import tempfile
import time
from collections.abc import Sequence
from typing import Any, cast
import os
from huggingface_hub import login
import gradio as gr
import numpy as np
import pillow_heif
import spaces
import torch
from gradio_image_annotation import image_annotator
from gradio_imageslider import ImageSlider
from PIL import Image
from pymatting.foreground.estimate_foreground_ml import estimate_foreground_ml
from refiners.fluxion.utils import no_grad
from refiners.solutions import BoxSegmenter
from transformers import GroundingDinoForObjectDetection, GroundingDinoProcessor
from diffusers import FluxPipeline
BoundingBox = tuple[int, int, int, int]
# 초기화 및 설정
pillow_heif.register_heif_opener()
pillow_heif.register_avif_opener()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# HF 토큰 설정
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN is None:
raise ValueError("Please set the HF_TOKEN environment variable")
try:
login(token=HF_TOKEN)
except Exception as e:
raise ValueError(f"Failed to login to Hugging Face: {str(e)}")
# 모델 초기화
segmenter = BoxSegmenter(device="cpu")
segmenter.device = device
segmenter.model = segmenter.model.to(device=segmenter.device)
gd_model_path = "IDEA-Research/grounding-dino-base"
gd_processor = GroundingDinoProcessor.from_pretrained(gd_model_path)
gd_model = GroundingDinoForObjectDetection.from_pretrained(gd_model_path, torch_dtype=torch.float32)
gd_model = gd_model.to(device=device)
assert isinstance(gd_model, GroundingDinoForObjectDetection)
# FLUX 파이프라인 초기화
pipe = FluxPipeline.from_pretrained(
"black-forest-labs/FLUX.1-dev",
torch_dtype=torch.bfloat16,
use_auth_token=HF_TOKEN
)
pipe.load_lora_weights(
hf_hub_download(
"ByteDance/Hyper-SD",
"Hyper-FLUX.1-dev-8steps-lora.safetensors",
use_auth_token=HF_TOKEN
)
)
pipe.fuse_lora(lora_scale=0.125)
pipe.to(device="cuda", dtype=torch.bfloat16)
def bbox_union(bboxes: Sequence[list[int]]) -> BoundingBox | None:
if not bboxes:
return None
for bbox in bboxes:
assert len(bbox) == 4
assert all(isinstance(x, int) for x in bbox)
return (
min(bbox[0] for bbox in bboxes),
min(bbox[1] for bbox in bboxes),
max(bbox[2] for bbox in bboxes),
max(bbox[3] for bbox in bboxes),
)
def corners_to_pixels_format(bboxes: torch.Tensor, width: int, height: int) -> torch.Tensor:
x1, y1, x2, y2 = bboxes.round().to(torch.int32).unbind(-1)
return torch.stack((x1.clamp_(0, width), y1.clamp_(0, height), x2.clamp_(0, width), y2.clamp_(0, height)), dim=-1)
def gd_detect(img: Image.Image, prompt: str) -> BoundingBox | None:
inputs = gd_processor(images=img, text=f"{prompt}.", return_tensors="pt").to(device=device)
with no_grad():
outputs = gd_model(**inputs)
width, height = img.size
results: dict[str, Any] = gd_processor.post_process_grounded_object_detection(
outputs,
inputs["input_ids"],
target_sizes=[(height, width)],
)[0]
assert "boxes" in results and isinstance(results["boxes"], torch.Tensor)
bboxes = corners_to_pixels_format(results["boxes"].cpu(), width, height)
return bbox_union(bboxes.numpy().tolist())
def apply_mask(img: Image.Image, mask_img: Image.Image, defringe: bool = True) -> Image.Image:
assert img.size == mask_img.size
img = img.convert("RGB")
mask_img = mask_img.convert("L")
if defringe:
rgb, alpha = np.asarray(img) / 255.0, np.asarray(mask_img) / 255.0
foreground = cast(np.ndarray[Any, np.dtype[np.uint8]], estimate_foreground_ml(rgb, alpha))
img = Image.fromarray((foreground * 255).astype("uint8"))
result = Image.new("RGBA", img.size)
result.paste(img, (0, 0), mask_img)
return result
def generate_background(prompt: str, width: int, height: int) -> Image.Image:
"""배경 이미지 생성 함수"""
try:
with timer("Background generation"):
image = pipe(
prompt=prompt,
width=width,
height=height,
num_inference_steps=8,
guidance_scale=4.0,
).images[0]
return image
except Exception as e:
raise gr.Error(f"Background generation failed: {str(e)}")
def combine_with_background(foreground: Image.Image, background: Image.Image) -> Image.Image:
"""전경과 배경 합성 함수"""
background = background.resize(foreground.size)
return Image.alpha_composite(background.convert('RGBA'), foreground)
@spaces.GPU
def _gpu_process(img: Image.Image, prompt: str | BoundingBox | None) -> tuple[Image.Image, BoundingBox | None, list[str]]:
time_log: list[str] = []
if isinstance(prompt, str):
t0 = time.time()
bbox = gd_detect(img, prompt)
time_log.append(f"detect: {time.time() - t0}")
if not bbox:
print(time_log[0])
raise gr.Error("No object detected")
else:
bbox = prompt
t0 = time.time()
mask = segmenter(img, bbox)
time_log.append(f"segment: {time.time() - t0}")
return mask, bbox, time_log
def _process(img: Image.Image, prompt: str | BoundingBox | None, bg_prompt: str | None = None) -> tuple[tuple[Image.Image, Image.Image, Image.Image], gr.DownloadButton]:
if img.width > 2048 or img.height > 2048:
orig_res = max(img.width, img.height)
img.thumbnail((2048, 2048))
if isinstance(prompt, tuple):
x0, y0, x1, y1 = (int(x * 2048 / orig_res) for x in prompt)
prompt = (x0, y0, x1, y1)
mask, bbox, time_log = _gpu_process(img, prompt)
masked_alpha = apply_mask(img, mask, defringe=True)
if bg_prompt:
try:
background = generate_background(bg_prompt, img.width, img.height)
combined = combine_with_background(masked_alpha, background)
except Exception as e:
raise gr.Error(f"Background processing failed: {str(e)}")
else:
combined = Image.alpha_composite(Image.new("RGBA", masked_alpha.size, "white"), masked_alpha)
thresholded = mask.point(lambda p: 255 if p > 10 else 0)
bbox = thresholded.getbbox()
to_dl = masked_alpha.crop(bbox)
temp = tempfile.NamedTemporaryFile(delete=False, suffix=".png")
to_dl.save(temp, format="PNG")
temp.close()
return (img, combined, masked_alpha), gr.DownloadButton(value=temp.name, interactive=True)
def process_bbox(prompts: dict[str, Any]) -> tuple[tuple[Image.Image, Image.Image], gr.DownloadButton]:
assert isinstance(img := prompts["image"], Image.Image)
assert isinstance(boxes := prompts["boxes"], list)
if len(boxes) == 1:
assert isinstance(box := boxes[0], dict)
bbox = tuple(box[k] for k in ["xmin", "ymin", "xmax", "ymax"])
else:
assert len(boxes) == 0
bbox = None
return _process(img, bbox)
def on_change_bbox(prompts: dict[str, Any] | None):
return gr.update(interactive=prompts is not None)
def process_prompt(img: Image.Image, prompt: str, bg_prompt: str | None = None) -> tuple[tuple[Image.Image, Image.Image], gr.DownloadButton]:
return _process(img, prompt, bg_prompt)
def on_change_prompt(img: Image.Image | None, prompt: str | None, bg_prompt: str | None = None):
return gr.update(interactive=bool(img and prompt))
# CSS 스타일 정의
css = """
footer {display: none}
.main-title {
text-align: center;
margin: 2em 0;
}
.main-title h1 {
color: #2196F3;
font-size: 2.5em;
}
.container {
max-width: 1200px;
margin: auto;
padding: 20px;
}
"""
# Gradio UI
with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
gr.HTML("""
🎨 Advanced Image Object Extractor
Extract objects from images using text prompts or bounding boxes
""")
with gr.Tabs() as tabs:
with gr.Tab("✨ Extract by Text", id="tab_prompt"):
with gr.Row(equal_height=True):
with gr.Column(scale=1, min_width=400):
gr.HTML("
📥 Input Section
")
iimg = gr.Image(
type="pil",
label="Upload Image"
)
with gr.Group():
prompt = gr.Textbox(
label="🎯 Object to Extract",
placeholder="Enter what you want to extract..."
)
bg_prompt = gr.Textbox(
label="🖼️ Background Generation Prompt (optional)",
placeholder="Describe the background you want..."
)
btn = gr.Button(
"🚀 Process Image",
variant="primary",
interactive=False
)
with gr.Column(scale=1, min_width=400):
gr.HTML("
📤 Output Section
")
oimg = ImageSlider(
label="Results Preview",
show_download_button=False
)
dlbt = gr.DownloadButton(
"💾 Download Result",
interactive=False
)
with gr.Accordion("📚 Examples", open=False):
examples = [
["examples/text.jpg", "text", "white background"],
["examples/black-lamp.jpg", "black lamp", "minimalist interior"]
]
ex = gr.Examples(
examples=examples,
inputs=[iimg, prompt, bg_prompt],
outputs=[oimg, dlbt],
fn=process_prompt,
cache_examples=True
)
with gr.Tab("📏 Extract by Box", id="tab_bb"):
with gr.Row(equal_height=True):
with gr.Column(scale=1, min_width=400):
gr.HTML("