time.time()
|
548 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'CLIP Vision encoding ...'))))
|
549 |
|
550 |
try:
|
551 |
if not high_vram and not cpu_fallback_mode:
|
552 |
load_model_as_complete(image_encoder, target_device=device)
|
|
|
553 |
image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
|
554 |
image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
|
555 |
except Exception as e:
|
556 |
+
err = f"CLIP Vision encode error: {e}"
|
557 |
+
print(err)
|
558 |
traceback.print_exc()
|
559 |
+
stream.output_queue.push(('error', err))
|
560 |
stream.output_queue.push(('end', None))
|
561 |
return
|
562 |
|
563 |
+
# Convert dtype
|
564 |
try:
|
565 |
llama_vec = llama_vec.to(transformer.dtype)
|
566 |
llama_vec_n = llama_vec_n.to(transformer.dtype)
|
|
|
568 |
clip_l_pooler_n = clip_l_pooler_n.to(transformer.dtype)
|
569 |
image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
|
570 |
except Exception as e:
|
571 |
+
err = f"Data type conversion error: {e}"
|
572 |
+
print(err)
|
573 |
traceback.print_exc()
|
574 |
+
stream.output_queue.push(('error', err))
|
575 |
stream.output_queue.push(('end', None))
|
576 |
return
|
577 |
|
578 |
+
# Sampling
|
579 |
last_update_time = time.time()
|
580 |
+
stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'Starting sampling...'))))
|
581 |
|
582 |
rnd = torch.Generator("cpu").manual_seed(seed)
|
583 |
num_frames = latent_window_size * 4 - 3
|
584 |
|
585 |
try:
|
586 |
+
history_latents = torch.zeros(
|
587 |
+
size=(1, 16, 1 + 2 + 16, target_h // 8, target_w // 8),
|
588 |
+
dtype=torch.float32
|
589 |
+
).cpu()
|
590 |
history_pixels = None
|
591 |
total_generated_latent_frames = 0
|
592 |
except Exception as e:
|
593 |
+
err = f"Error initializing history latents: {e}"
|
594 |
+
print(err)
|
595 |
traceback.print_exc()
|
596 |
+
stream.output_queue.push(('error', err))
|
597 |
stream.output_queue.push(('end', None))
|
598 |
return
|
599 |
|
600 |
+
latent_paddings = list(reversed(range(total_latent_sections)))
|
|
|
601 |
if total_latent_sections > 4:
|
602 |
+
latent_paddings = [3] + [2]*(total_latent_sections - 3) + [1, 0]
|
|
|
|
|
|
|
|
|
|
|
603 |
|
604 |
for latent_padding in latent_paddings:
|
605 |
last_update_time = time.time()
|
606 |
+
is_last_section = (latent_padding == 0)
|
607 |
latent_padding_size = latent_padding * latent_window_size
|
608 |
|
609 |
if stream.input_queue.top() == 'end':
|
|
|
610 |
if history_pixels is not None and total_generated_latent_frames > 0:
|
611 |
try:
|
612 |
+
final_name = os.path.join(outputs_folder, f'{job_id}_final_{total_generated_latent_frames}.mp4')
|
613 |
+
save_bcthw_as_mp4(history_pixels, final_name, fps=30, crf=18)
|
614 |
+
stream.output_queue.push(('file', final_name))
|
615 |
except Exception as e:
|
616 |
+
print(f"Error saving final partial video: {e}")
|
|
|
617 |
stream.output_queue.push(('end', None))
|
618 |
return
|
619 |
|
620 |
+
print(f'latent_padding_size = {latent_padding_size}, is_last_section={is_last_section}')
|
621 |
|
622 |
try:
|
623 |
indices = torch.arange(0, sum([1, latent_padding_size, latent_window_size, 1, 2, 16])).unsqueeze(0)
|
624 |
+
(
|
625 |
+
cidx_pre,
|
626 |
+
blank_indices,
|
627 |
+
latent_indices,
|
628 |
+
cidx_post,
|
629 |
+
cidx_2x,
|
630 |
+
cidx_4x
|
631 |
+
) = indices.split([1, latent_padding_size, latent_window_size, 1, 2, 16], dim=1)
|
632 |
+
clean_latent_indices = torch.cat([cidx_pre, cidx_post], dim=1)
|
633 |
|
634 |
clean_latents_pre = start_latent.to(history_latents)
|
635 |
+
c_latents_post, c_latents_2x, c_latents_4x = history_latents[:, :, :1 + 2 + 16].split([1, 2, 16], dim=2)
|
636 |
+
clean_latents = torch.cat([clean_latents_pre, c_latents_post], dim=2)
|
637 |
except Exception as e:
|
638 |
+
err = f"Error preparing sampling data: {e}"
|
639 |
+
print(err)
|
640 |
traceback.print_exc()
|
|
|
641 |
if last_output_filename:
|
642 |
stream.output_queue.push(('file', last_output_filename))
|
643 |
continue
|
|
|
645 |
if not high_vram and not cpu_fallback_mode:
|
646 |
try:
|
647 |
unload_complete_models()
|
648 |
+
move_model_to_device_with_memory_preservation(
|
649 |
+
transformer, target_device=device, preserved_memory_gb=gpu_memory_preservation
|
650 |
+
)
|
651 |
except Exception as e:
|
652 |
+
print(f"Error moving transformer to GPU: {e}")
|
|
|
653 |
|
654 |
if use_teacache and not cpu_fallback_mode:
|
655 |
try:
|
656 |
transformer.initialize_teacache(enable_teacache=True, num_steps=steps)
|
657 |
except Exception as e:
|
658 |
+
print(f"Error initializing TeaCache: {e}")
|
|
|
659 |
transformer.initialize_teacache(enable_teacache=False)
|
660 |
else:
|
661 |
transformer.initialize_teacache(enable_teacache=False)
|
|
|
663 |
def callback(d):
|
664 |
global last_update_time
|
665 |
last_update_time = time.time()
|
|
|
666 |
try:
|
667 |
+
if stream.input_queue.top() == 'end':
|
668 |
+
stream.output_queue.push(('end', None))
|
669 |
+
raise KeyboardInterrupt('User requested stop.')
|
670 |
+
preview_latents = d['denoised']
|
671 |
+
preview_latents = vae_decode_fake(preview_latents)
|
672 |
+
preview_img = (preview_latents * 255.0).cpu().numpy().clip(0,255).astype(np.uint8)
|
673 |
+
preview_img = einops.rearrange(preview_img, 'b c t h w -> (b h) (t w) c')
|
674 |
+
|
675 |
+
curr_step = d['i'] + 1
|
676 |
+
percentage = int(100.0 * curr_step / steps)
|
677 |
+
hint = f'Sampling {curr_step}/{steps}'
|
678 |
+
desc = f'Generated frames so far: {int(max(0, total_generated_latent_frames * 4 - 3))}'
|
679 |
+
bar_html = make_progress_bar_html(percentage, hint)
|
680 |
+
stream.output_queue.push(('progress', (preview_img, desc, bar_html)))
|
681 |
+
except KeyboardInterrupt:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
682 |
raise
|
683 |
+
except Exception as exc:
|
684 |
+
print(f"Error in sampling callback: {exc}")
|
|
|
|
|
685 |
return
|
686 |
|
687 |
try:
|
688 |
+
print(f"Sampling: device={device}, dtype={transformer.dtype}, teacache={use_teacache}")
|
|
|
|
|
689 |
try:
|
|
|
690 |
generated_latents = sample_hunyuan(
|
691 |
transformer=transformer,
|
692 |
sampler='unipc',
|
693 |
+
width=target_w,
|
694 |
+
height=target_h,
|
695 |
frames=num_frames,
|
696 |
real_guidance_scale=cfg,
|
697 |
distilled_guidance_scale=gs,
|
698 |
guidance_rescale=rs,
|
|
|
699 |
num_inference_steps=steps,
|
700 |
generator=rnd,
|
701 |
prompt_embeds=llama_vec,
|
|
|
710 |
latent_indices=latent_indices,
|
711 |
clean_latents=clean_latents,
|
712 |
clean_latent_indices=clean_latent_indices,
|
713 |
+
clean_latents_2x=c_latents_2x,
|
714 |
+
clean_latent_2x_indices=cidx_2x,
|
715 |
+
clean_latents_4x=c_latents_4x,
|
716 |
+
clean_latent_4x_indices=cidx_4x,
|
717 |
+
callback=callback
|
718 |
)
|
|
|
|
|
719 |
except KeyboardInterrupt as e:
|
720 |
+
print(f"User interrupt: {e}")
|
|
|
|
|
|
|
|
|
721 |
if last_output_filename:
|
|
|
722 |
stream.output_queue.push(('file', last_output_filename))
|
723 |
+
err_msg = "User stopped generation; partial video returned."
|
724 |
else:
|
725 |
+
err_msg = "User stopped generation; no video produced."
|
726 |
+
stream.output_queue.push(('error', err_msg))
|
|
|
|
|
|
|
|
|
727 |
stream.output_queue.push(('end', None))
|
|
|
728 |
return
|
729 |
except Exception as e:
|
730 |
+
print(f"Error during sampling: {e}")
|
731 |
traceback.print_exc()
|
|
|
|
|
732 |
if last_output_filename:
|
733 |
stream.output_queue.push(('file', last_output_filename))
|
734 |
+
err_msg = f"Sampling error; partial video returned: {e}"
|
735 |
+
stream.output_queue.push(('error', err_msg))
|
|
|
|
|
736 |
else:
|
737 |
+
err_msg = f"Sampling error; no video produced: {e}"
|
738 |
+
stream.output_queue.push(('error', err_msg))
|
|
|
|
|
739 |
stream.output_queue.push(('end', None))
|
740 |
return
|
741 |
|
742 |
try:
|
743 |
if is_last_section:
|
744 |
generated_latents = torch.cat([start_latent.to(generated_latents), generated_latents], dim=2)
|
|
|
745 |
total_generated_latent_frames += int(generated_latents.shape[2])
|
746 |
history_latents = torch.cat([generated_latents.to(history_latents), history_latents], dim=2)
|
747 |
except Exception as e:
|
748 |
+
err = f"Error merging latent outputs: {e}"
|
749 |
+
print(err)
|
750 |
traceback.print_exc()
|
|
|
751 |
if last_output_filename:
|
752 |
stream.output_queue.push(('file', last_output_filename))
|
753 |
+
stream.output_queue.push(('error', err))
|
754 |
stream.output_queue.push(('end', None))
|
755 |
return
|
756 |
|
757 |
if not high_vram and not cpu_fallback_mode:
|
758 |
try:
|
759 |
+
offload_model_from_device_for_memory_preservation(
|
760 |
+
transformer, target_device=device, preserved_memory_gb=8
|
761 |
+
)
|
762 |
load_model_as_complete(vae, target_device=device)
|
763 |
except Exception as e:
|
764 |
+
print(f"Error managing model memory: {e}")
|
|
|
765 |
|
766 |
try:
|
767 |
+
real_history_latents = history_latents[:, :, :total_generated_latent_frames]
|
768 |
except Exception as e:
|
769 |
+
err = f"Error slicing latents history: {e}"
|
770 |
+
print(err)
|
|
|
771 |
if last_output_filename:
|
772 |
stream.output_queue.push(('file', last_output_filename))
|
773 |
continue
|
774 |
|
775 |
try:
|
|
|
|
|
|
|
776 |
if history_pixels is None:
|
777 |
history_pixels = vae_decode(real_history_latents, vae).cpu()
|
778 |
else:
|
779 |
section_latent_frames = (latent_window_size * 2 + 1) if is_last_section else (latent_window_size * 2)
|
780 |
overlapped_frames = latent_window_size * 4 - 3
|
|
|
781 |
current_pixels = vae_decode(real_history_latents[:, :, :section_latent_frames], vae).cpu()
|
782 |
history_pixels = soft_append_bcthw(current_pixels, history_pixels, overlapped_frames)
|
783 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
784 |
output_filename = os.path.join(outputs_folder, f'{job_id}_{total_generated_latent_frames}.mp4')
|
|
|
|
|
785 |
save_bcthw_as_mp4(history_pixels, output_filename, fps=30, crf=18)
|
|
|
|
|
|
|
|
|
786 |
last_output_filename = output_filename
|
787 |
stream.output_queue.push(('file', output_filename))
|
788 |
except Exception as e:
|
789 |
+
print(f"Error decoding/saving video: {e}")
|
790 |
traceback.print_exc()
|
|
|
|
|
791 |
if last_output_filename:
|
792 |
stream.output_queue.push(('file', last_output_filename))
|
793 |
+
err = f"Error decoding/saving video: {e}"
|
794 |
+
stream.output_queue.push(('error', err))
|
|
|
|
|
|
|
|
|
795 |
continue
|
796 |
|
797 |
if is_last_section:
|
798 |
break
|
799 |
except Exception as e:
|
800 |
+
print(f"Outer error: {e}, type={type(e)}")
|
|
|
801 |
traceback.print_exc()
|
|
|
|
|
|
|
|
|
|
|
802 |
if not high_vram and not cpu_fallback_mode:
|
803 |
try:
|
|
|
804 |
unload_complete_models(
|
805 |
text_encoder, text_encoder_2, image_encoder, vae, transformer
|
806 |
)
|
807 |
+
except Exception as ue:
|
808 |
+
print(f"Unload error: {ue}")
|
|
|
|
|
|
|
|
|
809 |
if last_output_filename:
|
|
|
810 |
stream.output_queue.push(('file', last_output_filename))
|
811 |
+
err = f"Error in worker: {e}"
|
812 |
+
stream.output_queue.push(('error', err))
|
|
|
|
|
|
|
|
|
|
|
813 |
|
814 |
+
print("Worker finished, pushing end.")
|
|
|
815 |
stream.output_queue.push(('end', None))
|
|
|
816 |
|
817 |
|
818 |
+
# Create a processing function with or without the HF Spaces GPU decorator
|
819 |
if IN_HF_SPACE and 'spaces' in globals():
|
820 |
@spaces.GPU
|
821 |
def process_with_gpu(input_image, prompt, n_prompt, seed, total_second_length, use_teacache):
|
822 |
global stream
|
823 |
+
assert input_image is not None, "No input image provided."
|
824 |
|
825 |
+
# Fix certain parameters for simplicity
|
826 |
latent_window_size = 9
|
827 |
steps = 25
|
828 |
cfg = 1.0
|
|
|
830 |
rs = 0.0
|
831 |
gpu_memory_preservation = 6
|
832 |
|
|
|
|
|
833 |
yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
|
|
|
834 |
try:
|
835 |
stream = AsyncStream()
|
836 |
+
async_run(
|
837 |
+
worker,
|
838 |
+
input_image, prompt, n_prompt, seed,
|
839 |
+
total_second_length, latent_window_size, steps,
|
840 |
+
cfg, gs, rs, gpu_memory_preservation, use_teacache
|
841 |
+
)
|
842 |
|
843 |
output_filename = None
|
844 |
prev_output_filename = None
|
845 |
error_message = None
|
846 |
|
|
|
847 |
while True:
|
848 |
try:
|
849 |
flag, data = stream.output_queue.next()
|
|
|
850 |
if flag == 'file':
|
851 |
output_filename = data
|
852 |
prev_output_filename = output_filename
|
|
|
853 |
yield output_filename, gr.update(), gr.update(), '', gr.update(interactive=False), gr.update(interactive=True)
|
854 |
+
elif flag == 'progress':
|
|
|
855 |
preview, desc, html = data
|
|
|
856 |
yield gr.update(), gr.update(visible=True, value=preview), desc, html, gr.update(interactive=False), gr.update(interactive=True)
|
857 |
+
elif flag == 'error':
|
|
|
858 |
error_message = data
|
859 |
+
print(f"Received error: {error_message}")
|
860 |
+
elif flag == 'end':
|
|
|
|
|
|
|
861 |
if output_filename is None and prev_output_filename is not None:
|
862 |
output_filename = prev_output_filename
|
|
|
|
|
863 |
if error_message:
|
864 |
yield output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
865 |
else:
|
|
|
866 |
yield output_filename, gr.update(visible=False), gr.update(), '', gr.update(interactive=True), gr.update(interactive=False)
|
867 |
break
|
868 |
except Exception as e:
|
869 |
+
print(f"Error processing output: {e}")
|
870 |
+
if (time.time() - last_update_time) > 60:
|
871 |
+
print(f"No updates for {(time.time()-last_update_time):.1f}s, likely hung.")
|
|
|
|
|
|
|
|
|
872 |
if prev_output_filename:
|
873 |
yield prev_output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
874 |
else:
|
875 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
876 |
break
|
|
|
877 |
except Exception as e:
|
878 |
+
print(f"Error starting process: {e}")
|
879 |
traceback.print_exc()
|
|
|
|
|
880 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
881 |
|
882 |
process = process_with_gpu
|
883 |
else:
|
884 |
def process(input_image, prompt, n_prompt, seed, total_second_length, use_teacache):
|
885 |
global stream
|
886 |
+
assert input_image is not None, "No input image provided."
|
887 |
|
888 |
latent_window_size = 9
|
889 |
steps = 25
|
|
|
892 |
rs = 0.0
|
893 |
gpu_memory_preservation = 6
|
894 |
|
|
|
895 |
yield None, None, '', '', gr.update(interactive=False), gr.update(interactive=True)
|
|
|
896 |
try:
|
897 |
stream = AsyncStream()
|
898 |
+
async_run(
|
899 |
+
worker,
|
900 |
+
input_image, prompt, n_prompt, seed,
|
901 |
+
total_second_length, latent_window_size, steps,
|
902 |
+
cfg, gs, rs, gpu_memory_preservation, use_teacache
|
903 |
+
)
|
904 |
|
905 |
output_filename = None
|
906 |
prev_output_filename = None
|
907 |
error_message = None
|
908 |
|
|
|
909 |
while True:
|
910 |
try:
|
911 |
flag, data = stream.output_queue.next()
|
|
|
912 |
if flag == 'file':
|
913 |
output_filename = data
|
914 |
prev_output_filename = output_filename
|
|
|
915 |
yield output_filename, gr.update(), gr.update(), '', gr.update(interactive=False), gr.update(interactive=True)
|
916 |
+
elif flag == 'progress':
|
|
|
917 |
preview, desc, html = data
|
|
|
918 |
yield gr.update(), gr.update(visible=True, value=preview), desc, html, gr.update(interactive=False), gr.update(interactive=True)
|
919 |
+
elif flag == 'error':
|
|
|
920 |
error_message = data
|
921 |
+
print(f"Received error: {error_message}")
|
922 |
+
elif flag == 'end':
|
|
|
|
|
|
|
923 |
if output_filename is None and prev_output_filename is not None:
|
924 |
output_filename = prev_output_filename
|
|
|
|
|
925 |
if error_message:
|
926 |
yield output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
927 |
else:
|
|
|
928 |
yield output_filename, gr.update(visible=False), gr.update(), '', gr.update(interactive=True), gr.update(interactive=False)
|
929 |
break
|
930 |
except Exception as e:
|
931 |
+
print(f"Error processing output: {e}")
|
932 |
+
if (time.time() - last_update_time) > 60:
|
933 |
+
print(f"No updates for {(time.time()-last_update_time):.1f}s, likely hung.")
|
|
|
|
|
|
|
|
|
934 |
if prev_output_filename:
|
935 |
yield prev_output_filename, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
936 |
else:
|
937 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
938 |
break
|
|
|
939 |
except Exception as e:
|
940 |
+
print(f"Error starting process: {e}")
|
941 |
traceback.print_exc()
|
|
|
|
|
942 |
yield None, gr.update(visible=False), gr.update(), gr.update(interactive=True), gr.update(interactive=False)
|
943 |
+
|
944 |
|
945 |
def end_process():
|
946 |
+
"""
|
947 |
+
Stop generation by pushing 'end' signal into the queue.
|
948 |
+
"""
|
949 |
+
print("User clicked the stop button, sending 'end' signal...")
|
950 |
+
global stream
|
951 |
if 'stream' in globals() and stream is not None:
|
|
|
952 |
try:
|
953 |
current_top = stream.input_queue.top()
|
954 |
+
print(f"Queue top signal: {current_top}")
|
955 |
except Exception as e:
|
956 |
+
print(f"Error checking queue status: {e}")
|
|
|
|
|
957 |
try:
|
958 |
stream.input_queue.push('end')
|
959 |
+
print("Successfully pushed 'end' signal.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
960 |
except Exception as e:
|
961 |
+
print(f"Error pushing 'end' signal: {e}")
|
962 |
else:
|
963 |
+
print("Warning: 'stream' is not initialized; cannot stop.")
|
964 |
return None
|
965 |
|
966 |
|
967 |
quick_prompts = [
|
968 |
+
["The camera smoothly orbits around the center of the scene, keeping the center point fixed and always in view"]
|
969 |
]
|
|
|
970 |
|
|
|
|
|
971 |
def make_custom_css():
|
972 |
+
base_progress_css = make_progress_bar_css()
|
973 |
+
enhanced_css = """
|
974 |
+
body {
|
975 |
+
background: #f9fafb !important;
|
976 |
+
font-family: "Noto Sans", sans-serif;
|
977 |
+
}
|
978 |
#app-container {
|
979 |
+
max-width: 1200px;
|
980 |
margin: 0 auto;
|
981 |
+
padding: 1rem;
|
982 |
+
position: relative;
|
983 |
}
|
|
|
|
|
|
|
984 |
h1 {
|
985 |
font-size: 2rem;
|
986 |
text-align: center;
|
987 |
margin-bottom: 1rem;
|
988 |
+
color: #2d3748;
|
989 |
+
font-weight: 700;
|
990 |
}
|
|
|
|
|
991 |
.start-btn, .stop-btn {
|
992 |
min-height: 45px;
|
993 |
font-size: 1rem;
|
994 |
+
font-weight: 600;
|
995 |
}
|
996 |
+
.start-btn {
|
997 |
+
background-color: #3182ce !important;
|
998 |
+
color: #fff !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
999 |
}
|
1000 |
+
.stop-btn {
|
1001 |
+
background-color: #e53e3e !important;
|
1002 |
+
color: #fff !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1003 |
}
|
1004 |
+
.button-container button:hover {
|
1005 |
+
filter: brightness(0.95);
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1006 |
}
|
1007 |
+
.preview-container, .video-container {
|
1008 |
+
border: 1px solid #cbd5e0;
|
1009 |
+
border-radius: 8px;
|
1010 |
+
overflow: hidden;
|
|
|
1011 |
}
|
|
|
|
|
1012 |
.progress-container {
|
1013 |
+
margin-top: 15px;
|
1014 |
+
margin-bottom: 15px;
|
1015 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1016 |
.error-message {
|
1017 |
+
background-color: #fff5f5;
|
1018 |
+
border: 1px solid #fed7d7;
|
1019 |
+
color: #e53e3e;
|
1020 |
padding: 10px;
|
1021 |
border-radius: 4px;
|
1022 |
margin-top: 10px;
|
|
|
|
|
|
|
|
|
|
|
|
|
1023 |
}
|
|
|
|
|
1024 |
.error-icon {
|
1025 |
+
color: #e53e3e;
|
|
|
1026 |
margin-right: 8px;
|
1027 |
}
|
1028 |
+
#error-message {
|
1029 |
+
color: #ff4444;
|
1030 |
+
font-weight: bold;
|
1031 |
+
padding: 10px;
|
1032 |
+
border-radius: 4px;
|
1033 |
+
margin-top: 10px;
|
|
|
1034 |
}
|
1035 |
+
@media (max-width: 768px) {
|
1036 |
+
#app-container {
|
1037 |
+
padding: 0.5rem;
|
1038 |
+
}
|
1039 |
+
.mobile-full-width {
|
1040 |
+
flex-direction: column !important;
|
1041 |
+
}
|
1042 |
+
.mobile-full-width > .gr-block {
|
1043 |
+
width: 100% !important;
|
1044 |
+
}
|
1045 |
}
|
1046 |
"""
|
1047 |
+
return base_progress_css + enhanced_css
|
|
|
|
|
|
|
|
|
1048 |
|
1049 |
css = make_custom_css()
|
1050 |
+
|
1051 |
block = gr.Blocks(css=css).queue()
|
1052 |
with block:
|
1053 |
+
gr.HTML("<h1>FramePack Rotate-Landscape - Generate Rotating Landscape Video</h1>")
|
1054 |
|
|
|
1055 |
with gr.Row(elem_classes="mobile-full-width"):
|
1056 |
+
with gr.Column(scale=1):
|
|
|
1057 |
input_image = gr.Image(
|
1058 |
+
sources='upload',
|
1059 |
+
type="numpy",
|
1060 |
+
label="Upload Image",
|
|
|
1061 |
height=320
|
1062 |
)
|
1063 |
+
|
1064 |
prompt = gr.Textbox(
|
1065 |
+
label="Prompt",
|
1066 |
+
value='The camera smoothly orbits around the center of the scene...',
|
|
|
1067 |
)
|
1068 |
+
|
1069 |
example_quick_prompts = gr.Dataset(
|
1070 |
+
samples=quick_prompts,
|
1071 |
+
label="Quick Prompts",
|
1072 |
+
samples_per_page=1000,
|
1073 |
components=[prompt]
|
1074 |
)
|
1075 |
+
example_quick_prompts.click(
|
1076 |
+
lambda x: x[0],
|
1077 |
+
inputs=[example_quick_prompts],
|
1078 |
+
outputs=prompt,
|
1079 |
+
show_progress=False,
|
1080 |
+
queue=False
|
1081 |
+
)
|
1082 |
|
|
|
1083 |
with gr.Row(elem_classes="button-container"):
|
1084 |
start_button = gr.Button(
|
1085 |
+
value="Generate",
|
1086 |
+
elem_classes="start-btn",
|
|
|
1087 |
variant="primary"
|
1088 |
)
|
|
|
1089 |
end_button = gr.Button(
|
1090 |
+
value="Stop",
|
1091 |
+
elem_classes="stop-btn",
|
|
|
1092 |
interactive=False
|
1093 |
)
|
1094 |
|
1095 |
+
use_teacache = gr.Checkbox(
|
1096 |
+
label="Use TeaCache",
|
1097 |
+
value=True,
|
1098 |
+
info="Faster speed, but possibly worse finger/hand generation."
|
1099 |
+
)
|
1100 |
+
n_prompt = gr.Textbox(label="Negative Prompt", value="", visible=False)
|
1101 |
+
seed = gr.Number(label="Seed", value=31337, precision=0)
|
1102 |
+
total_second_length = gr.Slider(
|
1103 |
+
label="Video length (max 3 seconds)",
|
1104 |
+
minimum=0.5, maximum=3, value=1.0, step=0.1
|
1105 |
+
)
|
|
|
|
|
|
|
|
|
1106 |
|
1107 |
+
with gr.Column(scale=1):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1108 |
preview_image = gr.Image(
|
1109 |
+
label="Preview",
|
1110 |
+
height=200,
|
1111 |
visible=False,
|
1112 |
elem_classes="preview-container"
|
1113 |
)
|
|
|
|
|
1114 |
result_video = gr.Video(
|
1115 |
+
label="Generated Video",
|
1116 |
+
autoplay=True,
|
|
|
|
|
1117 |
loop=True,
|
1118 |
+
show_share_button=True,
|
1119 |
+
height=512,
|
1120 |
+
elem_classes="video-container"
|
1121 |
)
|
1122 |
+
gr.HTML("""
|
1123 |
+
<div>
|
1124 |
+
Note: Due to reversed sampling, ending actions may appear before starting actions. If the start action is missing, please wait for further frames.
|
1125 |
+
</div>
|
1126 |
+
""")
|
1127 |
|
|
|
1128 |
with gr.Group(elem_classes="progress-container"):
|
1129 |
+
progress_desc = gr.Markdown('')
|
1130 |
+
progress_bar = gr.HTML('')
|
1131 |
+
|
|
|
1132 |
error_message = gr.HTML('', elem_id='error-message', visible=True)
|
1133 |
|
1134 |
+
# Inputs
|
1135 |
ips = [input_image, prompt, n_prompt, seed, total_second_length, use_teacache]
|
1136 |
+
start_button.click(
|
1137 |
+
fn=process,
|
1138 |
+
inputs=ips,
|
1139 |
+
outputs=[result_video, preview_image, progress_desc, progress_bar, start_button, end_button]
|
1140 |
+
)
|
1141 |
end_button.click(fn=end_process)
|
1142 |
|
1143 |
+
block.launch()
|
|
|
|