Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -23,12 +23,18 @@ from diffusers import FluxPipeline
|
|
23 |
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
|
24 |
|
25 |
import gc
|
|
|
26 |
|
|
|
27 |
def clear_memory():
|
28 |
"""๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ ํจ์"""
|
|
|
|
|
|
|
29 |
gc.collect()
|
30 |
-
|
31 |
-
|
|
|
32 |
|
33 |
|
34 |
model_name = "Helsinki-NLP/opus-mt-ko-en"
|
@@ -79,13 +85,17 @@ gd_model = GroundingDinoForObjectDetection.from_pretrained(gd_model_path, torch_
|
|
79 |
gd_model = gd_model.to(device=device)
|
80 |
assert isinstance(gd_model, GroundingDinoForObjectDetection)
|
81 |
|
82 |
-
# FLUX ํ์ดํ๋ผ์ธ ์ด๊ธฐํ
|
83 |
pipe = FluxPipeline.from_pretrained(
|
84 |
"black-forest-labs/FLUX.1-dev",
|
85 |
-
torch_dtype=torch.float16, #
|
86 |
use_auth_token=HF_TOKEN,
|
87 |
-
device_map="balanced"
|
88 |
)
|
|
|
|
|
|
|
|
|
89 |
pipe.load_lora_weights(
|
90 |
hf_hub_download(
|
91 |
"ByteDance/Hyper-SD",
|
@@ -95,7 +105,8 @@ pipe.load_lora_weights(
|
|
95 |
)
|
96 |
pipe.fuse_lora(lora_scale=0.125)
|
97 |
|
98 |
-
|
|
|
99 |
|
100 |
|
101 |
class timer:
|
@@ -171,37 +182,32 @@ def calculate_dimensions(aspect_ratio: str, base_size: int = 512) -> tuple[int,
|
|
171 |
return base_size, base_size
|
172 |
|
173 |
def generate_background(prompt: str, aspect_ratio: str) -> Image.Image:
|
174 |
-
"""๋ฐฐ๊ฒฝ ์ด๋ฏธ์ง ์์ฑ ํจ์"""
|
175 |
try:
|
176 |
-
# ์ ํ๋ ๋น์จ์ ๋ฐ๋ผ ํฌ๊ธฐ ๊ณ์ฐ
|
177 |
width, height = calculate_dimensions(aspect_ratio)
|
178 |
-
|
179 |
-
# 8์ ๋ฐฐ์๋ก ์กฐ์
|
180 |
width, height = adjust_size_to_multiple_of_8(width, height)
|
181 |
|
182 |
-
#
|
183 |
-
|
184 |
-
|
185 |
-
|
|
|
|
|
|
|
|
|
186 |
with timer("Background generation"):
|
187 |
-
|
188 |
image = pipe(
|
189 |
prompt=prompt,
|
190 |
width=width,
|
191 |
height=height,
|
192 |
num_inference_steps=8,
|
193 |
guidance_scale=4.0,
|
194 |
-
max_length=77,
|
195 |
).images[0]
|
196 |
-
except Exception as e:
|
197 |
-
print(f"Pipeline error: {str(e)}")
|
198 |
-
# ์ค๋ฅ ๋ฐ์ ์ ๊ธฐ๋ณธ ํฐ์ ๋ฐฐ๊ฒฝ ์์ฑ
|
199 |
-
image = Image.new('RGB', (width, height), 'white')
|
200 |
|
201 |
return image
|
202 |
except Exception as e:
|
203 |
print(f"Background generation error: {str(e)}")
|
204 |
-
# ์ตํ์ ํด๋ฐฑ: ๊ธฐ๋ณธ ํฐ์ ๋ฐฐ๊ฒฝ ๋ฐํ
|
205 |
return Image.new('RGB', (512, 512), 'white')
|
206 |
|
207 |
# FLUX ํ์ดํ๋ผ์ธ ์ด๊ธฐํ ๋ถ๋ถ ์์
|
@@ -296,34 +302,32 @@ def _gpu_process(img: Image.Image, prompt: str | BoundingBox | None) -> tuple[Im
|
|
296 |
|
297 |
def _process(img: Image.Image, prompt: str | BoundingBox | None, bg_prompt: str | None = None, aspect_ratio: str = "1:1") -> tuple[tuple[Image.Image, Image.Image, Image.Image], gr.DownloadButton]:
|
298 |
try:
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
|
|
|
|
|
|
|
|
305 |
|
306 |
-
|
307 |
-
|
|
|
308 |
|
309 |
if bg_prompt:
|
310 |
-
# ๋ฐฐ๊ฒฝ๋ง ์์ฑํ๏ฟฝ๏ฟฝ๏ฟฝ ํฉ์ฑ์ ํ์ง ์์
|
311 |
background = generate_background(bg_prompt, aspect_ratio)
|
312 |
-
combined = background
|
313 |
else:
|
314 |
combined = Image.alpha_composite(Image.new("RGBA", masked_alpha.size, "white"), masked_alpha)
|
315 |
|
316 |
-
|
317 |
-
bbox = thresholded.getbbox()
|
318 |
-
to_dl = masked_alpha.crop(bbox)
|
319 |
-
|
320 |
-
temp = tempfile.NamedTemporaryFile(delete=False, suffix=".png")
|
321 |
-
to_dl.save(temp, format="PNG")
|
322 |
-
temp.close()
|
323 |
|
324 |
return (img, combined, masked_alpha), gr.DownloadButton(value=temp.name, interactive=True)
|
325 |
-
|
326 |
except Exception as e:
|
|
|
327 |
raise gr.Error(f"Processing failed: {str(e)}")
|
328 |
|
329 |
def on_change_bbox(prompts: dict[str, Any] | None):
|
@@ -683,12 +687,14 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
683 |
</div>
|
684 |
</div>
|
685 |
""")
|
686 |
-
|
|
|
687 |
demo.launch(
|
688 |
server_name="0.0.0.0",
|
689 |
server_port=7860,
|
690 |
share=False,
|
691 |
enable_queue=True,
|
692 |
-
max_threads=
|
693 |
-
allowed_paths=["examples"]
|
|
|
694 |
)
|
|
|
23 |
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
|
24 |
|
25 |
import gc
|
26 |
+
import torch.cuda.amp as amp
|
27 |
|
28 |
+
# ๋ฉ๋ชจ๋ฆฌ ๊ด๋ฆฌ ํจ์ ๊ฐํ
|
29 |
def clear_memory():
|
30 |
"""๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ ํจ์"""
|
31 |
+
if torch.cuda.is_available():
|
32 |
+
torch.cuda.empty_cache()
|
33 |
+
torch.cuda.synchronize()
|
34 |
gc.collect()
|
35 |
+
|
36 |
+
# ์๋ ํผํฉ ์ ๋ฐ๋(Automatic Mixed Precision) ์ค์
|
37 |
+
scaler = amp.GradScaler()
|
38 |
|
39 |
|
40 |
model_name = "Helsinki-NLP/opus-mt-ko-en"
|
|
|
85 |
gd_model = gd_model.to(device=device)
|
86 |
assert isinstance(gd_model, GroundingDinoForObjectDetection)
|
87 |
|
88 |
+
# FLUX ํ์ดํ๋ผ์ธ ์ด๊ธฐํ
|
89 |
pipe = FluxPipeline.from_pretrained(
|
90 |
"black-forest-labs/FLUX.1-dev",
|
91 |
+
torch_dtype=torch.float16, # A100์ ์ต์ ํ๋ float16 ์ฌ์ฉ
|
92 |
use_auth_token=HF_TOKEN,
|
93 |
+
device_map="balanced"
|
94 |
)
|
95 |
+
pipe.enable_attention_slicing(slice_size="auto") # ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋ ์ต์ ํ
|
96 |
+
pipe.enable_sequential_cpu_offload() # CPU ์คํ๋ก๋ฉ ํ์ฑํ
|
97 |
+
|
98 |
+
|
99 |
pipe.load_lora_weights(
|
100 |
hf_hub_download(
|
101 |
"ByteDance/Hyper-SD",
|
|
|
105 |
)
|
106 |
pipe.fuse_lora(lora_scale=0.125)
|
107 |
|
108 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = "0" # ๋จ์ผ GPU ์ฌ์ฉ
|
109 |
+
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:512" # CUDA ๋ฉ๋ชจ๋ฆฌ ํ ๋น ์ค์
|
110 |
|
111 |
|
112 |
class timer:
|
|
|
182 |
return base_size, base_size
|
183 |
|
184 |
def generate_background(prompt: str, aspect_ratio: str) -> Image.Image:
|
|
|
185 |
try:
|
|
|
186 |
width, height = calculate_dimensions(aspect_ratio)
|
|
|
|
|
187 |
width, height = adjust_size_to_multiple_of_8(width, height)
|
188 |
|
189 |
+
# A100 ๋ฉ๋ชจ๋ฆฌ ์ ํ์ ๊ณ ๋ คํ ์ต๋ ํฌ๊ธฐ ์ค์
|
190 |
+
max_size = 768
|
191 |
+
if width > max_size or height > max_size:
|
192 |
+
ratio = max_size / max(width, height)
|
193 |
+
width = int(width * ratio)
|
194 |
+
height = int(height * ratio)
|
195 |
+
width, height = adjust_size_to_multiple_of_8(width, height)
|
196 |
+
|
197 |
with timer("Background generation"):
|
198 |
+
with torch.cuda.amp.autocast(): # ์๋ ํผํฉ ์ ๋ฐ๋ ์ฌ์ฉ
|
199 |
image = pipe(
|
200 |
prompt=prompt,
|
201 |
width=width,
|
202 |
height=height,
|
203 |
num_inference_steps=8,
|
204 |
guidance_scale=4.0,
|
205 |
+
max_length=77,
|
206 |
).images[0]
|
|
|
|
|
|
|
|
|
207 |
|
208 |
return image
|
209 |
except Exception as e:
|
210 |
print(f"Background generation error: {str(e)}")
|
|
|
211 |
return Image.new('RGB', (512, 512), 'white')
|
212 |
|
213 |
# FLUX ํ์ดํ๋ผ์ธ ์ด๊ธฐํ ๋ถ๋ถ ์์
|
|
|
302 |
|
303 |
def _process(img: Image.Image, prompt: str | BoundingBox | None, bg_prompt: str | None = None, aspect_ratio: str = "1:1") -> tuple[tuple[Image.Image, Image.Image, Image.Image], gr.DownloadButton]:
|
304 |
try:
|
305 |
+
# ์
๋ ฅ ์ด๋ฏธ์ง ํฌ๊ธฐ ์ ํ
|
306 |
+
max_size = 1024
|
307 |
+
if img.width > max_size or img.height > max_size:
|
308 |
+
ratio = max_size / max(img.width, img.height)
|
309 |
+
new_size = (int(img.width * ratio), int(img.height * ratio))
|
310 |
+
img = img.resize(new_size, Image.LANCZOS)
|
311 |
+
|
312 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋ ๋ชจ๋ํฐ๋ง
|
313 |
+
if torch.cuda.is_available():
|
314 |
+
torch.cuda.reset_peak_memory_stats()
|
315 |
|
316 |
+
with torch.cuda.amp.autocast():
|
317 |
+
mask, bbox, time_log = _gpu_process(img, prompt)
|
318 |
+
masked_alpha = apply_mask(img, mask, defringe=True)
|
319 |
|
320 |
if bg_prompt:
|
|
|
321 |
background = generate_background(bg_prompt, aspect_ratio)
|
322 |
+
combined = background
|
323 |
else:
|
324 |
combined = Image.alpha_composite(Image.new("RGBA", masked_alpha.size, "white"), masked_alpha)
|
325 |
|
326 |
+
clear_memory() # ์ค๊ฐ ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
|
|
|
|
|
|
|
|
|
|
|
|
327 |
|
328 |
return (img, combined, masked_alpha), gr.DownloadButton(value=temp.name, interactive=True)
|
|
|
329 |
except Exception as e:
|
330 |
+
clear_memory()
|
331 |
raise gr.Error(f"Processing failed: {str(e)}")
|
332 |
|
333 |
def on_change_bbox(prompts: dict[str, Any] | None):
|
|
|
687 |
</div>
|
688 |
</div>
|
689 |
""")
|
690 |
+
|
691 |
+
demo.queue(max_size=10) # ํ ํฌ๊ธฐ ์ ํ
|
692 |
demo.launch(
|
693 |
server_name="0.0.0.0",
|
694 |
server_port=7860,
|
695 |
share=False,
|
696 |
enable_queue=True,
|
697 |
+
max_threads=2, # ์ค๋ ๋ ์ ์ ํ
|
698 |
+
allowed_paths=["examples"],
|
699 |
+
memory_limit=0.8 # ๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋ ์ ํ (80%)
|
700 |
)
|