Update app.py
Browse files
app.py
CHANGED
@@ -18,6 +18,32 @@ import gc
|
|
18 |
# 메모리 관리 설정 추가
|
19 |
import torch.backends.cuda
|
20 |
torch.backends.cuda.max_split_size_mb = 128 # 메모리 분할 크기 제한
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
# 메모리 관리 설정
|
22 |
torch.cuda.empty_cache()
|
23 |
gc.collect()
|
@@ -30,6 +56,16 @@ def clear_memory():
|
|
30 |
torch.cuda.synchronize()
|
31 |
gc.collect()
|
32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
# 상수 정의
|
34 |
MAX_SEED = 2**32 - 1
|
35 |
BASE_MODEL = "black-forest-labs/FLUX.1-dev"
|
@@ -194,13 +230,98 @@ def generate_fashion(prompt, mode, cfg_scale, steps, randomize_seed, seed, width
|
|
194 |
clear_memory() # 오류 발생 시에도 메모리 정리
|
195 |
raise e
|
196 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
197 |
def leffa_predict_vt(src_image_path, ref_image_path):
|
198 |
return leffa_predict(src_image_path, ref_image_path, "virtual_tryon")
|
199 |
|
200 |
def leffa_predict_pt(src_image_path, ref_image_path):
|
201 |
return leffa_predict(src_image_path, ref_image_path, "pose_transfer")
|
202 |
-
|
203 |
-
|
204 |
# Gradio 인터페이스
|
205 |
with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange") as demo:
|
206 |
gr.Markdown("# 🎭 Fashion Studio & Virtual Try-on")
|
|
|
18 |
# 메모리 관리 설정 추가
|
19 |
import torch.backends.cuda
|
20 |
torch.backends.cuda.max_split_size_mb = 128 # 메모리 분할 크기 제한
|
21 |
+
|
22 |
+
|
23 |
+
# 전역 변수로 모델들을 선언
|
24 |
+
fashion_pipe = None
|
25 |
+
translator = None
|
26 |
+
mask_predictor = None
|
27 |
+
densepose_predictor = None
|
28 |
+
vt_model = None
|
29 |
+
pt_model = None
|
30 |
+
vt_inference = None
|
31 |
+
pt_inference = None
|
32 |
+
|
33 |
+
# 초기화 함수
|
34 |
+
def initialize_models():
|
35 |
+
global fashion_pipe
|
36 |
+
if fashion_pipe is None:
|
37 |
+
fashion_pipe = DiffusionPipeline.from_pretrained(
|
38 |
+
BASE_MODEL,
|
39 |
+
torch_dtype=torch.float16,
|
40 |
+
use_auth_token=HF_TOKEN
|
41 |
+
)
|
42 |
+
fashion_pipe.to(device)
|
43 |
+
|
44 |
+
# 앱 시작 시 모델 초기화
|
45 |
+
initialize_models()
|
46 |
+
|
47 |
# 메모리 관리 설정
|
48 |
torch.cuda.empty_cache()
|
49 |
gc.collect()
|
|
|
56 |
torch.cuda.synchronize()
|
57 |
gc.collect()
|
58 |
|
59 |
+
# 모델 사용 후 메모리 해제
|
60 |
+
def unload_models():
|
61 |
+
global fashion_pipe, translator, mask_predictor, densepose_predictor, vt_model, pt_model
|
62 |
+
fashion_pipe = None
|
63 |
+
translator = None
|
64 |
+
mask_predictor = None
|
65 |
+
densepose_predictor = None
|
66 |
+
vt_model = None
|
67 |
+
pt_model = None
|
68 |
+
clear_memory()
|
69 |
# 상수 정의
|
70 |
MAX_SEED = 2**32 - 1
|
71 |
BASE_MODEL = "black-forest-labs/FLUX.1-dev"
|
|
|
230 |
clear_memory() # 오류 발생 시에도 메모리 정리
|
231 |
raise e
|
232 |
|
233 |
+
|
234 |
+
|
235 |
+
def leffa_predict(src_image_path, ref_image_path, control_type):
|
236 |
+
global mask_predictor, densepose_predictor, vt_model, pt_model, vt_inference, pt_inference
|
237 |
+
|
238 |
+
clear_memory()
|
239 |
+
|
240 |
+
try:
|
241 |
+
# 필요한 모델 초기화
|
242 |
+
if control_type == "virtual_tryon" and vt_model is None:
|
243 |
+
vt_model = LeffaModel(
|
244 |
+
pretrained_model_name_or_path="./ckpts/stable-diffusion-inpainting",
|
245 |
+
pretrained_model="./ckpts/virtual_tryon.pth"
|
246 |
+
)
|
247 |
+
vt_model.to(device)
|
248 |
+
vt_inference = LeffaInference(model=vt_model)
|
249 |
+
|
250 |
+
elif control_type == "pose_transfer" and pt_model is None:
|
251 |
+
pt_model = LeffaModel(
|
252 |
+
pretrained_model_name_or_path="./ckpts/stable-diffusion-xl-1.0-inpainting-0.1",
|
253 |
+
pretrained_model="./ckpts/pose_transfer.pth"
|
254 |
+
)
|
255 |
+
pt_model.to(device)
|
256 |
+
pt_inference = LeffaInference(model=pt_model)
|
257 |
+
|
258 |
+
if mask_predictor is None:
|
259 |
+
mask_predictor = AutoMasker(
|
260 |
+
densepose_path="./ckpts/densepose",
|
261 |
+
schp_path="./ckpts/schp",
|
262 |
+
)
|
263 |
+
|
264 |
+
if densepose_predictor is None:
|
265 |
+
densepose_predictor = DensePosePredictor(
|
266 |
+
config_path="./ckpts/densepose/densepose_rcnn_R_50_FPN_s1x.yaml",
|
267 |
+
weights_path="./ckpts/densepose/model_final_162be9.pkl",
|
268 |
+
)
|
269 |
+
|
270 |
+
# 이미지 처리
|
271 |
+
src_image = Image.open(src_image_path)
|
272 |
+
ref_image = Image.open(ref_image_path)
|
273 |
+
src_image = resize_and_center(src_image, 768, 1024)
|
274 |
+
ref_image = resize_and_center(ref_image, 768, 1024)
|
275 |
+
|
276 |
+
src_image_array = np.array(src_image)
|
277 |
+
ref_image_array = np.array(ref_image)
|
278 |
+
|
279 |
+
# Mask 생성
|
280 |
+
if control_type == "virtual_tryon":
|
281 |
+
src_image = src_image.convert("RGB")
|
282 |
+
mask = mask_predictor(src_image, "upper")["mask"]
|
283 |
+
else:
|
284 |
+
mask = Image.fromarray(np.ones_like(src_image_array) * 255)
|
285 |
+
|
286 |
+
# DensePose 예측
|
287 |
+
src_image_iuv_array = densepose_predictor.predict_iuv(src_image_array)
|
288 |
+
src_image_seg_array = densepose_predictor.predict_seg(src_image_array)
|
289 |
+
src_image_iuv = Image.fromarray(src_image_iuv_array)
|
290 |
+
src_image_seg = Image.fromarray(src_image_seg_array)
|
291 |
+
|
292 |
+
if control_type == "virtual_tryon":
|
293 |
+
densepose = src_image_seg
|
294 |
+
inference = vt_inference
|
295 |
+
else:
|
296 |
+
densepose = src_image_iuv
|
297 |
+
inference = pt_inference
|
298 |
+
|
299 |
+
# Leffa 변환 및 추론
|
300 |
+
transform = LeffaTransform()
|
301 |
+
data = {
|
302 |
+
"src_image": [src_image],
|
303 |
+
"ref_image": [ref_image],
|
304 |
+
"mask": [mask],
|
305 |
+
"densepose": [densepose],
|
306 |
+
}
|
307 |
+
data = transform(data)
|
308 |
+
|
309 |
+
output = inference(data)
|
310 |
+
gen_image = output["generated_image"][0]
|
311 |
+
|
312 |
+
clear_memory()
|
313 |
+
return np.array(gen_image)
|
314 |
+
|
315 |
+
except Exception as e:
|
316 |
+
clear_memory()
|
317 |
+
raise e
|
318 |
+
|
319 |
def leffa_predict_vt(src_image_path, ref_image_path):
|
320 |
return leffa_predict(src_image_path, ref_image_path, "virtual_tryon")
|
321 |
|
322 |
def leffa_predict_pt(src_image_path, ref_image_path):
|
323 |
return leffa_predict(src_image_path, ref_image_path, "pose_transfer")
|
324 |
+
|
|
|
325 |
# Gradio 인터페이스
|
326 |
with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange") as demo:
|
327 |
gr.Markdown("# 🎭 Fashion Studio & Virtual Try-on")
|