Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -3,6 +3,7 @@ import time
|
|
3 |
from collections.abc import Sequence
|
4 |
from typing import Any, cast
|
5 |
import os
|
|
|
6 |
from huggingface_hub import login, hf_hub_download
|
7 |
|
8 |
import gradio as gr
|
@@ -19,35 +20,36 @@ from refiners.solutions import BoxSegmenter
|
|
19 |
from transformers import GroundingDinoForObjectDetection, GroundingDinoProcessor
|
20 |
from diffusers import FluxPipeline
|
21 |
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
|
22 |
-
import gc
|
23 |
|
|
|
|
|
24 |
def clear_memory():
|
25 |
-
"""메모리 정리 함수"""
|
26 |
gc.collect()
|
27 |
try:
|
28 |
if torch.cuda.is_available():
|
29 |
with torch.cuda.device(0): # 명시적으로 device 0 사용
|
30 |
torch.cuda.empty_cache()
|
31 |
-
except:
|
32 |
pass
|
33 |
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
# GPU 설정을 try-except로 감싸기
|
38 |
if torch.cuda.is_available():
|
39 |
try:
|
40 |
with torch.cuda.device(0):
|
41 |
torch.cuda.empty_cache()
|
42 |
torch.backends.cudnn.benchmark = True
|
43 |
torch.backends.cuda.matmul.allow_tf32 = True
|
44 |
-
except:
|
45 |
print("Warning: Could not configure CUDA settings")
|
46 |
|
47 |
-
|
|
|
48 |
model_name = "Helsinki-NLP/opus-mt-ko-en"
|
49 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
50 |
-
|
|
|
51 |
translator = pipeline("translation", model=model, tokenizer=tokenizer, device=-1)
|
52 |
|
53 |
def translate_to_english(text: str) -> str:
|
@@ -67,6 +69,7 @@ BoundingBox = tuple[int, int, int, int]
|
|
67 |
pillow_heif.register_heif_opener()
|
68 |
pillow_heif.register_avif_opener()
|
69 |
|
|
|
70 |
# HF 토큰 설정
|
71 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
72 |
if HF_TOKEN is None:
|
@@ -77,7 +80,8 @@ try:
|
|
77 |
except Exception as e:
|
78 |
raise ValueError(f"Failed to login to Hugging Face: {str(e)}")
|
79 |
|
80 |
-
|
|
|
81 |
segmenter = BoxSegmenter(device="cpu")
|
82 |
segmenter.device = device
|
83 |
segmenter.model = segmenter.model.to(device=segmenter.device)
|
@@ -88,15 +92,14 @@ gd_model = GroundingDinoForObjectDetection.from_pretrained(gd_model_path, torch_
|
|
88 |
gd_model = gd_model.to(device=device)
|
89 |
assert isinstance(gd_model, GroundingDinoForObjectDetection)
|
90 |
|
91 |
-
|
|
|
92 |
pipe = FluxPipeline.from_pretrained(
|
93 |
"black-forest-labs/FLUX.1-dev",
|
94 |
torch_dtype=torch.float16,
|
95 |
use_auth_token=HF_TOKEN
|
96 |
)
|
97 |
pipe.enable_attention_slicing(slice_size="auto")
|
98 |
-
|
99 |
-
# LoRA 가중치 로드
|
100 |
pipe.load_lora_weights(
|
101 |
hf_hub_download(
|
102 |
"ByteDance/Hyper-SD",
|
@@ -105,14 +108,14 @@ pipe.load_lora_weights(
|
|
105 |
)
|
106 |
)
|
107 |
pipe.fuse_lora(lora_scale=0.125)
|
108 |
-
|
109 |
-
# GPU 설정을 try-except로 감싸기
|
110 |
try:
|
111 |
if torch.cuda.is_available():
|
112 |
-
pipe = pipe.to("cuda:0") # 명시적으로 cuda:0
|
113 |
except Exception as e:
|
114 |
print(f"Warning: Could not move pipeline to CUDA: {str(e)}")
|
115 |
|
|
|
|
|
116 |
class timer:
|
117 |
def __init__(self, method_name="timed process"):
|
118 |
self.method = method_name
|
@@ -123,6 +126,8 @@ class timer:
|
|
123 |
end = time.time()
|
124 |
print(f"{self.method} took {str(round(end - self.start, 2))}s")
|
125 |
|
|
|
|
|
126 |
def bbox_union(bboxes: Sequence[list[int]]) -> BoundingBox | None:
|
127 |
if not bboxes:
|
128 |
return None
|
@@ -166,15 +171,12 @@ def apply_mask(img: Image.Image, mask_img: Image.Image, defringe: bool = True) -
|
|
166 |
result.paste(img, (0, 0), mask_img)
|
167 |
return result
|
168 |
|
169 |
-
|
170 |
def adjust_size_to_multiple_of_8(width: int, height: int) -> tuple[int, int]:
|
171 |
-
"""이미지 크기를 8의 배수로 조정하는 함수"""
|
172 |
new_width = ((width + 7) // 8) * 8
|
173 |
new_height = ((height + 7) // 8) * 8
|
174 |
return new_width, new_height
|
175 |
|
176 |
def calculate_dimensions(aspect_ratio: str, base_size: int = 512) -> tuple[int, int]:
|
177 |
-
"""선택된 비율에 따라 이미지 크기 계산"""
|
178 |
if aspect_ratio == "1:1":
|
179 |
return base_size, base_size
|
180 |
elif aspect_ratio == "16:9":
|
@@ -185,7 +187,9 @@ def calculate_dimensions(aspect_ratio: str, base_size: int = 512) -> tuple[int,
|
|
185 |
return base_size * 4 // 3, base_size
|
186 |
return base_size, base_size
|
187 |
|
188 |
-
|
|
|
|
|
189 |
def generate_background(prompt: str, aspect_ratio: str) -> Image.Image:
|
190 |
try:
|
191 |
width, height = calculate_dimensions(aspect_ratio)
|
@@ -197,7 +201,7 @@ def generate_background(prompt: str, aspect_ratio: str) -> Image.Image:
|
|
197 |
width = int(width * ratio)
|
198 |
height = int(height * ratio)
|
199 |
width, height = adjust_size_to_multiple_of_8(width, height)
|
200 |
-
|
201 |
with timer("Background generation"):
|
202 |
try:
|
203 |
with torch.inference_mode():
|
@@ -211,7 +215,6 @@ def generate_background(prompt: str, aspect_ratio: str) -> Image.Image:
|
|
211 |
except Exception as e:
|
212 |
print(f"Pipeline error: {str(e)}")
|
213 |
return Image.new('RGB', (width, height), 'white')
|
214 |
-
|
215 |
return image
|
216 |
except Exception as e:
|
217 |
print(f"Background generation error: {str(e)}")
|
@@ -233,7 +236,6 @@ def create_position_grid():
|
|
233 |
"""
|
234 |
|
235 |
def calculate_object_position(position: str, bg_size: tuple[int, int], obj_size: tuple[int, int]) -> tuple[int, int]:
|
236 |
-
"""오브젝트의 위치 계산"""
|
237 |
bg_width, bg_height = bg_size
|
238 |
obj_width, obj_height = obj_size
|
239 |
|
@@ -252,28 +254,21 @@ def calculate_object_position(position: str, bg_size: tuple[int, int], obj_size:
|
|
252 |
return positions.get(position, positions["bottom-center"])
|
253 |
|
254 |
def resize_object(image: Image.Image, scale_percent: float) -> Image.Image:
|
255 |
-
"""오브젝트 크기 조정"""
|
256 |
width = int(image.width * scale_percent / 100)
|
257 |
height = int(image.height * scale_percent / 100)
|
258 |
return image.resize((width, height), Image.Resampling.LANCZOS)
|
259 |
|
260 |
def combine_with_background(foreground: Image.Image, background: Image.Image,
|
261 |
-
|
262 |
-
"""전경과 배경 합성 함수"""
|
263 |
-
# 배경 이미지 준비
|
264 |
result = background.convert('RGBA')
|
265 |
-
|
266 |
-
# 오브젝트 크기 조정
|
267 |
scaled_foreground = resize_object(foreground, scale_percent)
|
268 |
-
|
269 |
-
# 오브젝트 위치 계산
|
270 |
x, y = calculate_object_position(position, result.size, scaled_foreground.size)
|
271 |
-
|
272 |
-
# 합성
|
273 |
result.paste(scaled_foreground, (x, y), scaled_foreground)
|
274 |
return result
|
275 |
|
276 |
-
|
|
|
|
|
277 |
def _gpu_process(img: Image.Image, prompt: str | BoundingBox | None) -> tuple[Image.Image, BoundingBox | None, list[str]]:
|
278 |
time_log: list[str] = []
|
279 |
try:
|
@@ -294,6 +289,8 @@ def _gpu_process(img: Image.Image, prompt: str | BoundingBox | None) -> tuple[Im
|
|
294 |
print(f"GPU process error: {str(e)}")
|
295 |
raise
|
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 |
# 입력 이미지 크기 제한
|
@@ -302,8 +299,7 @@ def _process(img: Image.Image, prompt: str | BoundingBox | None, bg_prompt: str
|
|
302 |
ratio = max_size / max(img.width, img.height)
|
303 |
new_size = (int(img.width * ratio), int(img.height * ratio))
|
304 |
img = img.resize(new_size, Image.LANCZOS)
|
305 |
-
|
306 |
-
# CUDA 메모리 관리 수정
|
307 |
try:
|
308 |
if torch.cuda.is_available():
|
309 |
current_device = torch.cuda.current_device()
|
@@ -311,19 +307,19 @@ def _process(img: Image.Image, prompt: str | BoundingBox | None, bg_prompt: str
|
|
311 |
torch.cuda.empty_cache()
|
312 |
except Exception as e:
|
313 |
print(f"CUDA memory management failed: {e}")
|
314 |
-
|
315 |
with torch.cuda.amp.autocast(enabled=torch.cuda.is_available()):
|
316 |
mask, bbox, time_log = _gpu_process(img, prompt)
|
317 |
masked_alpha = apply_mask(img, mask, defringe=True)
|
318 |
-
|
319 |
if bg_prompt:
|
320 |
background = generate_background(bg_prompt, aspect_ratio)
|
321 |
combined = background
|
322 |
else:
|
323 |
combined = Image.alpha_composite(Image.new("RGBA", masked_alpha.size, "white"), masked_alpha)
|
324 |
-
|
325 |
clear_memory()
|
326 |
-
|
327 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp:
|
328 |
combined.save(temp.name)
|
329 |
return (img, combined, masked_alpha), gr.DownloadButton(value=temp.name, interactive=True)
|
@@ -335,15 +331,12 @@ def _process(img: Image.Image, prompt: str | BoundingBox | None, bg_prompt: str
|
|
335 |
def on_change_bbox(prompts: dict[str, Any] | None):
|
336 |
return gr.update(interactive=prompts is not None)
|
337 |
|
338 |
-
|
339 |
def on_change_prompt(img: Image.Image | None, prompt: str | None, bg_prompt: str | None = None):
|
340 |
return gr.update(interactive=bool(img and prompt))
|
341 |
|
342 |
-
|
343 |
-
|
344 |
def process_prompt(img: Image.Image, prompt: str, bg_prompt: str | None = None,
|
345 |
-
|
346 |
-
|
347 |
try:
|
348 |
if img is None or prompt.strip() == "":
|
349 |
raise gr.Error("Please provide both image and prompt")
|
@@ -379,7 +372,7 @@ def process_prompt(img: Image.Image, prompt: str, bg_prompt: str | None = None,
|
|
379 |
raise gr.Error(str(e))
|
380 |
finally:
|
381 |
clear_memory()
|
382 |
-
|
383 |
def process_bbox(img: Image.Image, box_input: str) -> tuple[Image.Image, Image.Image]:
|
384 |
try:
|
385 |
if img is None or box_input.strip() == "":
|
@@ -393,15 +386,11 @@ def process_bbox(img: Image.Image, box_input: str) -> tuple[Image.Image, Image.I
|
|
393 |
except:
|
394 |
raise gr.Error("Invalid box format. Please provide [xmin, ymin, xmax, ymax]")
|
395 |
|
396 |
-
# Process the image
|
397 |
results, _ = _process(img, bbox)
|
398 |
-
|
399 |
-
# 합성된 이미지와 추출된 이미지만 반환
|
400 |
return results[1], results[2]
|
401 |
except Exception as e:
|
402 |
raise gr.Error(str(e))
|
403 |
|
404 |
-
# Event handler functions 수정
|
405 |
def update_process_button(img, prompt):
|
406 |
return gr.update(
|
407 |
interactive=bool(img and prompt),
|
@@ -418,7 +407,7 @@ def update_box_button(img, box_input):
|
|
418 |
except:
|
419 |
return gr.update(interactive=False, variant="secondary")
|
420 |
|
421 |
-
|
422 |
# CSS 정의
|
423 |
css = """
|
424 |
footer {display: none}
|
@@ -482,9 +471,7 @@ button.primary:hover {
|
|
482 |
}
|
483 |
"""
|
484 |
|
485 |
-
|
486 |
-
# UI 구성 부분에서 process_btn을 위로 이동하고 position_grid.click 부분 제거
|
487 |
-
|
488 |
# UI 구성
|
489 |
with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
490 |
gr.HTML("""
|
@@ -493,7 +480,6 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
493 |
<p>AI Integrated Image Creator: Extract objects, generate backgrounds, and adjust ratios and positions to create complete images with AI.</p>
|
494 |
</div>
|
495 |
""")
|
496 |
-
|
497 |
with gr.Row():
|
498 |
with gr.Column(scale=1):
|
499 |
input_image = gr.Image(
|
@@ -521,7 +507,6 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
521 |
visible=True,
|
522 |
scale=1
|
523 |
)
|
524 |
-
|
525 |
with gr.Row(visible=False) as object_controls:
|
526 |
with gr.Column(scale=1):
|
527 |
with gr.Row():
|
@@ -545,17 +530,14 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
545 |
step=5,
|
546 |
label="Object Size (%)"
|
547 |
)
|
548 |
-
|
549 |
process_btn = gr.Button(
|
550 |
"Process",
|
551 |
variant="primary",
|
552 |
interactive=False
|
553 |
)
|
554 |
-
|
555 |
# 각 버튼에 대한 클릭 이벤트 처리
|
556 |
def update_position(new_position):
|
557 |
return new_position
|
558 |
-
|
559 |
btn_top_left.click(fn=lambda: update_position("top-left"), outputs=position)
|
560 |
btn_top_center.click(fn=lambda: update_position("top-center"), outputs=position)
|
561 |
btn_top_right.click(fn=lambda: update_position("top-right"), outputs=position)
|
@@ -565,7 +547,6 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
565 |
btn_bottom_left.click(fn=lambda: update_position("bottom-left"), outputs=position)
|
566 |
btn_bottom_center.click(fn=lambda: update_position("bottom-center"), outputs=position)
|
567 |
btn_bottom_right.click(fn=lambda: update_position("bottom-right"), outputs=position)
|
568 |
-
|
569 |
with gr.Column(scale=1):
|
570 |
with gr.Row():
|
571 |
combined_image = gr.Image(
|
@@ -581,7 +562,6 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
581 |
type="pil",
|
582 |
height=256
|
583 |
)
|
584 |
-
|
585 |
# Event bindings
|
586 |
input_image.change(
|
587 |
fn=update_process_button,
|
@@ -589,29 +569,24 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
589 |
outputs=process_btn,
|
590 |
queue=False
|
591 |
)
|
592 |
-
|
593 |
text_prompt.change(
|
594 |
fn=update_process_button,
|
595 |
inputs=[input_image, text_prompt],
|
596 |
outputs=process_btn,
|
597 |
queue=False
|
598 |
)
|
599 |
-
|
600 |
def update_controls(bg_prompt):
|
601 |
-
"""배경 프롬프트 입력 여부에 따라 컨트롤 표시 업데이트"""
|
602 |
is_visible = bool(bg_prompt)
|
603 |
return [
|
604 |
-
gr.update(visible=is_visible),
|
605 |
-
gr.update(visible=is_visible),
|
606 |
]
|
607 |
-
|
608 |
bg_prompt.change(
|
609 |
fn=update_controls,
|
610 |
inputs=bg_prompt,
|
611 |
outputs=[aspect_ratio, object_controls],
|
612 |
queue=False
|
613 |
)
|
614 |
-
|
615 |
process_btn.click(
|
616 |
fn=process_prompt,
|
617 |
inputs=[
|
@@ -625,12 +600,23 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
|
625 |
outputs=[combined_image, extracted_image],
|
626 |
queue=True
|
627 |
)
|
628 |
-
|
629 |
-
|
630 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
631 |
demo.launch(
|
632 |
server_name="0.0.0.0",
|
633 |
server_port=7860,
|
634 |
share=False,
|
635 |
-
max_threads=2
|
636 |
)
|
|
|
3 |
from collections.abc import Sequence
|
4 |
from typing import Any, cast
|
5 |
import os
|
6 |
+
import gc
|
7 |
from huggingface_hub import login, hf_hub_download
|
8 |
|
9 |
import gradio as gr
|
|
|
20 |
from transformers import GroundingDinoForObjectDetection, GroundingDinoProcessor
|
21 |
from diffusers import FluxPipeline
|
22 |
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
|
|
|
23 |
|
24 |
+
#############################################################
|
25 |
+
# 메모리 정리 함수
|
26 |
def clear_memory():
|
|
|
27 |
gc.collect()
|
28 |
try:
|
29 |
if torch.cuda.is_available():
|
30 |
with torch.cuda.device(0): # 명시적으로 device 0 사용
|
31 |
torch.cuda.empty_cache()
|
32 |
+
except Exception as e:
|
33 |
pass
|
34 |
|
35 |
+
#############################################################
|
36 |
+
# GPU 설정 (Zero GPU 환경)
|
37 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
|
|
38 |
if torch.cuda.is_available():
|
39 |
try:
|
40 |
with torch.cuda.device(0):
|
41 |
torch.cuda.empty_cache()
|
42 |
torch.backends.cudnn.benchmark = True
|
43 |
torch.backends.cuda.matmul.allow_tf32 = True
|
44 |
+
except Exception as e:
|
45 |
print("Warning: Could not configure CUDA settings")
|
46 |
|
47 |
+
#############################################################
|
48 |
+
# 번역 모델 초기화 (CPU에서 동작)
|
49 |
model_name = "Helsinki-NLP/opus-mt-ko-en"
|
50 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
51 |
+
# 번역 모델은 CPU에 올림
|
52 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to("cpu")
|
53 |
translator = pipeline("translation", model=model, tokenizer=tokenizer, device=-1)
|
54 |
|
55 |
def translate_to_english(text: str) -> str:
|
|
|
69 |
pillow_heif.register_heif_opener()
|
70 |
pillow_heif.register_avif_opener()
|
71 |
|
72 |
+
#############################################################
|
73 |
# HF 토큰 설정
|
74 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
75 |
if HF_TOKEN is None:
|
|
|
80 |
except Exception as e:
|
81 |
raise ValueError(f"Failed to login to Hugging Face: {str(e)}")
|
82 |
|
83 |
+
#############################################################
|
84 |
+
# 객체 분할 모델 초기화
|
85 |
segmenter = BoxSegmenter(device="cpu")
|
86 |
segmenter.device = device
|
87 |
segmenter.model = segmenter.model.to(device=segmenter.device)
|
|
|
92 |
gd_model = gd_model.to(device=device)
|
93 |
assert isinstance(gd_model, GroundingDinoForObjectDetection)
|
94 |
|
95 |
+
#############################################################
|
96 |
+
# FLUX 파이프라인 초기화 (Zero GPU용)
|
97 |
pipe = FluxPipeline.from_pretrained(
|
98 |
"black-forest-labs/FLUX.1-dev",
|
99 |
torch_dtype=torch.float16,
|
100 |
use_auth_token=HF_TOKEN
|
101 |
)
|
102 |
pipe.enable_attention_slicing(slice_size="auto")
|
|
|
|
|
103 |
pipe.load_lora_weights(
|
104 |
hf_hub_download(
|
105 |
"ByteDance/Hyper-SD",
|
|
|
108 |
)
|
109 |
)
|
110 |
pipe.fuse_lora(lora_scale=0.125)
|
|
|
|
|
111 |
try:
|
112 |
if torch.cuda.is_available():
|
113 |
+
pipe = pipe.to("cuda:0") # 명시적으로 cuda:0로 이동
|
114 |
except Exception as e:
|
115 |
print(f"Warning: Could not move pipeline to CUDA: {str(e)}")
|
116 |
|
117 |
+
#############################################################
|
118 |
+
# 타이머 클래스
|
119 |
class timer:
|
120 |
def __init__(self, method_name="timed process"):
|
121 |
self.method = method_name
|
|
|
126 |
end = time.time()
|
127 |
print(f"{self.method} took {str(round(end - self.start, 2))}s")
|
128 |
|
129 |
+
#############################################################
|
130 |
+
# 유틸리티 함수들
|
131 |
def bbox_union(bboxes: Sequence[list[int]]) -> BoundingBox | None:
|
132 |
if not bboxes:
|
133 |
return None
|
|
|
171 |
result.paste(img, (0, 0), mask_img)
|
172 |
return result
|
173 |
|
|
|
174 |
def adjust_size_to_multiple_of_8(width: int, height: int) -> tuple[int, int]:
|
|
|
175 |
new_width = ((width + 7) // 8) * 8
|
176 |
new_height = ((height + 7) // 8) * 8
|
177 |
return new_width, new_height
|
178 |
|
179 |
def calculate_dimensions(aspect_ratio: str, base_size: int = 512) -> tuple[int, int]:
|
|
|
180 |
if aspect_ratio == "1:1":
|
181 |
return base_size, base_size
|
182 |
elif aspect_ratio == "16:9":
|
|
|
187 |
return base_size * 4 // 3, base_size
|
188 |
return base_size, base_size
|
189 |
|
190 |
+
#############################################################
|
191 |
+
# 배경 생성 함수 (Zero GPU에 맞게 수정)
|
192 |
+
@spaces.GPU(duration=20)
|
193 |
def generate_background(prompt: str, aspect_ratio: str) -> Image.Image:
|
194 |
try:
|
195 |
width, height = calculate_dimensions(aspect_ratio)
|
|
|
201 |
width = int(width * ratio)
|
202 |
height = int(height * ratio)
|
203 |
width, height = adjust_size_to_multiple_of_8(width, height)
|
204 |
+
|
205 |
with timer("Background generation"):
|
206 |
try:
|
207 |
with torch.inference_mode():
|
|
|
215 |
except Exception as e:
|
216 |
print(f"Pipeline error: {str(e)}")
|
217 |
return Image.new('RGB', (width, height), 'white')
|
|
|
218 |
return image
|
219 |
except Exception as e:
|
220 |
print(f"Background generation error: {str(e)}")
|
|
|
236 |
"""
|
237 |
|
238 |
def calculate_object_position(position: str, bg_size: tuple[int, int], obj_size: tuple[int, int]) -> tuple[int, int]:
|
|
|
239 |
bg_width, bg_height = bg_size
|
240 |
obj_width, obj_height = obj_size
|
241 |
|
|
|
254 |
return positions.get(position, positions["bottom-center"])
|
255 |
|
256 |
def resize_object(image: Image.Image, scale_percent: float) -> Image.Image:
|
|
|
257 |
width = int(image.width * scale_percent / 100)
|
258 |
height = int(image.height * scale_percent / 100)
|
259 |
return image.resize((width, height), Image.Resampling.LANCZOS)
|
260 |
|
261 |
def combine_with_background(foreground: Image.Image, background: Image.Image,
|
262 |
+
position: str = "bottom-center", scale_percent: float = 100) -> Image.Image:
|
|
|
|
|
263 |
result = background.convert('RGBA')
|
|
|
|
|
264 |
scaled_foreground = resize_object(foreground, scale_percent)
|
|
|
|
|
265 |
x, y = calculate_object_position(position, result.size, scaled_foreground.size)
|
|
|
|
|
266 |
result.paste(scaled_foreground, (x, y), scaled_foreground)
|
267 |
return result
|
268 |
|
269 |
+
#############################################################
|
270 |
+
# GPU 처리 함수 (Zero GPU에 맞게 수정)
|
271 |
+
@spaces.GPU(duration=30)
|
272 |
def _gpu_process(img: Image.Image, prompt: str | BoundingBox | None) -> tuple[Image.Image, BoundingBox | None, list[str]]:
|
273 |
time_log: list[str] = []
|
274 |
try:
|
|
|
289 |
print(f"GPU process error: {str(e)}")
|
290 |
raise
|
291 |
|
292 |
+
#############################################################
|
293 |
+
# 전체 처리 함수
|
294 |
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]:
|
295 |
try:
|
296 |
# 입력 이미지 크기 제한
|
|
|
299 |
ratio = max_size / max(img.width, img.height)
|
300 |
new_size = (int(img.width * ratio), int(img.height * ratio))
|
301 |
img = img.resize(new_size, Image.LANCZOS)
|
302 |
+
|
|
|
303 |
try:
|
304 |
if torch.cuda.is_available():
|
305 |
current_device = torch.cuda.current_device()
|
|
|
307 |
torch.cuda.empty_cache()
|
308 |
except Exception as e:
|
309 |
print(f"CUDA memory management failed: {e}")
|
310 |
+
|
311 |
with torch.cuda.amp.autocast(enabled=torch.cuda.is_available()):
|
312 |
mask, bbox, time_log = _gpu_process(img, prompt)
|
313 |
masked_alpha = apply_mask(img, mask, defringe=True)
|
314 |
+
|
315 |
if bg_prompt:
|
316 |
background = generate_background(bg_prompt, aspect_ratio)
|
317 |
combined = background
|
318 |
else:
|
319 |
combined = Image.alpha_composite(Image.new("RGBA", masked_alpha.size, "white"), masked_alpha)
|
320 |
+
|
321 |
clear_memory()
|
322 |
+
|
323 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp:
|
324 |
combined.save(temp.name)
|
325 |
return (img, combined, masked_alpha), gr.DownloadButton(value=temp.name, interactive=True)
|
|
|
331 |
def on_change_bbox(prompts: dict[str, Any] | None):
|
332 |
return gr.update(interactive=prompts is not None)
|
333 |
|
|
|
334 |
def on_change_prompt(img: Image.Image | None, prompt: str | None, bg_prompt: str | None = None):
|
335 |
return gr.update(interactive=bool(img and prompt))
|
336 |
|
|
|
|
|
337 |
def process_prompt(img: Image.Image, prompt: str, bg_prompt: str | None = None,
|
338 |
+
aspect_ratio: str = "1:1", position: str = "bottom-center",
|
339 |
+
scale_percent: float = 100) -> tuple[Image.Image, Image.Image]:
|
340 |
try:
|
341 |
if img is None or prompt.strip() == "":
|
342 |
raise gr.Error("Please provide both image and prompt")
|
|
|
372 |
raise gr.Error(str(e))
|
373 |
finally:
|
374 |
clear_memory()
|
375 |
+
|
376 |
def process_bbox(img: Image.Image, box_input: str) -> tuple[Image.Image, Image.Image]:
|
377 |
try:
|
378 |
if img is None or box_input.strip() == "":
|
|
|
386 |
except:
|
387 |
raise gr.Error("Invalid box format. Please provide [xmin, ymin, xmax, ymax]")
|
388 |
|
|
|
389 |
results, _ = _process(img, bbox)
|
|
|
|
|
390 |
return results[1], results[2]
|
391 |
except Exception as e:
|
392 |
raise gr.Error(str(e))
|
393 |
|
|
|
394 |
def update_process_button(img, prompt):
|
395 |
return gr.update(
|
396 |
interactive=bool(img and prompt),
|
|
|
407 |
except:
|
408 |
return gr.update(interactive=False, variant="secondary")
|
409 |
|
410 |
+
#############################################################
|
411 |
# CSS 정의
|
412 |
css = """
|
413 |
footer {display: none}
|
|
|
471 |
}
|
472 |
"""
|
473 |
|
474 |
+
#############################################################
|
|
|
|
|
475 |
# UI 구성
|
476 |
with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
|
477 |
gr.HTML("""
|
|
|
480 |
<p>AI Integrated Image Creator: Extract objects, generate backgrounds, and adjust ratios and positions to create complete images with AI.</p>
|
481 |
</div>
|
482 |
""")
|
|
|
483 |
with gr.Row():
|
484 |
with gr.Column(scale=1):
|
485 |
input_image = gr.Image(
|
|
|
507 |
visible=True,
|
508 |
scale=1
|
509 |
)
|
|
|
510 |
with gr.Row(visible=False) as object_controls:
|
511 |
with gr.Column(scale=1):
|
512 |
with gr.Row():
|
|
|
530 |
step=5,
|
531 |
label="Object Size (%)"
|
532 |
)
|
|
|
533 |
process_btn = gr.Button(
|
534 |
"Process",
|
535 |
variant="primary",
|
536 |
interactive=False
|
537 |
)
|
|
|
538 |
# 각 버튼에 대한 클릭 이벤트 처리
|
539 |
def update_position(new_position):
|
540 |
return new_position
|
|
|
541 |
btn_top_left.click(fn=lambda: update_position("top-left"), outputs=position)
|
542 |
btn_top_center.click(fn=lambda: update_position("top-center"), outputs=position)
|
543 |
btn_top_right.click(fn=lambda: update_position("top-right"), outputs=position)
|
|
|
547 |
btn_bottom_left.click(fn=lambda: update_position("bottom-left"), outputs=position)
|
548 |
btn_bottom_center.click(fn=lambda: update_position("bottom-center"), outputs=position)
|
549 |
btn_bottom_right.click(fn=lambda: update_position("bottom-right"), outputs=position)
|
|
|
550 |
with gr.Column(scale=1):
|
551 |
with gr.Row():
|
552 |
combined_image = gr.Image(
|
|
|
562 |
type="pil",
|
563 |
height=256
|
564 |
)
|
|
|
565 |
# Event bindings
|
566 |
input_image.change(
|
567 |
fn=update_process_button,
|
|
|
569 |
outputs=process_btn,
|
570 |
queue=False
|
571 |
)
|
|
|
572 |
text_prompt.change(
|
573 |
fn=update_process_button,
|
574 |
inputs=[input_image, text_prompt],
|
575 |
outputs=process_btn,
|
576 |
queue=False
|
577 |
)
|
|
|
578 |
def update_controls(bg_prompt):
|
|
|
579 |
is_visible = bool(bg_prompt)
|
580 |
return [
|
581 |
+
gr.update(visible=is_visible),
|
582 |
+
gr.update(visible=is_visible),
|
583 |
]
|
|
|
584 |
bg_prompt.change(
|
585 |
fn=update_controls,
|
586 |
inputs=bg_prompt,
|
587 |
outputs=[aspect_ratio, object_controls],
|
588 |
queue=False
|
589 |
)
|
|
|
590 |
process_btn.click(
|
591 |
fn=process_prompt,
|
592 |
inputs=[
|
|
|
600 |
outputs=[combined_image, extracted_image],
|
601 |
queue=True
|
602 |
)
|
603 |
+
# 예제 섹션 추가
|
604 |
+
with gr.Accordion("Show Example", open=True):
|
605 |
+
gr.Markdown("### Example")
|
606 |
+
with gr.Row():
|
607 |
+
with gr.Column():
|
608 |
+
gr.Markdown("**Upload Image(aa1.png)**")
|
609 |
+
gr.Image(value="aa1.png", label="Upload")
|
610 |
+
with gr.Column():
|
611 |
+
gr.Markdown("**Cut Object (aa2.png)**<br>(Prompt: 'text')", elem_classes="center")
|
612 |
+
gr.Image(value="aa2.png", label="Object")
|
613 |
+
with gr.Column():
|
614 |
+
gr.Markdown("**Generated Image (aa3.png)**<br>(Background Prompt: 'alps mountain')", elem_classes="center")
|
615 |
+
gr.Image(value="aa3.png", label="Output")
|
616 |
+
demo.queue(max_size=5)
|
617 |
demo.launch(
|
618 |
server_name="0.0.0.0",
|
619 |
server_port=7860,
|
620 |
share=False,
|
621 |
+
max_threads=2
|
622 |
)
|