Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -3,17 +3,55 @@ import cv2
|
|
3 |
import numpy as np
|
4 |
import torch
|
5 |
import gradio as gr
|
6 |
-
import spaces
|
7 |
-
|
|
|
8 |
from PIL import Image, ImageOps
|
9 |
from transformers import AutoModelForImageSegmentation
|
10 |
from torchvision import transforms
|
11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
torch.set_float32_matmul_precision('high')
|
13 |
torch.jit.script = lambda f: f
|
14 |
-
|
15 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
16 |
|
|
|
17 |
def refine_foreground(image, mask, r=90):
|
18 |
if mask.size != image.size:
|
19 |
mask = mask.resize(image.size)
|
@@ -33,14 +71,11 @@ def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90):
|
|
33 |
if isinstance(image, Image.Image):
|
34 |
image = np.array(image) / 255.0
|
35 |
blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
|
36 |
-
|
37 |
blurred_FA = cv2.blur(F * alpha, (r, r))
|
38 |
blurred_F = blurred_FA / (blurred_alpha + 1e-5)
|
39 |
-
|
40 |
blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
|
41 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
42 |
-
F = blurred_F + alpha *
|
43 |
-
(image - alpha * blurred_F - (1 - alpha) * blurred_B)
|
44 |
F = np.clip(F, 0, 1)
|
45 |
return F, blurred_B
|
46 |
|
@@ -67,42 +102,33 @@ def remove_background_wrapper(image):
|
|
67 |
if image is None:
|
68 |
raise gr.Error("Please upload an image.")
|
69 |
image_ori = Image.fromarray(image).convert('RGB')
|
70 |
-
# Call the processing function
|
71 |
foreground, background, pred_pil, reverse_mask = remove_background(image_ori)
|
72 |
return foreground, background, pred_pil, reverse_mask
|
73 |
|
74 |
-
@spaces.GPU
|
75 |
def remove_background(image_ori):
|
76 |
original_size = image_ori.size
|
77 |
-
|
78 |
-
# Preprocess the image
|
79 |
image_preprocessor = ImagePreprocessor(resolution=(1024, 1024))
|
80 |
image_proc = image_preprocessor.proc(image_ori)
|
81 |
image_proc = image_proc.unsqueeze(0)
|
82 |
-
|
83 |
-
# Prediction
|
84 |
with torch.no_grad():
|
85 |
preds = birefnet(image_proc.to(device))[-1].sigmoid().cpu()
|
86 |
pred = preds[0].squeeze()
|
87 |
-
|
88 |
-
# Process Results
|
89 |
pred_pil = transforms.ToPILImage()(pred)
|
90 |
-
pred_pil = pred_pil.resize(original_size, Image.BICUBIC)
|
91 |
-
|
92 |
-
# Create reverse mask (background mask)
|
93 |
reverse_mask = ImageOps.invert(pred_pil)
|
94 |
-
|
95 |
-
# Create foreground image (object with transparent background)
|
96 |
foreground = image_ori.copy()
|
97 |
foreground.putalpha(pred_pil)
|
98 |
-
|
99 |
-
# Create background image
|
100 |
background = image_ori.copy()
|
101 |
background.putalpha(reverse_mask)
|
102 |
-
|
103 |
torch.cuda.empty_cache()
|
104 |
-
|
105 |
-
# Return images in the specified order
|
106 |
return foreground, background, pred_pil, reverse_mask
|
107 |
|
108 |
# Custom CSS for button styling
|
@@ -123,11 +149,12 @@ custom_css = """
|
|
123 |
animation: gradient-animation 15s ease infinite;
|
124 |
border-radius: 12px;
|
125 |
color: black;
|
|
|
126 |
}
|
127 |
"""
|
128 |
|
129 |
-
with
|
130 |
-
|
131 |
with gr.Row():
|
132 |
with gr.Column():
|
133 |
image_input = gr.Image(type="numpy", sources=['upload'], label="Upload Image")
|
@@ -138,7 +165,6 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Default()) as demo:
|
|
138 |
output_foreground_mask = gr.Image(type="pil", label="Foreground Mask")
|
139 |
output_background_mask = gr.Image(type="pil", label="Background Mask")
|
140 |
|
141 |
-
# Link the button to the processing function
|
142 |
btn.click(fn=remove_background_wrapper, inputs=image_input, outputs=[
|
143 |
output_foreground, output_background, output_foreground_mask, output_background_mask])
|
144 |
|
|
|
3 |
import numpy as np
|
4 |
import torch
|
5 |
import gradio as gr
|
6 |
+
import spaces
|
7 |
+
from gradio.themes.base import Base
|
8 |
+
from gradio.themes.utils import colors, fonts, sizes
|
9 |
from PIL import Image, ImageOps
|
10 |
from transformers import AutoModelForImageSegmentation
|
11 |
from torchvision import transforms
|
12 |
|
13 |
+
# Custom White Theme with Inter font
|
14 |
+
class WhiteTheme(Base):
|
15 |
+
def __init__(
|
16 |
+
self,
|
17 |
+
*,
|
18 |
+
primary_hue: colors.Color | str = colors.orange,
|
19 |
+
font: fonts.Font | str = fonts.GoogleFont("Inter"),
|
20 |
+
font_mono: fonts.Font | str = fonts.GoogleFont("Inter")
|
21 |
+
):
|
22 |
+
super().__init__(
|
23 |
+
primary_hue=primary_hue,
|
24 |
+
font=font,
|
25 |
+
font_mono=font_mono,
|
26 |
+
)
|
27 |
+
|
28 |
+
self.set(
|
29 |
+
body_background_fill="white",
|
30 |
+
block_background_fill="white",
|
31 |
+
panel_background_fill="white",
|
32 |
+
body_text_color="black",
|
33 |
+
block_label_text_color="black",
|
34 |
+
block_border_color="white",
|
35 |
+
panel_border_color="white",
|
36 |
+
input_border_color="lightgray",
|
37 |
+
button_primary_background_fill="*primary_500",
|
38 |
+
button_primary_background_fill_hover="*primary_600",
|
39 |
+
button_primary_text_color="white",
|
40 |
+
button_secondary_background_fill="white",
|
41 |
+
button_secondary_border_color="lightgray",
|
42 |
+
block_shadow="none",
|
43 |
+
button_shadow="none",
|
44 |
+
input_shadow="none",
|
45 |
+
slider_color="*primary_500",
|
46 |
+
slider_track_color="lightgray",
|
47 |
+
)
|
48 |
+
|
49 |
+
# Your existing setup code
|
50 |
torch.set_float32_matmul_precision('high')
|
51 |
torch.jit.script = lambda f: f
|
|
|
52 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
53 |
|
54 |
+
# Keep all your existing functions unchanged
|
55 |
def refine_foreground(image, mask, r=90):
|
56 |
if mask.size != image.size:
|
57 |
mask = mask.resize(image.size)
|
|
|
71 |
if isinstance(image, Image.Image):
|
72 |
image = np.array(image) / 255.0
|
73 |
blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
|
|
|
74 |
blurred_FA = cv2.blur(F * alpha, (r, r))
|
75 |
blurred_F = blurred_FA / (blurred_alpha + 1e-5)
|
|
|
76 |
blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
|
77 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
78 |
+
F = blurred_F + alpha * (image - alpha * blurred_F - (1 - alpha) * blurred_B)
|
|
|
79 |
F = np.clip(F, 0, 1)
|
80 |
return F, blurred_B
|
81 |
|
|
|
102 |
if image is None:
|
103 |
raise gr.Error("Please upload an image.")
|
104 |
image_ori = Image.fromarray(image).convert('RGB')
|
|
|
105 |
foreground, background, pred_pil, reverse_mask = remove_background(image_ori)
|
106 |
return foreground, background, pred_pil, reverse_mask
|
107 |
|
108 |
+
@spaces.GPU
|
109 |
def remove_background(image_ori):
|
110 |
original_size = image_ori.size
|
|
|
|
|
111 |
image_preprocessor = ImagePreprocessor(resolution=(1024, 1024))
|
112 |
image_proc = image_preprocessor.proc(image_ori)
|
113 |
image_proc = image_proc.unsqueeze(0)
|
114 |
+
|
|
|
115 |
with torch.no_grad():
|
116 |
preds = birefnet(image_proc.to(device))[-1].sigmoid().cpu()
|
117 |
pred = preds[0].squeeze()
|
118 |
+
|
|
|
119 |
pred_pil = transforms.ToPILImage()(pred)
|
120 |
+
pred_pil = pred_pil.resize(original_size, Image.BICUBIC)
|
121 |
+
|
|
|
122 |
reverse_mask = ImageOps.invert(pred_pil)
|
123 |
+
|
|
|
124 |
foreground = image_ori.copy()
|
125 |
foreground.putalpha(pred_pil)
|
126 |
+
|
|
|
127 |
background = image_ori.copy()
|
128 |
background.putalpha(reverse_mask)
|
129 |
+
|
130 |
torch.cuda.empty_cache()
|
131 |
+
|
|
|
132 |
return foreground, background, pred_pil, reverse_mask
|
133 |
|
134 |
# Custom CSS for button styling
|
|
|
149 |
animation: gradient-animation 15s ease infinite;
|
150 |
border-radius: 12px;
|
151 |
color: black;
|
152 |
+
font-family: 'Inter', sans-serif;
|
153 |
}
|
154 |
"""
|
155 |
|
156 |
+
# Create the interface with the custom theme
|
157 |
+
with gr.Blocks(css=custom_css, theme=WhiteTheme()) as demo:
|
158 |
with gr.Row():
|
159 |
with gr.Column():
|
160 |
image_input = gr.Image(type="numpy", sources=['upload'], label="Upload Image")
|
|
|
165 |
output_foreground_mask = gr.Image(type="pil", label="Foreground Mask")
|
166 |
output_background_mask = gr.Image(type="pil", label="Background Mask")
|
167 |
|
|
|
168 |
btn.click(fn=remove_background_wrapper, inputs=image_input, outputs=[
|
169 |
output_foreground, output_background, output_foreground_mask, output_background_mask])
|
170 |
|