Spaces:
Running
on
L40S
Running
on
L40S
Update app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,10 @@
|
|
1 |
import gradio as gr
|
2 |
import spaces
|
3 |
from gradio_litmodel3d import LitModel3D
|
|
|
4 |
import os
|
|
|
|
|
5 |
import torch
|
6 |
import numpy as np
|
7 |
import imageio
|
@@ -11,162 +14,28 @@ from PIL import Image
|
|
11 |
from trellis.pipelines import TrellisImageTo3DPipeline
|
12 |
from trellis.representations import Gaussian, MeshExtractResult
|
13 |
from trellis.utils import render_utils, postprocessing_utils
|
14 |
-
from transformers import pipeline as translation_pipeline
|
15 |
-
from diffusers import FluxPipeline
|
16 |
-
from typing import *
|
17 |
|
18 |
|
19 |
MAX_SEED = np.iinfo(np.int32).max
|
20 |
TMP_DIR = "/tmp/Trellis-demo"
|
21 |
-
os.makedirs(TMP_DIR, exist_ok=True)
|
22 |
-
|
23 |
-
# GPU 메모리 관련 환경 변수
|
24 |
-
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128' # 더 작은 값으로 설정
|
25 |
-
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
|
26 |
-
os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
|
27 |
-
os.environ['PYTORCH_NO_CUDA_MEMORY_CACHING'] = '1'
|
28 |
-
os.environ['CUDA_CACHE_DISABLE'] = '1'
|
29 |
-
|
30 |
-
def initialize_models():
|
31 |
-
global pipeline, translator, flux_pipe
|
32 |
-
|
33 |
-
try:
|
34 |
-
# CUDA 설정
|
35 |
-
if torch.cuda.is_available():
|
36 |
-
torch.backends.cudnn.benchmark = True
|
37 |
-
torch.backends.cuda.matmul.allow_tf32 = True
|
38 |
-
torch.backends.cudnn.allow_tf32 = True
|
39 |
-
|
40 |
-
print("Initializing Trellis pipeline...")
|
41 |
-
try:
|
42 |
-
pipeline = TrellisImageTo3DPipeline.from_pretrained(
|
43 |
-
"JeffreyXiang/TRELLIS-image-large"
|
44 |
-
)
|
45 |
-
|
46 |
-
if pipeline is None:
|
47 |
-
raise ValueError("Pipeline initialization returned None")
|
48 |
-
|
49 |
-
if torch.cuda.is_available():
|
50 |
-
pipeline = pipeline.to("cuda")
|
51 |
-
# Half precision으로 변환
|
52 |
-
pipeline = pipeline.half()
|
53 |
-
|
54 |
-
except Exception as e:
|
55 |
-
print(f"Error initializing Trellis pipeline: {str(e)}")
|
56 |
-
raise
|
57 |
-
|
58 |
-
print("Initializing translator...")
|
59 |
-
try:
|
60 |
-
translator = translation_pipeline(
|
61 |
-
"translation",
|
62 |
-
model="Helsinki-NLP/opus-mt-ko-en",
|
63 |
-
device=0 if torch.cuda.is_available() else -1
|
64 |
-
)
|
65 |
-
except Exception as e:
|
66 |
-
print(f"Error initializing translator: {str(e)}")
|
67 |
-
raise
|
68 |
-
|
69 |
-
flux_pipe = None
|
70 |
-
|
71 |
-
print("Models initialized successfully")
|
72 |
-
return True
|
73 |
-
|
74 |
-
except Exception as e:
|
75 |
-
print(f"Model initialization error: {str(e)}")
|
76 |
-
free_memory()
|
77 |
-
return False
|
78 |
-
|
79 |
-
def get_flux_pipe():
|
80 |
-
"""Flux 파이프라인을 필요할 때만 로드하는 함수"""
|
81 |
-
global flux_pipe
|
82 |
-
if flux_pipe is None:
|
83 |
-
try:
|
84 |
-
free_memory()
|
85 |
-
flux_pipe = FluxPipeline.from_pretrained(
|
86 |
-
"black-forest-labs/FLUX.1-dev",
|
87 |
-
use_safetensors=True
|
88 |
-
)
|
89 |
-
if torch.cuda.is_available():
|
90 |
-
flux_pipe = flux_pipe.to("cuda")
|
91 |
-
flux_pipe.enable_model_cpu_offload() # CPU 오프로딩 활성화
|
92 |
-
except Exception as e:
|
93 |
-
print(f"Error loading Flux pipeline: {e}")
|
94 |
-
return None
|
95 |
-
return flux_pipe
|
96 |
-
|
97 |
-
def free_memory():
|
98 |
-
"""강화된 메모리 정리 함수"""
|
99 |
-
import gc
|
100 |
-
import os
|
101 |
-
|
102 |
-
# Python 가비지 컬렉션
|
103 |
-
gc.collect()
|
104 |
-
|
105 |
-
# CUDA 메모리 정리
|
106 |
-
if torch.cuda.is_available():
|
107 |
-
torch.cuda.empty_cache()
|
108 |
-
torch.cuda.synchronize()
|
109 |
-
|
110 |
-
# 임시 파일 정리
|
111 |
-
tmp_dirs = ['/tmp/transformers_cache', '/tmp/torch_home',
|
112 |
-
'/tmp/huggingface', '/tmp/cache', TMP_DIR]
|
113 |
-
|
114 |
-
for dir_path in tmp_dirs:
|
115 |
-
if os.path.exists(dir_path):
|
116 |
-
try:
|
117 |
-
for file in os.listdir(dir_path):
|
118 |
-
file_path = os.path.join(dir_path, file)
|
119 |
-
if os.path.isfile(file_path):
|
120 |
-
try:
|
121 |
-
os.unlink(file_path)
|
122 |
-
except:
|
123 |
-
pass
|
124 |
-
except:
|
125 |
-
pass
|
126 |
|
127 |
-
|
128 |
-
def setup_gpu_model(model):
|
129 |
-
"""GPU 설정이 필요한 모델을 처리하는 함수"""
|
130 |
-
if torch.cuda.is_available():
|
131 |
-
model = model.to("cuda")
|
132 |
-
return model
|
133 |
-
|
134 |
-
|
135 |
-
def translate_if_korean(text):
|
136 |
-
if any(ord('가') <= ord(char) <= ord('힣') for char in text):
|
137 |
-
translated = translator(text)[0]['translation_text']
|
138 |
-
return translated
|
139 |
-
return text
|
140 |
|
141 |
|
142 |
def preprocess_image(image: Image.Image) -> Tuple[str, Image.Image]:
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
# 이미지 전처리
|
157 |
-
processed_image = pipeline.preprocess_image(image)
|
158 |
-
if processed_image is None:
|
159 |
-
raise Exception("Failed to process image")
|
160 |
-
|
161 |
-
# 임시 파일 저장
|
162 |
-
save_path = os.path.join(TMP_DIR, f"{trial_id}.png")
|
163 |
-
processed_image.save(save_path)
|
164 |
-
|
165 |
-
return trial_id, processed_image
|
166 |
-
|
167 |
-
except Exception as e:
|
168 |
-
print(f"Error in preprocess_image: {str(e)}")
|
169 |
-
return None, None
|
170 |
|
171 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult, trial_id: str) -> dict:
|
172 |
return {
|
@@ -184,7 +53,8 @@ def pack_state(gs: Gaussian, mesh: MeshExtractResult, trial_id: str) -> dict:
|
|
184 |
},
|
185 |
'trial_id': trial_id,
|
186 |
}
|
187 |
-
|
|
|
188 |
def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
|
189 |
gs = Gaussian(
|
190 |
aabb=state['gaussian']['aabb'],
|
@@ -207,190 +77,116 @@ def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
|
|
207 |
|
208 |
return gs, mesh, state['trial_id']
|
209 |
|
210 |
-
def image_to_3d(trial_id: str, seed: int, randomize_seed: bool, ss_guidance_strength: float,
|
211 |
-
ss_sampling_steps: int, slat_guidance_strength: float, slat_sampling_steps: int):
|
212 |
-
try:
|
213 |
-
if randomize_seed:
|
214 |
-
seed = np.random.randint(0, MAX_SEED)
|
215 |
-
|
216 |
-
input_image = Image.open(f"{TMP_DIR}/{trial_id}.png")
|
217 |
-
|
218 |
-
# L40S에 맞게 이미지 크기 제한 조정
|
219 |
-
max_size = 768
|
220 |
-
if max(input_image.size) > max_size:
|
221 |
-
ratio = max_size / max(input_image.size)
|
222 |
-
input_image = input_image.resize(
|
223 |
-
(int(input_image.size[0] * ratio),
|
224 |
-
int(input_image.size[1] * ratio)),
|
225 |
-
Image.LANCZOS
|
226 |
-
)
|
227 |
-
|
228 |
-
if torch.cuda.is_available():
|
229 |
-
pipeline.to("cuda")
|
230 |
-
|
231 |
-
try:
|
232 |
-
outputs = pipeline.run(
|
233 |
-
input_image,
|
234 |
-
seed=seed,
|
235 |
-
formats=["gaussian", "mesh"],
|
236 |
-
preprocess_image=False,
|
237 |
-
sparse_structure_sampler_params={
|
238 |
-
"steps": min(ss_sampling_steps, 20),
|
239 |
-
"cfg_strength": ss_guidance_strength,
|
240 |
-
},
|
241 |
-
slat_sampler_params={
|
242 |
-
"steps": min(slat_sampling_steps, 20),
|
243 |
-
"cfg_strength": slat_guidance_strength,
|
244 |
-
}
|
245 |
-
)
|
246 |
-
except RuntimeError as e:
|
247 |
-
print(f"Runtime error in pipeline.run: {str(e)}")
|
248 |
-
free_memory()
|
249 |
-
raise e
|
250 |
-
|
251 |
-
# 비디오 생성
|
252 |
-
video = render_utils.render_video(outputs['gaussian'][0], num_frames=40)['color']
|
253 |
-
video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=40)['normal']
|
254 |
-
video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
|
255 |
-
|
256 |
-
trial_id = str(uuid.uuid4())
|
257 |
-
video_path = f"{TMP_DIR}/{trial_id}.mp4"
|
258 |
-
os.makedirs(os.path.dirname(video_path), exist_ok=True)
|
259 |
-
imageio.mimsave(video_path, video, fps=20)
|
260 |
-
|
261 |
-
state = pack_state(outputs['gaussian'][0], outputs['mesh'][0], trial_id)
|
262 |
-
|
263 |
-
if torch.cuda.is_available():
|
264 |
-
pipeline.to("cpu")
|
265 |
-
|
266 |
-
return state, video_path
|
267 |
-
|
268 |
-
except Exception as e:
|
269 |
-
print(f"Error in image_to_3d: {str(e)}")
|
270 |
-
if torch.cuda.is_available():
|
271 |
-
pipeline.to("cpu")
|
272 |
-
raise e
|
273 |
-
|
274 |
-
|
275 |
-
def generate_image_from_text(prompt, height, width, guidance_scale, num_steps):
|
276 |
-
try:
|
277 |
-
free_memory()
|
278 |
-
|
279 |
-
flux_pipe = get_flux_pipe()
|
280 |
-
if flux_pipe is None:
|
281 |
-
raise Exception("Failed to load Flux pipeline")
|
282 |
-
|
283 |
-
# L40S에 맞게 크기 제한 조정
|
284 |
-
height = min(height, 1024)
|
285 |
-
width = min(width, 1024)
|
286 |
-
|
287 |
-
translated_prompt = translate_if_korean(prompt)
|
288 |
-
final_prompt = f"{translated_prompt}, wbgmsst, 3D, white background"
|
289 |
-
|
290 |
-
with torch.cuda.amp.autocast():
|
291 |
-
output = flux_pipe(
|
292 |
-
prompt=[final_prompt],
|
293 |
-
height=height,
|
294 |
-
width=width,
|
295 |
-
guidance_scale=guidance_scale,
|
296 |
-
num_inference_steps=num_steps,
|
297 |
-
generator=torch.Generator(device='cuda')
|
298 |
-
)
|
299 |
-
|
300 |
-
image = output.images[0]
|
301 |
-
|
302 |
-
free_memory()
|
303 |
-
return image
|
304 |
-
|
305 |
-
except Exception as e:
|
306 |
-
print(f"Error in generate_image_from_text: {str(e)}")
|
307 |
-
free_memory()
|
308 |
-
raise e
|
309 |
-
|
310 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
311 |
def extract_glb(state: dict, mesh_simplify: float, texture_size: int) -> Tuple[str, str]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
312 |
gs, mesh, trial_id = unpack_state(state)
|
313 |
glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
|
314 |
glb_path = f"{TMP_DIR}/{trial_id}.glb"
|
315 |
glb.export(glb_path)
|
316 |
return glb_path, glb_path
|
317 |
|
|
|
318 |
def activate_button() -> gr.Button:
|
319 |
return gr.Button(interactive=True)
|
320 |
|
|
|
321 |
def deactivate_button() -> gr.Button:
|
322 |
return gr.Button(interactive=False)
|
323 |
|
324 |
-
css = """
|
325 |
-
footer {
|
326 |
-
visibility: hidden;
|
327 |
-
}
|
328 |
-
"""
|
329 |
|
330 |
-
|
331 |
-
with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
|
332 |
gr.Markdown("""
|
333 |
-
|
|
|
|
|
334 |
""")
|
335 |
|
336 |
-
with gr.
|
337 |
-
with gr.
|
338 |
-
|
339 |
-
|
340 |
-
|
341 |
-
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
|
359 |
-
|
360 |
-
extract_glb_btn = gr.Button("Extract GLB", interactive=False)
|
361 |
-
|
362 |
-
with gr.Column():
|
363 |
-
video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
|
364 |
-
model_output = LitModel3D(label="Extracted GLB", exposure=20.0, height=300)
|
365 |
-
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
366 |
|
367 |
-
with gr.
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
|
372 |
-
placeholder="Enter your image description...",
|
373 |
-
lines=3
|
374 |
-
)
|
375 |
-
|
376 |
-
with gr.Row():
|
377 |
-
txt2img_height = gr.Slider(256, 1024, value=512, step=64, label="Height")
|
378 |
-
txt2img_width = gr.Slider(256, 1024, value=512, step=64, label="Width")
|
379 |
-
|
380 |
-
with gr.Row():
|
381 |
-
guidance_scale = gr.Slider(1.0, 20.0, value=7.5, label="Guidance Scale")
|
382 |
-
num_steps = gr.Slider(1, 50, value=20, label="Number of Steps")
|
383 |
-
|
384 |
-
generate_txt2img_btn = gr.Button("Generate Image")
|
385 |
-
|
386 |
-
with gr.Column():
|
387 |
-
txt2img_output = gr.Image(label="Generated Image")
|
388 |
-
|
389 |
trial_id = gr.Textbox(visible=False)
|
390 |
output_buf = gr.State()
|
391 |
|
392 |
-
|
393 |
-
# Example images
|
394 |
with gr.Row():
|
395 |
examples = gr.Examples(
|
396 |
examples=[
|
@@ -401,11 +197,8 @@ with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
|
|
401 |
fn=preprocess_image,
|
402 |
outputs=[trial_id, image_prompt],
|
403 |
run_on_click=True,
|
404 |
-
examples_per_page=
|
405 |
-
cache_examples=False # 예제 캐싱 비활성화는 Examples 컴포넌트에서 설정
|
406 |
)
|
407 |
-
|
408 |
-
|
409 |
|
410 |
# Handlers
|
411 |
image_prompt.upload(
|
@@ -413,7 +206,6 @@ with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
|
|
413 |
inputs=[image_prompt],
|
414 |
outputs=[trial_id, image_prompt],
|
415 |
)
|
416 |
-
|
417 |
image_prompt.clear(
|
418 |
lambda: '',
|
419 |
outputs=[trial_id],
|
@@ -421,62 +213,39 @@ with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
|
|
421 |
|
422 |
generate_btn.click(
|
423 |
image_to_3d,
|
424 |
-
inputs=[trial_id, seed, randomize_seed, ss_guidance_strength, ss_sampling_steps,
|
425 |
-
slat_guidance_strength, slat_sampling_steps],
|
426 |
outputs=[output_buf, video_output],
|
427 |
-
concurrency_limit=1
|
428 |
).then(
|
429 |
activate_button,
|
430 |
-
outputs=[extract_glb_btn]
|
|
|
|
|
|
|
|
|
|
|
431 |
)
|
432 |
|
433 |
extract_glb_btn.click(
|
434 |
extract_glb,
|
435 |
inputs=[output_buf, mesh_simplify, texture_size],
|
436 |
outputs=[model_output, download_glb],
|
437 |
-
concurrency_limit=1
|
438 |
).then(
|
439 |
activate_button,
|
440 |
-
outputs=[download_glb]
|
441 |
-
)
|
442 |
-
generate_txt2img_btn.click(
|
443 |
-
generate_image_from_text,
|
444 |
-
inputs=[text_prompt, txt2img_height, txt2img_width, guidance_scale, num_steps],
|
445 |
-
outputs=[txt2img_output],
|
446 |
-
concurrency_limit=1,
|
447 |
-
show_progress=True # 진행 상황 표시
|
448 |
)
|
449 |
|
|
|
|
|
|
|
|
|
|
|
450 |
|
|
|
451 |
if __name__ == "__main__":
|
452 |
-
|
453 |
-
|
454 |
-
|
455 |
-
#
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
# CUDA 메모리 설정
|
461 |
-
torch.cuda.set_per_process_memory_fraction(0.8) # GPU 메모리 사용량 제한
|
462 |
-
|
463 |
-
# 디렉토리 생성
|
464 |
-
os.makedirs(TMP_DIR, exist_ok=True)
|
465 |
-
|
466 |
-
# 메모리 정리
|
467 |
-
free_memory()
|
468 |
-
|
469 |
-
# 모델 초기화
|
470 |
-
if not initialize_models():
|
471 |
-
print("Failed to initialize models")
|
472 |
-
exit(1)
|
473 |
-
|
474 |
-
# Gradio 앱 실행
|
475 |
-
demo.queue(max_size=1).launch(
|
476 |
-
share=True,
|
477 |
-
max_threads=2,
|
478 |
-
show_error=True,
|
479 |
-
server_port=7860,
|
480 |
-
server_name="0.0.0.0",
|
481 |
-
enable_queue=True
|
482 |
-
)
|
|
|
1 |
import gradio as gr
|
2 |
import spaces
|
3 |
from gradio_litmodel3d import LitModel3D
|
4 |
+
|
5 |
import os
|
6 |
+
os.environ['SPCONV_ALGO'] = 'native'
|
7 |
+
from typing import *
|
8 |
import torch
|
9 |
import numpy as np
|
10 |
import imageio
|
|
|
14 |
from trellis.pipelines import TrellisImageTo3DPipeline
|
15 |
from trellis.representations import Gaussian, MeshExtractResult
|
16 |
from trellis.utils import render_utils, postprocessing_utils
|
|
|
|
|
|
|
17 |
|
18 |
|
19 |
MAX_SEED = np.iinfo(np.int32).max
|
20 |
TMP_DIR = "/tmp/Trellis-demo"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
+
os.makedirs(TMP_DIR, exist_ok=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
|
25 |
def preprocess_image(image: Image.Image) -> Tuple[str, Image.Image]:
|
26 |
+
"""
|
27 |
+
Preprocess the input image.
|
28 |
+
Args:
|
29 |
+
image (Image.Image): The input image.
|
30 |
+
Returns:
|
31 |
+
str: uuid of the trial.
|
32 |
+
Image.Image: The preprocessed image.
|
33 |
+
"""
|
34 |
+
trial_id = str(uuid.uuid4())
|
35 |
+
processed_image = pipeline.preprocess_image(image)
|
36 |
+
processed_image.save(f"{TMP_DIR}/{trial_id}.png")
|
37 |
+
return trial_id, processed_image
|
38 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
40 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult, trial_id: str) -> dict:
|
41 |
return {
|
|
|
53 |
},
|
54 |
'trial_id': trial_id,
|
55 |
}
|
56 |
+
|
57 |
+
|
58 |
def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
|
59 |
gs = Gaussian(
|
60 |
aabb=state['gaussian']['aabb'],
|
|
|
77 |
|
78 |
return gs, mesh, state['trial_id']
|
79 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
|
81 |
+
@spaces.GPU
|
82 |
+
def image_to_3d(trial_id: str, seed: int, randomize_seed: bool, ss_guidance_strength: float, ss_sampling_steps: int, slat_guidance_strength: float, slat_sampling_steps: int) -> Tuple[dict, str]:
|
83 |
+
"""
|
84 |
+
Convert an image to a 3D model.
|
85 |
+
Args:
|
86 |
+
trial_id (str): The uuid of the trial.
|
87 |
+
seed (int): The random seed.
|
88 |
+
randomize_seed (bool): Whether to randomize the seed.
|
89 |
+
ss_guidance_strength (float): The guidance strength for sparse structure generation.
|
90 |
+
ss_sampling_steps (int): The number of sampling steps for sparse structure generation.
|
91 |
+
slat_guidance_strength (float): The guidance strength for structured latent generation.
|
92 |
+
slat_sampling_steps (int): The number of sampling steps for structured latent generation.
|
93 |
+
Returns:
|
94 |
+
dict: The information of the generated 3D model.
|
95 |
+
str: The path to the video of the 3D model.
|
96 |
+
"""
|
97 |
+
if randomize_seed:
|
98 |
+
seed = np.random.randint(0, MAX_SEED)
|
99 |
+
outputs = pipeline.run(
|
100 |
+
Image.open(f"{TMP_DIR}/{trial_id}.png"),
|
101 |
+
seed=seed,
|
102 |
+
formats=["gaussian", "mesh"],
|
103 |
+
preprocess_image=False,
|
104 |
+
sparse_structure_sampler_params={
|
105 |
+
"steps": ss_sampling_steps,
|
106 |
+
"cfg_strength": ss_guidance_strength,
|
107 |
+
},
|
108 |
+
slat_sampler_params={
|
109 |
+
"steps": slat_sampling_steps,
|
110 |
+
"cfg_strength": slat_guidance_strength,
|
111 |
+
},
|
112 |
+
)
|
113 |
+
video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
|
114 |
+
video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
|
115 |
+
video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
|
116 |
+
trial_id = uuid.uuid4()
|
117 |
+
video_path = f"{TMP_DIR}/{trial_id}.mp4"
|
118 |
+
os.makedirs(os.path.dirname(video_path), exist_ok=True)
|
119 |
+
imageio.mimsave(video_path, video, fps=15)
|
120 |
+
state = pack_state(outputs['gaussian'][0], outputs['mesh'][0], trial_id)
|
121 |
+
return state, video_path
|
122 |
+
|
123 |
+
|
124 |
+
@spaces.GPU
|
125 |
def extract_glb(state: dict, mesh_simplify: float, texture_size: int) -> Tuple[str, str]:
|
126 |
+
"""
|
127 |
+
Extract a GLB file from the 3D model.
|
128 |
+
Args:
|
129 |
+
state (dict): The state of the generated 3D model.
|
130 |
+
mesh_simplify (float): The mesh simplification factor.
|
131 |
+
texture_size (int): The texture resolution.
|
132 |
+
Returns:
|
133 |
+
str: The path to the extracted GLB file.
|
134 |
+
"""
|
135 |
gs, mesh, trial_id = unpack_state(state)
|
136 |
glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
|
137 |
glb_path = f"{TMP_DIR}/{trial_id}.glb"
|
138 |
glb.export(glb_path)
|
139 |
return glb_path, glb_path
|
140 |
|
141 |
+
|
142 |
def activate_button() -> gr.Button:
|
143 |
return gr.Button(interactive=True)
|
144 |
|
145 |
+
|
146 |
def deactivate_button() -> gr.Button:
|
147 |
return gr.Button(interactive=False)
|
148 |
|
|
|
|
|
|
|
|
|
|
|
149 |
|
150 |
+
with gr.Blocks() as demo:
|
|
|
151 |
gr.Markdown("""
|
152 |
+
## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
|
153 |
+
* Upload an image and click "Generate" to create a 3D asset. If the image has alpha channel, it be used as the mask. Otherwise, we use `rembg` to remove the background.
|
154 |
+
* If you find the generated 3D asset satisfactory, click "Extract GLB" to extract the GLB file and download it.
|
155 |
""")
|
156 |
|
157 |
+
with gr.Row():
|
158 |
+
with gr.Column():
|
159 |
+
image_prompt = gr.Image(label="Image Prompt", image_mode="RGBA", type="pil", height=300)
|
160 |
+
|
161 |
+
with gr.Accordion(label="Generation Settings", open=False):
|
162 |
+
seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
|
163 |
+
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
164 |
+
gr.Markdown("Stage 1: Sparse Structure Generation")
|
165 |
+
with gr.Row():
|
166 |
+
ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
|
167 |
+
ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
|
168 |
+
gr.Markdown("Stage 2: Structured Latent Generation")
|
169 |
+
with gr.Row():
|
170 |
+
slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
|
171 |
+
slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
|
172 |
+
|
173 |
+
generate_btn = gr.Button("Generate")
|
174 |
+
|
175 |
+
with gr.Accordion(label="GLB Extraction Settings", open=False):
|
176 |
+
mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
|
177 |
+
texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
|
178 |
+
|
179 |
+
extract_glb_btn = gr.Button("Extract GLB", interactive=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
180 |
|
181 |
+
with gr.Column():
|
182 |
+
video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
|
183 |
+
model_output = LitModel3D(label="Extracted GLB", exposure=20.0, height=300)
|
184 |
+
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
185 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
186 |
trial_id = gr.Textbox(visible=False)
|
187 |
output_buf = gr.State()
|
188 |
|
189 |
+
# Example images at the bottom of the page
|
|
|
190 |
with gr.Row():
|
191 |
examples = gr.Examples(
|
192 |
examples=[
|
|
|
197 |
fn=preprocess_image,
|
198 |
outputs=[trial_id, image_prompt],
|
199 |
run_on_click=True,
|
200 |
+
examples_per_page=64,
|
|
|
201 |
)
|
|
|
|
|
202 |
|
203 |
# Handlers
|
204 |
image_prompt.upload(
|
|
|
206 |
inputs=[image_prompt],
|
207 |
outputs=[trial_id, image_prompt],
|
208 |
)
|
|
|
209 |
image_prompt.clear(
|
210 |
lambda: '',
|
211 |
outputs=[trial_id],
|
|
|
213 |
|
214 |
generate_btn.click(
|
215 |
image_to_3d,
|
216 |
+
inputs=[trial_id, seed, randomize_seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
|
|
|
217 |
outputs=[output_buf, video_output],
|
|
|
218 |
).then(
|
219 |
activate_button,
|
220 |
+
outputs=[extract_glb_btn],
|
221 |
+
)
|
222 |
+
|
223 |
+
video_output.clear(
|
224 |
+
deactivate_button,
|
225 |
+
outputs=[extract_glb_btn],
|
226 |
)
|
227 |
|
228 |
extract_glb_btn.click(
|
229 |
extract_glb,
|
230 |
inputs=[output_buf, mesh_simplify, texture_size],
|
231 |
outputs=[model_output, download_glb],
|
|
|
232 |
).then(
|
233 |
activate_button,
|
234 |
+
outputs=[download_glb],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
)
|
236 |
|
237 |
+
model_output.clear(
|
238 |
+
deactivate_button,
|
239 |
+
outputs=[download_glb],
|
240 |
+
)
|
241 |
+
|
242 |
|
243 |
+
# Launch the Gradio app
|
244 |
if __name__ == "__main__":
|
245 |
+
pipeline = TrellisImageTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-image-large")
|
246 |
+
pipeline.cuda()
|
247 |
+
try:
|
248 |
+
pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))) # Preload rembg
|
249 |
+
except:
|
250 |
+
pass
|
251 |
+
demo.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|