Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -11,525 +11,444 @@ from PIL import Image
|
|
11 |
import random
|
12 |
import logging
|
13 |
import gc
|
14 |
-
import time
|
15 |
-
import hashlib
|
16 |
-
from dataclasses import dataclass
|
17 |
-
from typing import Optional, Tuple
|
18 |
-
from functools import wraps
|
19 |
-
import threading
|
20 |
-
import os
|
21 |
-
|
22 |
-
# GPU ๋ฉ๋ชจ๋ฆฌ ๊ด๋ฆฌ ์ค์
|
23 |
-
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:256' # ๋ ์์ ์ฒญํฌ ์ฌ์ฉ
|
24 |
|
25 |
# ๋ก๊น
์ค์
|
26 |
logging.basicConfig(level=logging.INFO)
|
27 |
logger = logging.getLogger(__name__)
|
28 |
|
29 |
-
# ์ค์
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
max_area: float = 320.0 * 320.0 # Zero GPU์ ์ต์ ํ
|
40 |
-
slider_min_h: int = 128
|
41 |
-
slider_max_h: int = 512 # ๋ ๋ฎ์ ์ต๋๊ฐ
|
42 |
-
slider_min_w: int = 128
|
43 |
-
slider_max_w: int = 512 # ๋ ๋ฎ์ ์ต๋๊ฐ
|
44 |
-
fixed_fps: int = 24
|
45 |
-
min_frames: int = 8
|
46 |
-
max_frames: int = 30 # ๋ ๋ฎ์ ์ต๋ ํ๋ ์ (1.25์ด)
|
47 |
-
default_prompt: str = "make this image move, smooth motion"
|
48 |
-
default_negative_prompt: str = "static, blur"
|
49 |
-
# GPU ๋ฉ๋ชจ๋ฆฌ ์ต์ ํ ์ค์
|
50 |
-
enable_model_cpu_offload: bool = True
|
51 |
-
enable_vae_slicing: bool = True
|
52 |
-
enable_vae_tiling: bool = True
|
53 |
-
|
54 |
-
@property
|
55 |
-
def max_duration(self):
|
56 |
-
"""์ต๋ ํ์ฉ duration (์ด)"""
|
57 |
-
return self.max_frames / self.fixed_fps
|
58 |
-
|
59 |
-
@property
|
60 |
-
def min_duration(self):
|
61 |
-
"""์ต์ ํ์ฉ duration (์ด)"""
|
62 |
-
return self.min_frames / self.fixed_fps
|
63 |
|
64 |
-
|
|
|
65 |
MAX_SEED = np.iinfo(np.int32).max
|
66 |
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
aspect_ratio = orig_h / orig_w
|
103 |
-
|
104 |
-
# Zero GPU์ ์ต์ ํ๋ ๋งค์ฐ ์์ ํด์๋
|
105 |
-
max_area = 320.0 * 320.0 # 102,400 ํฝ์
|
106 |
-
|
107 |
-
# ์ข
ํก๋น๊ฐ ๋๋ฌด ๊ทน๋จ์ ์ธ ๊ฒฝ์ฐ ์กฐ์
|
108 |
-
if aspect_ratio > 2.0:
|
109 |
-
aspect_ratio = 2.0
|
110 |
-
elif aspect_ratio < 0.5:
|
111 |
-
aspect_ratio = 0.5
|
112 |
-
|
113 |
-
calc_h = round(np.sqrt(max_area * aspect_ratio))
|
114 |
-
calc_w = round(np.sqrt(max_area / aspect_ratio))
|
115 |
-
|
116 |
-
# mod_value์ ๋ง์ถค
|
117 |
-
calc_h = max(self.config.mod_value, (calc_h // self.config.mod_value) * self.config.mod_value)
|
118 |
-
calc_w = max(self.config.mod_value, (calc_w // self.config.mod_value) * self.config.mod_value)
|
119 |
-
|
120 |
-
# ์ต๋ 512๋ก ์ ํ
|
121 |
-
new_h = int(np.clip(calc_h, self.config.slider_min_h, 512))
|
122 |
-
new_w = int(np.clip(calc_w, self.config.slider_min_w, 512))
|
123 |
-
|
124 |
-
# mod_value์ ๋ง์ถค
|
125 |
-
new_h = (new_h // self.config.mod_value) * self.config.mod_value
|
126 |
-
new_w = (new_w // self.config.mod_value) * self.config.mod_value
|
127 |
-
|
128 |
-
# ์ต์ข
ํฝ์
์ ํ์ธ
|
129 |
-
if new_h * new_w > 102400: # 320x320
|
130 |
-
# ๋น์จ์ ์ ์งํ๋ฉด์ ์ถ์
|
131 |
-
scale = np.sqrt(102400 / (new_h * new_w))
|
132 |
-
new_h = int((new_h * scale) // self.config.mod_value) * self.config.mod_value
|
133 |
-
new_w = int((new_w * scale) // self.config.mod_value) * self.config.mod_value
|
134 |
-
|
135 |
-
return new_h, new_w
|
136 |
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
# Zero GPU์ ์ต์ ํ๋ ์ ํ
|
149 |
-
if duration < 0.3:
|
150 |
-
return False, "โฑ๏ธ Duration too short (min 0.3s)"
|
151 |
-
|
152 |
-
if duration > 1.2: # ๋ ์งง์ ์ต๋ duration
|
153 |
-
return False, "โฑ๏ธ Duration too long (max 1.2s for stability)"
|
154 |
-
|
155 |
-
# ํฝ์
์ ์ ํ (๋ ๋ณด์์ ์ผ๋ก)
|
156 |
-
max_pixels = 320 * 320 # 102,400 ํฝ์
|
157 |
-
if height * width > max_pixels:
|
158 |
-
return False, f"๐ Total pixels limited to {max_pixels:,} (e.g., 320ร320, 256ร384)"
|
159 |
-
|
160 |
-
if height > 512 or width > 512: # ๋ ๋ฎ์ ์ต๋๊ฐ
|
161 |
-
return False, "๐ Maximum dimension is 512 pixels"
|
162 |
-
|
163 |
-
# ์ข
ํก๋น ์ฒดํฌ
|
164 |
-
aspect_ratio = max(height/width, width/height)
|
165 |
-
if aspect_ratio > 2.0:
|
166 |
-
return False, "๐ Aspect ratio too extreme (max 2:1 or 1:2)"
|
167 |
-
|
168 |
-
if steps > 5: # ๋ ๋ฎ์ ์ต๋ ์คํ
|
169 |
-
return False, "๐ง Maximum 5 steps in Zero GPU environment"
|
170 |
-
|
171 |
-
return True, None
|
172 |
|
173 |
-
|
174 |
-
|
175 |
-
unique_str = f"{timestamp}_{seed}_{random.randint(1000, 9999)}"
|
176 |
-
hash_obj = hashlib.md5(unique_str.encode())
|
177 |
-
return f"video_{hash_obj.hexdigest()[:8]}.mp4"
|
178 |
-
|
179 |
-
video_generator = VideoGenerator(config)
|
180 |
-
|
181 |
-
# Gradio ํจ์๋ค
|
182 |
-
def handle_image_upload(image):
|
183 |
-
if image is None:
|
184 |
-
return gr.update(value=config.default_height), gr.update(value=config.default_width)
|
185 |
|
|
|
|
|
|
|
|
|
|
|
186 |
try:
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
|
|
191 |
return gr.update(value=new_h), gr.update(value=new_w)
|
192 |
-
|
193 |
except Exception as e:
|
194 |
-
|
195 |
-
gr.
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
pixels = height * width
|
205 |
-
if pixels > 147456: # 384x384 ์ด์
|
206 |
-
base_duration += 20
|
207 |
-
elif pixels > 100000: # ~316x316 ์ด์
|
208 |
-
base_duration += 10
|
209 |
-
|
210 |
-
# ์คํ
์์ ๋ฐ๋ฅธ ์ถ๊ฐ ์๊ฐ
|
211 |
-
if steps > 4:
|
212 |
-
base_duration += 15
|
213 |
-
elif steps > 2:
|
214 |
-
base_duration += 10
|
215 |
-
|
216 |
-
# ์ข
ํก๋น๊ฐ ๊ทน๋จ์ ์ธ ๊ฒฝ์ฐ ์ถ๊ฐ ์๊ฐ
|
217 |
-
aspect_ratio = max(height/width, width/height)
|
218 |
-
if aspect_ratio > 1.5: # 3:2 ์ด์์ ๋น์จ
|
219 |
-
base_duration += 10
|
220 |
|
221 |
-
|
222 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
223 |
|
224 |
@spaces.GPU(duration=get_duration)
|
225 |
-
@measure_time
|
226 |
def generate_video(input_image, prompt, height, width,
|
227 |
-
negative_prompt=
|
228 |
-
|
229 |
-
seed=42, randomize_seed=False,
|
230 |
progress=gr.Progress(track_tqdm=True)):
|
231 |
|
232 |
-
|
233 |
-
|
234 |
-
# ๋์ ์คํ ๋ฐฉ์ง
|
235 |
-
if not generation_lock.acquire(blocking=False):
|
236 |
-
raise gr.Error("โณ Another video is being generated. Please wait...")
|
237 |
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
input_image, prompt, height, width, duration_seconds, steps
|
245 |
-
)
|
246 |
-
if not is_valid:
|
247 |
-
logger.warning(f"Validation failed: {error_msg}")
|
248 |
-
raise gr.Error(error_msg)
|
249 |
-
|
250 |
-
# ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
251 |
-
clear_gpu_memory()
|
252 |
-
|
253 |
-
progress(0.1, desc="๐ Loading model...")
|
254 |
-
|
255 |
-
# ๋ชจ๋ธ ๋ก๋ฉ (GPU ํจ์ ๋ด์์)
|
256 |
-
if pipe is None:
|
257 |
-
try:
|
258 |
-
logger.info("Loading model components...")
|
259 |
-
|
260 |
-
# ์ปดํฌ๋ํธ ๋ก๋
|
261 |
-
image_encoder = CLIPVisionModel.from_pretrained(
|
262 |
-
config.model_id,
|
263 |
-
subfolder="image_encoder",
|
264 |
-
torch_dtype=torch.float16,
|
265 |
-
low_cpu_mem_usage=True
|
266 |
-
)
|
267 |
-
|
268 |
-
vae = AutoencoderKLWan.from_pretrained(
|
269 |
-
config.model_id,
|
270 |
-
subfolder="vae",
|
271 |
-
torch_dtype=torch.float16,
|
272 |
-
low_cpu_mem_usage=True
|
273 |
-
)
|
274 |
-
|
275 |
-
pipe = WanImageToVideoPipeline.from_pretrained(
|
276 |
-
config.model_id,
|
277 |
-
vae=vae,
|
278 |
-
image_encoder=image_encoder,
|
279 |
-
torch_dtype=torch.bfloat16,
|
280 |
-
low_cpu_mem_usage=True,
|
281 |
-
use_safetensors=True
|
282 |
-
)
|
283 |
-
|
284 |
-
# ์ค์ผ์ค๋ฌ ์ค์
|
285 |
-
pipe.scheduler = UniPCMultistepScheduler.from_config(
|
286 |
-
pipe.scheduler.config, flow_shift=8.0
|
287 |
-
)
|
288 |
-
|
289 |
-
# LoRA ๋ก๋ ๊ฑด๋๋ฐ๊ธฐ (์์ ์ฑ์ ์ํด)
|
290 |
-
logger.info("Skipping LoRA for stability")
|
291 |
-
|
292 |
-
# GPU๋ก ์ด๋
|
293 |
-
pipe.to("cuda")
|
294 |
-
|
295 |
-
# ์ต์ ํ ํ์ฑํ
|
296 |
-
pipe.enable_vae_slicing()
|
297 |
-
pipe.enable_vae_tiling()
|
298 |
-
|
299 |
-
# ๋ชจ๋ธ CPU ์คํ๋ก๋ ํ์ฑํ (๋ฉ๋ชจ๋ฆฌ ์ ์ฝ)
|
300 |
-
pipe.enable_model_cpu_offload()
|
301 |
-
|
302 |
-
logger.info("Model loaded successfully")
|
303 |
-
|
304 |
-
except Exception as e:
|
305 |
-
logger.error(f"Model loading failed: {e}")
|
306 |
-
raise gr.Error("Failed to load model")
|
307 |
-
|
308 |
-
progress(0.3, desc="๐ฏ Preparing image...")
|
309 |
-
|
310 |
-
# ์ด๋ฏธ์ง ์ค๋น
|
311 |
-
target_h = max(config.mod_value, (int(height) // config.mod_value) * config.mod_value)
|
312 |
-
target_w = max(config.mod_value, (int(width) // config.mod_value) * config.mod_value)
|
313 |
-
|
314 |
-
# ํ๋ ์ ์ ๊ณ์ฐ (๋งค์ฐ ๋ณด์์ )
|
315 |
-
num_frames = min(
|
316 |
-
int(round(duration_seconds * config.fixed_fps)),
|
317 |
-
24 # ์ต๋ 24ํ๋ ์ (1์ด)
|
318 |
-
)
|
319 |
-
num_frames = max(8, num_frames) # ์ต์ 8ํ๋ ์
|
320 |
-
|
321 |
-
logger.info(f"Generating {num_frames} frames at {target_h}x{target_w}")
|
322 |
-
|
323 |
-
current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
|
324 |
-
|
325 |
-
# ์ด๋ฏธ์ง ๋ฆฌ์ฌ์ด์ฆ
|
326 |
-
resized_image = input_image.resize((target_w, target_h), Image.Resampling.LANCZOS)
|
327 |
-
|
328 |
-
progress(0.4, desc="๐ฌ Generating video...")
|
329 |
-
|
330 |
-
# ๋น๋์ค ์์ฑ
|
331 |
-
with torch.inference_mode(), torch.amp.autocast('cuda', enabled=True, dtype=torch.float16):
|
332 |
-
try:
|
333 |
-
# ๋ฉ๋ชจ๋ฆฌ ํจ์จ์ ์ํ ์ค์
|
334 |
-
torch.cuda.empty_cache()
|
335 |
-
|
336 |
-
# ์์ฑ ํ๋ผ๋ฏธํฐ ์ต์ ํ
|
337 |
-
output_frames_list = pipe(
|
338 |
-
image=resized_image,
|
339 |
-
prompt=prompt[:150], # ํ๋กฌํํธ ๊ธธ์ด ์ ํ
|
340 |
-
negative_prompt=negative_prompt[:50] if negative_prompt else "",
|
341 |
-
height=target_h,
|
342 |
-
width=target_w,
|
343 |
-
num_frames=num_frames,
|
344 |
-
guidance_scale=float(guidance_scale),
|
345 |
-
num_inference_steps=int(steps),
|
346 |
-
generator=torch.Generator(device="cuda").manual_seed(current_seed),
|
347 |
-
return_dict=True,
|
348 |
-
# ์ถ๊ฐ ์ต์ ํ ํ๋ผ๋ฏธํฐ
|
349 |
-
output_type="pil"
|
350 |
-
).frames[0]
|
351 |
-
|
352 |
-
logger.info("Video generation completed successfully")
|
353 |
-
|
354 |
-
except torch.cuda.OutOfMemoryError:
|
355 |
-
logger.error("GPU OOM error")
|
356 |
-
clear_gpu_memory()
|
357 |
-
raise gr.Error("๐พ GPU out of memory. Try smaller dimensions (256x256 recommended).")
|
358 |
-
except RuntimeError as e:
|
359 |
-
if "out of memory" in str(e).lower():
|
360 |
-
logger.error("Runtime OOM error")
|
361 |
-
clear_gpu_memory()
|
362 |
-
raise gr.Error("๐พ GPU memory error. Please try again with smaller settings.")
|
363 |
-
else:
|
364 |
-
logger.error(f"Runtime error: {e}")
|
365 |
-
raise gr.Error(f"โ Generation failed: {str(e)[:50]}")
|
366 |
-
except Exception as e:
|
367 |
-
logger.error(f"Generation error: {type(e).__name__}: {e}")
|
368 |
-
raise gr.Error(f"โ Generation failed. Try reducing resolution or steps.")
|
369 |
-
|
370 |
-
progress(0.9, desc="๐พ Saving video...")
|
371 |
-
|
372 |
-
# ๋น๋์ค ์ ์ฅ
|
373 |
-
try:
|
374 |
-
filename = video_generator.generate_unique_filename(current_seed)
|
375 |
-
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
|
376 |
-
video_path = tmpfile.name
|
377 |
-
|
378 |
-
export_to_video(output_frames_list, video_path, fps=config.fixed_fps)
|
379 |
-
logger.info(f"Video saved: {video_path}")
|
380 |
-
except Exception as e:
|
381 |
-
logger.error(f"Save error: {e}")
|
382 |
-
raise gr.Error("Failed to save video")
|
383 |
|
384 |
-
|
385 |
-
|
|
|
|
|
386 |
|
387 |
-
#
|
388 |
-
|
389 |
-
|
390 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
391 |
gc.collect()
|
392 |
-
|
393 |
-
|
394 |
-
|
395 |
-
except gr.Error:
|
396 |
-
raise
|
397 |
except Exception as e:
|
398 |
-
logger.error(f"
|
399 |
-
raise gr.Error(f"
|
400 |
-
|
401 |
-
|
402 |
-
|
403 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
404 |
|
405 |
-
# CSS
|
406 |
css = """
|
407 |
.container {
|
408 |
-
max-width:
|
409 |
margin: auto;
|
410 |
padding: 20px;
|
411 |
}
|
412 |
|
413 |
.header {
|
414 |
text-align: center;
|
415 |
-
margin-bottom:
|
416 |
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
|
417 |
-
padding:
|
418 |
-
border-radius:
|
419 |
color: white;
|
420 |
-
box-shadow: 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
}
|
422 |
|
423 |
.header h1 {
|
424 |
-
font-size:
|
425 |
margin-bottom: 10px;
|
|
|
|
|
|
|
426 |
}
|
427 |
|
428 |
-
.
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
436 |
}
|
437 |
|
438 |
.generate-btn {
|
439 |
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
|
440 |
color: white;
|
441 |
-
font-size: 1.
|
442 |
-
padding:
|
443 |
-
border-radius:
|
444 |
border: none;
|
445 |
cursor: pointer;
|
|
|
|
|
446 |
width: 100%;
|
447 |
-
margin-top:
|
448 |
}
|
449 |
|
450 |
.generate-btn:hover {
|
451 |
transform: translateY(-2px);
|
452 |
-
box-shadow: 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
453 |
}
|
454 |
"""
|
455 |
|
456 |
-
# Gradio UI
|
457 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
458 |
with gr.Column(elem_classes="container"):
|
459 |
-
# Header
|
460 |
gr.HTML("""
|
461 |
<div class="header">
|
462 |
-
<h1>๐ฌ AI Video
|
463 |
-
<p>Transform images into videos with Wan 2.1
|
|
|
464 |
</div>
|
465 |
""")
|
466 |
|
467 |
-
# ๊ฒฝ๊ณ
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
468 |
gr.HTML("""
|
469 |
-
<div class="
|
470 |
-
<strong
|
471 |
-
<
|
472 |
-
<li>
|
473 |
-
<li>
|
474 |
-
<li>
|
475 |
-
<li>
|
476 |
-
|
477 |
-
</ul>
|
478 |
</div>
|
479 |
""")
|
480 |
|
481 |
-
with gr.Row():
|
482 |
with gr.Column(scale=1):
|
483 |
-
|
484 |
-
type="pil",
|
485 |
-
label="๐ผ๏ธ Upload Image"
|
486 |
-
)
|
487 |
-
|
488 |
-
prompt_input = gr.Textbox(
|
489 |
-
label="โจ Animation Prompt",
|
490 |
-
value=config.default_prompt,
|
491 |
-
placeholder="Describe the motion...",
|
492 |
-
lines=2,
|
493 |
-
max_lines=3
|
494 |
-
)
|
495 |
-
|
496 |
-
duration_input = gr.Slider(
|
497 |
-
minimum=0.3,
|
498 |
-
maximum=1.2,
|
499 |
-
step=0.1,
|
500 |
-
value=0.8,
|
501 |
-
label="โฑ๏ธ Duration (seconds)"
|
502 |
-
)
|
503 |
|
504 |
-
with gr.
|
505 |
-
|
506 |
-
|
507 |
-
|
508 |
-
|
509 |
)
|
510 |
|
511 |
-
|
512 |
-
|
513 |
-
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
label="Height"
|
518 |
-
)
|
519 |
-
width_slider = gr.Slider(
|
520 |
-
minimum=128,
|
521 |
-
maximum=512,
|
522 |
-
step=32,
|
523 |
-
value=256,
|
524 |
-
label="Width"
|
525 |
-
)
|
526 |
|
527 |
-
|
528 |
-
minimum=1,
|
529 |
-
maximum=5,
|
530 |
-
step=1,
|
531 |
value=2,
|
532 |
-
label="
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
533 |
)
|
534 |
|
535 |
with gr.Row():
|
@@ -538,19 +457,43 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
|
538 |
maximum=MAX_SEED,
|
539 |
step=1,
|
540 |
value=42,
|
541 |
-
label="Seed"
|
542 |
)
|
543 |
randomize_seed = gr.Checkbox(
|
544 |
-
label="
|
545 |
value=True
|
546 |
)
|
547 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
548 |
guidance_scale = gr.Slider(
|
549 |
minimum=0.0,
|
550 |
-
maximum=
|
551 |
step=0.5,
|
552 |
value=1.0,
|
553 |
-
label="Guidance Scale",
|
554 |
visible=False
|
555 |
)
|
556 |
|
@@ -561,44 +504,63 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
|
561 |
)
|
562 |
|
563 |
with gr.Column(scale=1):
|
|
|
564 |
video_output = gr.Video(
|
565 |
-
label="
|
566 |
-
autoplay=True
|
|
|
567 |
)
|
568 |
|
569 |
-
gr.
|
570 |
-
|
571 |
-
|
572 |
-
|
573 |
-
- **Duration**: 0.8s is optimal
|
574 |
-
- **Prompts**: Keep short and simple
|
575 |
-
- **Important**: Wait for completion!
|
576 |
-
|
577 |
-
### โ ๏ธ If GPU stops:
|
578 |
-
- Reduce resolution to 256ร256
|
579 |
-
- Use only 2 steps
|
580 |
-
- Keep duration under 1 second
|
581 |
-
- Avoid extreme aspect ratios
|
582 |
""")
|
583 |
|
584 |
-
#
|
585 |
-
|
586 |
-
|
587 |
-
|
588 |
-
|
589 |
-
)
|
590 |
-
|
591 |
-
generate_btn.click(
|
592 |
-
fn=generate_video,
|
593 |
-
inputs=[
|
594 |
-
input_image, prompt_input, height_slider, width_slider,
|
595 |
-
negative_prompt, duration_input, guidance_scale,
|
596 |
-
steps_slider, seed, randomize_seed
|
597 |
],
|
598 |
-
|
|
|
|
|
|
|
599 |
)
|
600 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
601 |
if __name__ == "__main__":
|
602 |
-
|
603 |
-
demo.queue(max_size=2) # ์์ ํ ์ฌ์ด์ฆ
|
604 |
-
demo.launch()
|
|
|
11 |
import random
|
12 |
import logging
|
13 |
import gc
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
# ๋ก๊น
์ค์
|
16 |
logging.basicConfig(level=logging.INFO)
|
17 |
logger = logging.getLogger(__name__)
|
18 |
|
19 |
+
# ๋ชจ๋ธ ์ค์
|
20 |
+
MODEL_ID = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
|
21 |
+
LORA_REPO_ID = "Kijai/WanVideo_comfy"
|
22 |
+
LORA_FILENAME = "Wan21_CausVid_14B_T2V_lora_rank32.safetensors"
|
23 |
+
|
24 |
+
# ํ๋ผ๋ฏธํฐ ์ค์
|
25 |
+
MOD_VALUE = 32
|
26 |
+
DEFAULT_H_SLIDER_VALUE = 512
|
27 |
+
DEFAULT_W_SLIDER_VALUE = 512 # Zero GPU๋ฅผ ์ํด ์ ์ฌ๊ฐํ ๊ธฐ๋ณธ๊ฐ
|
28 |
+
NEW_FORMULA_MAX_AREA = 480.0 * 832.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
+
SLIDER_MIN_H, SLIDER_MAX_H = 128, 896
|
31 |
+
SLIDER_MIN_W, SLIDER_MAX_W = 128, 896
|
32 |
MAX_SEED = np.iinfo(np.int32).max
|
33 |
|
34 |
+
FIXED_FPS = 24
|
35 |
+
MIN_FRAMES_MODEL = 8
|
36 |
+
MAX_FRAMES_MODEL = 81
|
37 |
+
|
38 |
+
default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
|
39 |
+
default_negative_prompt = "static, blurred, low quality, watermark, text"
|
40 |
+
|
41 |
+
# ๋ชจ๋ธ ๊ธ๋ก๋ฒ ๋ก๋ฉ
|
42 |
+
logger.info("Loading model components...")
|
43 |
+
image_encoder = CLIPVisionModel.from_pretrained(MODEL_ID, subfolder="image_encoder", torch_dtype=torch.float32)
|
44 |
+
vae = AutoencoderKLWan.from_pretrained(MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
|
45 |
+
pipe = WanImageToVideoPipeline.from_pretrained(
|
46 |
+
MODEL_ID, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16
|
47 |
+
)
|
48 |
+
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=8.0)
|
49 |
+
pipe.to("cuda")
|
50 |
+
|
51 |
+
# LoRA ๋ก๋ฉ
|
52 |
+
try:
|
53 |
+
causvid_path = hf_hub_download(repo_id=LORA_REPO_ID, filename=LORA_FILENAME)
|
54 |
+
pipe.load_lora_weights(causvid_path, adapter_name="causvid_lora")
|
55 |
+
pipe.set_adapters(["causvid_lora"], adapter_weights=[0.95])
|
56 |
+
pipe.fuse_lora()
|
57 |
+
logger.info("LoRA loaded successfully")
|
58 |
+
except Exception as e:
|
59 |
+
logger.warning(f"LoRA loading failed: {e}")
|
60 |
+
|
61 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ต์ ํ ํ์ฑํ
|
62 |
+
pipe.enable_vae_slicing()
|
63 |
+
pipe.enable_vae_tiling()
|
64 |
+
pipe.enable_model_cpu_offload()
|
65 |
+
|
66 |
+
logger.info("Model loaded and ready")
|
67 |
+
|
68 |
+
def _calculate_new_dimensions_wan(pil_image, mod_val, calculation_max_area,
|
69 |
+
min_slider_h, max_slider_h,
|
70 |
+
min_slider_w, max_slider_w,
|
71 |
+
default_h, default_w):
|
72 |
+
orig_w, orig_h = pil_image.size
|
73 |
+
if orig_w <= 0 or orig_h <= 0:
|
74 |
+
return default_h, default_w
|
75 |
+
|
76 |
+
aspect_ratio = orig_h / orig_w
|
77 |
|
78 |
+
# Zero GPU๋ฅผ ์ํ ๋ณด์์ ์ธ ๊ณ์ฐ
|
79 |
+
if hasattr(spaces, 'GPU'):
|
80 |
+
# ๋ ์์ max_area ์ฌ์ฉ
|
81 |
+
calculation_max_area = min(calculation_max_area, 320.0 * 320.0)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
|
83 |
+
calc_h = round(np.sqrt(calculation_max_area * aspect_ratio))
|
84 |
+
calc_w = round(np.sqrt(calculation_max_area / aspect_ratio))
|
85 |
+
|
86 |
+
calc_h = max(mod_val, (calc_h // mod_val) * mod_val)
|
87 |
+
calc_w = max(mod_val, (calc_w // mod_val) * mod_val)
|
88 |
+
|
89 |
+
# Zero GPU ํ๊ฒฝ์์ ์ถ๊ฐ ์ ํ
|
90 |
+
if hasattr(spaces, 'GPU'):
|
91 |
+
max_slider_h = min(max_slider_h, 640)
|
92 |
+
max_slider_w = min(max_slider_w, 640)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
|
94 |
+
new_h = int(np.clip(calc_h, min_slider_h, (max_slider_h // mod_val) * mod_val))
|
95 |
+
new_w = int(np.clip(calc_w, min_slider_w, (max_slider_w // mod_val) * mod_val))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
|
97 |
+
return new_h, new_w
|
98 |
+
|
99 |
+
def handle_image_upload_for_dims_wan(uploaded_pil_image, current_h_val, current_w_val):
|
100 |
+
if uploaded_pil_image is None:
|
101 |
+
return gr.update(value=DEFAULT_H_SLIDER_VALUE), gr.update(value=DEFAULT_W_SLIDER_VALUE)
|
102 |
try:
|
103 |
+
new_h, new_w = _calculate_new_dimensions_wan(
|
104 |
+
uploaded_pil_image, MOD_VALUE, NEW_FORMULA_MAX_AREA,
|
105 |
+
SLIDER_MIN_H, SLIDER_MAX_H, SLIDER_MIN_W, SLIDER_MAX_W,
|
106 |
+
DEFAULT_H_SLIDER_VALUE, DEFAULT_W_SLIDER_VALUE
|
107 |
+
)
|
108 |
return gr.update(value=new_h), gr.update(value=new_w)
|
|
|
109 |
except Exception as e:
|
110 |
+
gr.Warning("Error attempting to calculate new dimensions")
|
111 |
+
return gr.update(value=DEFAULT_H_SLIDER_VALUE), gr.update(value=DEFAULT_W_SLIDER_VALUE)
|
112 |
+
|
113 |
+
def get_duration(input_image, prompt, height, width,
|
114 |
+
negative_prompt, duration_seconds,
|
115 |
+
guidance_scale, steps,
|
116 |
+
seed, randomize_seed,
|
117 |
+
progress):
|
118 |
+
# Zero GPU๋ฅผ ์ํ ๋ณด์์ ์ธ ์๊ฐ ํ ๋น
|
119 |
+
base_time = 60
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
|
121 |
+
if hasattr(spaces, 'GPU'):
|
122 |
+
# Zero GPU ํ๊ฒฝ์์ ๋ ๋ง์ ์๊ฐ ํ ๋น
|
123 |
+
if steps > 4 and duration_seconds > 2:
|
124 |
+
return 90
|
125 |
+
elif steps > 4 or duration_seconds > 2:
|
126 |
+
return 80
|
127 |
+
else:
|
128 |
+
return 70
|
129 |
+
else:
|
130 |
+
# ์ผ๋ฐ GPU ํ๊ฒฝ
|
131 |
+
if steps > 4 and duration_seconds > 2:
|
132 |
+
return 90
|
133 |
+
elif steps > 4 or duration_seconds > 2:
|
134 |
+
return 75
|
135 |
+
else:
|
136 |
+
return 60
|
137 |
|
138 |
@spaces.GPU(duration=get_duration)
|
|
|
139 |
def generate_video(input_image, prompt, height, width,
|
140 |
+
negative_prompt=default_negative_prompt, duration_seconds = 2,
|
141 |
+
guidance_scale = 1, steps = 4,
|
142 |
+
seed = 42, randomize_seed = False,
|
143 |
progress=gr.Progress(track_tqdm=True)):
|
144 |
|
145 |
+
if input_image is None:
|
146 |
+
raise gr.Error("Please upload an input image.")
|
|
|
|
|
|
|
147 |
|
148 |
+
# Zero GPU ํ๊ฒฝ์์ ์ถ๊ฐ ๊ฒ์ฆ
|
149 |
+
if hasattr(spaces, 'GPU'):
|
150 |
+
# ํฝ์
์ ํ
|
151 |
+
max_pixels = 409600 # 640x640
|
152 |
+
if height * width > max_pixels:
|
153 |
+
raise gr.Error(f"Resolution too high for Zero GPU. Maximum {max_pixels:,} pixels (e.g., 640ร640)")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
|
155 |
+
# Duration ์ ํ
|
156 |
+
if duration_seconds > 2.5:
|
157 |
+
duration_seconds = 2.5
|
158 |
+
gr.Warning("Duration limited to 2.5s in Zero GPU environment")
|
159 |
|
160 |
+
# Steps ์ ํ
|
161 |
+
if steps > 8:
|
162 |
+
steps = 8
|
163 |
+
gr.Warning("Steps limited to 8 in Zero GPU environment")
|
164 |
+
|
165 |
+
target_h = max(MOD_VALUE, (int(height) // MOD_VALUE) * MOD_VALUE)
|
166 |
+
target_w = max(MOD_VALUE, (int(width) // MOD_VALUE) * MOD_VALUE)
|
167 |
+
|
168 |
+
num_frames = np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
|
169 |
+
|
170 |
+
# Zero GPU์์ ํ๋ ์ ์ ์ถ๊ฐ ์ ํ
|
171 |
+
if hasattr(spaces, 'GPU'):
|
172 |
+
max_frames_zerogpu = int(2.5 * FIXED_FPS) # 2.5์ด
|
173 |
+
num_frames = min(num_frames, max_frames_zerogpu)
|
174 |
+
|
175 |
+
current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
|
176 |
+
|
177 |
+
logger.info(f"Generating video: {target_h}x{target_w}, {num_frames} frames, seed={current_seed}")
|
178 |
+
|
179 |
+
# ์ด๋ฏธ์ง ๋ฆฌ์ฌ์ด์ฆ
|
180 |
+
resized_image = input_image.resize((target_w, target_h), Image.Resampling.LANCZOS)
|
181 |
+
|
182 |
+
try:
|
183 |
+
with torch.inference_mode():
|
184 |
+
output_frames_list = pipe(
|
185 |
+
image=resized_image,
|
186 |
+
prompt=prompt,
|
187 |
+
negative_prompt=negative_prompt,
|
188 |
+
height=target_h,
|
189 |
+
width=target_w,
|
190 |
+
num_frames=num_frames,
|
191 |
+
guidance_scale=float(guidance_scale),
|
192 |
+
num_inference_steps=int(steps),
|
193 |
+
generator=torch.Generator(device="cuda").manual_seed(current_seed)
|
194 |
+
).frames[0]
|
195 |
+
except torch.cuda.OutOfMemoryError:
|
196 |
gc.collect()
|
197 |
+
torch.cuda.empty_cache()
|
198 |
+
raise gr.Error("GPU out of memory. Try smaller resolution or shorter duration.")
|
|
|
|
|
|
|
199 |
except Exception as e:
|
200 |
+
logger.error(f"Generation failed: {e}")
|
201 |
+
raise gr.Error(f"Video generation failed: {str(e)[:100]}")
|
202 |
+
|
203 |
+
# ๋น๋์ค ์ ์ฅ
|
204 |
+
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
|
205 |
+
video_path = tmpfile.name
|
206 |
+
export_to_video(output_frames_list, video_path, fps=FIXED_FPS)
|
207 |
+
|
208 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
209 |
+
del output_frames_list
|
210 |
+
gc.collect()
|
211 |
+
if torch.cuda.is_available():
|
212 |
+
torch.cuda.empty_cache()
|
213 |
+
|
214 |
+
return video_path, current_seed
|
215 |
|
216 |
+
# CSS ์คํ์ผ (๊ธฐ์กด UI ์ ์ง)
|
217 |
css = """
|
218 |
.container {
|
219 |
+
max-width: 1200px;
|
220 |
margin: auto;
|
221 |
padding: 20px;
|
222 |
}
|
223 |
|
224 |
.header {
|
225 |
text-align: center;
|
226 |
+
margin-bottom: 30px;
|
227 |
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
|
228 |
+
padding: 40px;
|
229 |
+
border-radius: 20px;
|
230 |
color: white;
|
231 |
+
box-shadow: 0 10px 30px rgba(0,0,0,0.2);
|
232 |
+
position: relative;
|
233 |
+
overflow: hidden;
|
234 |
+
}
|
235 |
+
|
236 |
+
.header::before {
|
237 |
+
content: '';
|
238 |
+
position: absolute;
|
239 |
+
top: -50%;
|
240 |
+
left: -50%;
|
241 |
+
width: 200%;
|
242 |
+
height: 200%;
|
243 |
+
background: radial-gradient(circle, rgba(255,255,255,0.1) 0%, transparent 70%);
|
244 |
+
animation: pulse 4s ease-in-out infinite;
|
245 |
+
}
|
246 |
+
|
247 |
+
@keyframes pulse {
|
248 |
+
0%, 100% { transform: scale(1); opacity: 0.5; }
|
249 |
+
50% { transform: scale(1.1); opacity: 0.8; }
|
250 |
}
|
251 |
|
252 |
.header h1 {
|
253 |
+
font-size: 3em;
|
254 |
margin-bottom: 10px;
|
255 |
+
text-shadow: 2px 2px 4px rgba(0,0,0,0.3);
|
256 |
+
position: relative;
|
257 |
+
z-index: 1;
|
258 |
}
|
259 |
|
260 |
+
.header p {
|
261 |
+
font-size: 1.2em;
|
262 |
+
opacity: 0.95;
|
263 |
+
position: relative;
|
264 |
+
z-index: 1;
|
265 |
+
}
|
266 |
+
|
267 |
+
.gpu-status {
|
268 |
+
position: absolute;
|
269 |
+
top: 10px;
|
270 |
+
right: 10px;
|
271 |
+
background: rgba(0,0,0,0.3);
|
272 |
+
padding: 5px 15px;
|
273 |
+
border-radius: 20px;
|
274 |
+
font-size: 0.8em;
|
275 |
+
}
|
276 |
+
|
277 |
+
.main-content {
|
278 |
+
background: rgba(255, 255, 255, 0.95);
|
279 |
+
border-radius: 20px;
|
280 |
+
padding: 30px;
|
281 |
+
box-shadow: 0 5px 20px rgba(0,0,0,0.1);
|
282 |
+
backdrop-filter: blur(10px);
|
283 |
+
}
|
284 |
+
|
285 |
+
.input-section {
|
286 |
+
background: linear-gradient(135deg, #f5f7fa 0%, #c3cfe2 100%);
|
287 |
+
padding: 25px;
|
288 |
+
border-radius: 15px;
|
289 |
+
margin-bottom: 20px;
|
290 |
}
|
291 |
|
292 |
.generate-btn {
|
293 |
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
|
294 |
color: white;
|
295 |
+
font-size: 1.3em;
|
296 |
+
padding: 15px 40px;
|
297 |
+
border-radius: 30px;
|
298 |
border: none;
|
299 |
cursor: pointer;
|
300 |
+
transition: all 0.3s ease;
|
301 |
+
box-shadow: 0 5px 15px rgba(102, 126, 234, 0.4);
|
302 |
width: 100%;
|
303 |
+
margin-top: 20px;
|
304 |
}
|
305 |
|
306 |
.generate-btn:hover {
|
307 |
transform: translateY(-2px);
|
308 |
+
box-shadow: 0 7px 20px rgba(102, 126, 234, 0.6);
|
309 |
+
}
|
310 |
+
|
311 |
+
.generate-btn:active {
|
312 |
+
transform: translateY(0);
|
313 |
+
}
|
314 |
+
|
315 |
+
.video-output {
|
316 |
+
background: #f8f9fa;
|
317 |
+
padding: 20px;
|
318 |
+
border-radius: 15px;
|
319 |
+
text-align: center;
|
320 |
+
min-height: 400px;
|
321 |
+
display: flex;
|
322 |
+
align-items: center;
|
323 |
+
justify-content: center;
|
324 |
+
}
|
325 |
+
|
326 |
+
.accordion {
|
327 |
+
background: rgba(255, 255, 255, 0.7);
|
328 |
+
border-radius: 10px;
|
329 |
+
margin-top: 15px;
|
330 |
+
padding: 15px;
|
331 |
+
}
|
332 |
+
|
333 |
+
.slider-container {
|
334 |
+
background: rgba(255, 255, 255, 0.5);
|
335 |
+
padding: 15px;
|
336 |
+
border-radius: 10px;
|
337 |
+
margin: 10px 0;
|
338 |
+
}
|
339 |
+
|
340 |
+
body {
|
341 |
+
background: linear-gradient(-45deg, #ee7752, #e73c7e, #23a6d5, #23d5ab);
|
342 |
+
background-size: 400% 400%;
|
343 |
+
animation: gradient 15s ease infinite;
|
344 |
+
}
|
345 |
+
|
346 |
+
@keyframes gradient {
|
347 |
+
0% { background-position: 0% 50%; }
|
348 |
+
50% { background-position: 100% 50%; }
|
349 |
+
100% { background-position: 0% 50%; }
|
350 |
+
}
|
351 |
+
|
352 |
+
.warning-box {
|
353 |
+
background: rgba(255, 193, 7, 0.1);
|
354 |
+
border: 1px solid rgba(255, 193, 7, 0.3);
|
355 |
+
border-radius: 10px;
|
356 |
+
padding: 15px;
|
357 |
+
margin: 10px 0;
|
358 |
+
color: #856404;
|
359 |
+
font-size: 0.9em;
|
360 |
+
}
|
361 |
+
|
362 |
+
.info-box {
|
363 |
+
background: rgba(52, 152, 219, 0.1);
|
364 |
+
border: 1px solid rgba(52, 152, 219, 0.3);
|
365 |
+
border-radius: 10px;
|
366 |
+
padding: 15px;
|
367 |
+
margin: 10px 0;
|
368 |
+
color: #2c5282;
|
369 |
+
font-size: 0.9em;
|
370 |
+
}
|
371 |
+
|
372 |
+
.footer {
|
373 |
+
text-align: center;
|
374 |
+
margin-top: 30px;
|
375 |
+
color: #666;
|
376 |
+
font-size: 0.9em;
|
377 |
}
|
378 |
"""
|
379 |
|
380 |
+
# Gradio UI (๊ธฐ์กด ๊ตฌ์กฐ ์ ์ง)
|
381 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
382 |
with gr.Column(elem_classes="container"):
|
383 |
+
# Header with GPU status
|
384 |
gr.HTML("""
|
385 |
<div class="header">
|
386 |
+
<h1>๐ฌ AI Video Magic Studio</h1>
|
387 |
+
<p>Transform your images into captivating videos with Wan 2.1 + CausVid LoRA</p>
|
388 |
+
<div class="gpu-status">๐ฅ๏ธ Zero GPU Optimized</div>
|
389 |
</div>
|
390 |
""")
|
391 |
|
392 |
+
# GPU ๋ฉ๋ชจ๋ฆฌ ๊ฒฝ๊ณ
|
393 |
+
if hasattr(spaces, 'GPU'):
|
394 |
+
gr.HTML("""
|
395 |
+
<div class="warning-box">
|
396 |
+
<strong>๐ก Zero GPU Performance Tips:</strong>
|
397 |
+
<ul style="margin: 5px 0; padding-left: 20px;">
|
398 |
+
<li>Maximum duration: 2.5 seconds</li>
|
399 |
+
<li>Maximum resolution: 640ร640 pixels</li>
|
400 |
+
<li>Recommended: 512ร512 at 2 seconds</li>
|
401 |
+
<li>Use 4-6 steps for optimal speed/quality balance</li>
|
402 |
+
<li>Processing time: ~60-90 seconds</li>
|
403 |
+
</ul>
|
404 |
+
</div>
|
405 |
+
""")
|
406 |
+
|
407 |
+
# ์ ๋ณด ๋ฐ์ค
|
408 |
gr.HTML("""
|
409 |
+
<div class="info-box">
|
410 |
+
<strong>๐ฏ Quick Start Guide:</strong>
|
411 |
+
<ol style="margin: 5px 0; padding-left: 20px;">
|
412 |
+
<li>Upload your image - AI will calculate optimal dimensions</li>
|
413 |
+
<li>Enter a creative prompt or use the default</li>
|
414 |
+
<li>Adjust duration (2s recommended for best results)</li>
|
415 |
+
<li>Click Generate and wait for completion</li>
|
416 |
+
</ol>
|
|
|
417 |
</div>
|
418 |
""")
|
419 |
|
420 |
+
with gr.Row(elem_classes="main-content"):
|
421 |
with gr.Column(scale=1):
|
422 |
+
gr.Markdown("### ๐ธ Input Settings")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
423 |
|
424 |
+
with gr.Column(elem_classes="input-section"):
|
425 |
+
input_image = gr.Image(
|
426 |
+
type="pil",
|
427 |
+
label="๐ผ๏ธ Upload Your Image",
|
428 |
+
elem_classes="image-upload"
|
429 |
)
|
430 |
|
431 |
+
prompt_input = gr.Textbox(
|
432 |
+
label="โจ Animation Prompt",
|
433 |
+
value=default_prompt_i2v,
|
434 |
+
placeholder="Describe how you want your image to move...",
|
435 |
+
lines=2
|
436 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
437 |
|
438 |
+
duration_input = gr.Slider(
|
439 |
+
minimum=round(MIN_FRAMES_MODEL/FIXED_FPS, 1),
|
440 |
+
maximum=round(MAX_FRAMES_MODEL/FIXED_FPS, 1) if not hasattr(spaces, 'GPU') else 2.5,
|
441 |
+
step=0.1,
|
442 |
value=2,
|
443 |
+
label=f"โฑ๏ธ Video Duration (seconds) - Clamped to {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps",
|
444 |
+
elem_classes="slider-container"
|
445 |
+
)
|
446 |
+
|
447 |
+
with gr.Accordion("๐๏ธ Advanced Settings", open=False, elem_classes="accordion"):
|
448 |
+
negative_prompt = gr.Textbox(
|
449 |
+
label="๐ซ Negative Prompt",
|
450 |
+
value=default_negative_prompt,
|
451 |
+
lines=3
|
452 |
)
|
453 |
|
454 |
with gr.Row():
|
|
|
457 |
maximum=MAX_SEED,
|
458 |
step=1,
|
459 |
value=42,
|
460 |
+
label="๐ฒ Seed"
|
461 |
)
|
462 |
randomize_seed = gr.Checkbox(
|
463 |
+
label="๐ Randomize",
|
464 |
value=True
|
465 |
)
|
466 |
|
467 |
+
with gr.Row():
|
468 |
+
height_slider = gr.Slider(
|
469 |
+
minimum=SLIDER_MIN_H,
|
470 |
+
maximum=SLIDER_MAX_H if not hasattr(spaces, 'GPU') else 640,
|
471 |
+
step=MOD_VALUE,
|
472 |
+
value=DEFAULT_H_SLIDER_VALUE,
|
473 |
+
label=f"๐ Height (multiple of {MOD_VALUE})"
|
474 |
+
)
|
475 |
+
width_slider = gr.Slider(
|
476 |
+
minimum=SLIDER_MIN_W,
|
477 |
+
maximum=SLIDER_MAX_W if not hasattr(spaces, 'GPU') else 640,
|
478 |
+
step=MOD_VALUE,
|
479 |
+
value=DEFAULT_W_SLIDER_VALUE,
|
480 |
+
label=f"๐ Width (multiple of {MOD_VALUE})"
|
481 |
+
)
|
482 |
+
|
483 |
+
steps_slider = gr.Slider(
|
484 |
+
minimum=1,
|
485 |
+
maximum=30 if not hasattr(spaces, 'GPU') else 8,
|
486 |
+
step=1,
|
487 |
+
value=4,
|
488 |
+
label="๐ง Quality Steps (4-6 recommended)"
|
489 |
+
)
|
490 |
+
|
491 |
guidance_scale = gr.Slider(
|
492 |
minimum=0.0,
|
493 |
+
maximum=20.0,
|
494 |
step=0.5,
|
495 |
value=1.0,
|
496 |
+
label="๐ฏ Guidance Scale",
|
497 |
visible=False
|
498 |
)
|
499 |
|
|
|
504 |
)
|
505 |
|
506 |
with gr.Column(scale=1):
|
507 |
+
gr.Markdown("### ๐ฅ Generated Video")
|
508 |
video_output = gr.Video(
|
509 |
+
label="",
|
510 |
+
autoplay=True,
|
511 |
+
elem_classes="video-output"
|
512 |
)
|
513 |
|
514 |
+
gr.HTML("""
|
515 |
+
<div class="footer">
|
516 |
+
<p>๐ก Tip: For best results, use clear images with good lighting and distinct subjects</p>
|
517 |
+
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
518 |
""")
|
519 |
|
520 |
+
# Examples
|
521 |
+
gr.Examples(
|
522 |
+
examples=[
|
523 |
+
["peng.png", "a penguin playfully dancing in the snow, Antarctica", 512, 512],
|
524 |
+
["forg.jpg", "the frog jumps around", 448, 576],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
525 |
],
|
526 |
+
inputs=[input_image, prompt_input, height_slider, width_slider],
|
527 |
+
outputs=[video_output, seed],
|
528 |
+
fn=generate_video,
|
529 |
+
cache_examples=False # ์บ์ ๋นํ์ฑํ๋ก ๋ฉ๋ชจ๋ฆฌ ์ ์ฝ
|
530 |
)
|
531 |
|
532 |
+
# ๊ฐ์ ์ฌํญ ์์ฝ
|
533 |
+
gr.HTML("""
|
534 |
+
<div style="background: rgba(255,255,255,0.9); border-radius: 10px; padding: 15px; margin-top: 20px; font-size: 0.8em; text-align: center;">
|
535 |
+
<p style="margin: 0; color: #666;">
|
536 |
+
<strong style="color: #667eea;">Powered by:</strong>
|
537 |
+
Wan 2.1 I2V (14B) + CausVid LoRA โข ๐ 4-8 steps fast inference โข ๐ฌ Up to 81 frames
|
538 |
+
</p>
|
539 |
+
</div>
|
540 |
+
""")
|
541 |
+
|
542 |
+
# Event handlers
|
543 |
+
input_image.upload(
|
544 |
+
fn=handle_image_upload_for_dims_wan,
|
545 |
+
inputs=[input_image, height_slider, width_slider],
|
546 |
+
outputs=[height_slider, width_slider]
|
547 |
+
)
|
548 |
+
|
549 |
+
input_image.clear(
|
550 |
+
fn=handle_image_upload_for_dims_wan,
|
551 |
+
inputs=[input_image, height_slider, width_slider],
|
552 |
+
outputs=[height_slider, width_slider]
|
553 |
+
)
|
554 |
+
|
555 |
+
generate_btn.click(
|
556 |
+
fn=generate_video,
|
557 |
+
inputs=[
|
558 |
+
input_image, prompt_input, height_slider, width_slider,
|
559 |
+
negative_prompt, duration_input, guidance_scale,
|
560 |
+
steps_slider, seed, randomize_seed
|
561 |
+
],
|
562 |
+
outputs=[video_output, seed]
|
563 |
+
)
|
564 |
+
|
565 |
if __name__ == "__main__":
|
566 |
+
demo.queue().launch()
|
|
|
|