Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -4,132 +4,119 @@ import torch
|
|
4 |
import cv2
|
5 |
from PIL import Image
|
6 |
from torchvision import transforms
|
7 |
-
from cloth_segmentation.networks.u2net import U2NET
|
8 |
-
import matplotlib.colors as mcolors
|
9 |
-
import traceback
|
10 |
|
11 |
-
# Load U²-Net
|
12 |
model_path = "cloth_segmentation/networks/u2net.pth"
|
13 |
model = U2NET(3, 1)
|
14 |
-
state_dict = torch.load(model_path, map_location=torch.device(
|
15 |
-
state_dict = {k.replace(
|
16 |
model.load_state_dict(state_dict)
|
17 |
model.eval()
|
18 |
|
19 |
-
# Util to get BGR color from name
|
20 |
-
def get_bgr_from_color_name(color_name):
|
21 |
-
try:
|
22 |
-
rgb = mcolors.to_rgb(color_name.lower())
|
23 |
-
return tuple(int(255 * c) for c in rgb[::-1]) # Convert to BGR
|
24 |
-
except:
|
25 |
-
return (0, 0, 255) # Default to red
|
26 |
-
|
27 |
-
# Mask refinement
|
28 |
-
def refine_mask(mask):
|
29 |
-
close_kernel = np.ones((5, 5), np.uint8)
|
30 |
-
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, close_kernel)
|
31 |
-
erode_kernel = np.ones((3, 3), np.uint8)
|
32 |
-
mask = cv2.erode(mask, erode_kernel, iterations=1)
|
33 |
-
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, close_kernel)
|
34 |
-
return cv2.GaussianBlur(mask, (5, 5), 1.5)
|
35 |
-
|
36 |
-
# U²-Net segmentation
|
37 |
def segment_dress(image_np):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
transform_pipeline = transforms.Compose([
|
39 |
transforms.ToTensor(),
|
40 |
transforms.Resize((320, 320))
|
41 |
])
|
|
|
42 |
image = Image.fromarray(image_np).convert("RGB")
|
43 |
input_tensor = transform_pipeline(image).unsqueeze(0)
|
44 |
|
45 |
with torch.no_grad():
|
46 |
output = model(input_tensor)[0][0].squeeze().cpu().numpy()
|
47 |
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
|
|
|
|
69 |
img_lab = cv2.cvtColor(image_np, cv2.COLOR_RGB2LAB)
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
feathered_mask = cv2.GaussianBlur(dress_mask, (21, 21), 7)
|
84 |
-
lightness_mask = (img_lab[..., 0] / 255.0) ** 0.7
|
85 |
-
adaptive_feather = (feathered_mask * lightness_mask).astype(np.uint8)
|
86 |
-
|
87 |
-
return (image_np * (1 - adaptive_feather[..., None] / 255) + img_recolored * (adaptive_feather[..., None] / 255)).astype(np.uint8)
|
88 |
-
|
89 |
-
# Main function with enhanced error handling
|
90 |
-
def change_dress_color(img, color_prompt):
|
91 |
-
if img is None or not color_prompt:
|
92 |
-
return img
|
93 |
-
|
94 |
-
try:
|
95 |
-
img_np = np.array(img)
|
96 |
-
target_bgr = get_bgr_from_color_name(color_prompt)
|
97 |
-
|
98 |
-
dress_mask = segment_dress(img_np)
|
99 |
-
if np.sum(dress_mask) < 1000:
|
100 |
-
return img
|
101 |
-
|
102 |
-
dress_mask = apply_grabcut(img_np, dress_mask)
|
103 |
-
img_recolored = recolor_dress(img_np, dress_mask, target_bgr)
|
104 |
-
return Image.fromarray(img_recolored)
|
105 |
-
|
106 |
-
except Exception as e:
|
107 |
-
print(f"Error: {e}")
|
108 |
-
traceback.print_exc()
|
109 |
-
return img
|
110 |
-
|
111 |
-
# Create a simple function that wraps the main functionality
|
112 |
-
def process_image(image, color_name):
|
113 |
-
if image is None:
|
114 |
return None
|
115 |
-
return change_dress_color(image, color_name)
|
116 |
|
117 |
-
|
118 |
-
|
119 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
inputs=[
|
121 |
-
gr.Image(type="
|
122 |
-
gr.
|
123 |
],
|
124 |
-
outputs=gr.Image(type="pil", label="
|
125 |
-
title="
|
126 |
-
description="Upload an image of a
|
127 |
-
examples=[
|
128 |
-
["examples/dress1.jpg", "red"],
|
129 |
-
["examples/dress2.jpg", "blue"]
|
130 |
-
]
|
131 |
)
|
132 |
|
133 |
-
# Launch the application
|
134 |
if __name__ == "__main__":
|
135 |
-
|
|
|
4 |
import cv2
|
5 |
from PIL import Image
|
6 |
from torchvision import transforms
|
7 |
+
from cloth_segmentation.networks.u2net import U2NET # Import U²-Net
|
|
|
|
|
8 |
|
9 |
+
# Load U²-Net model
|
10 |
model_path = "cloth_segmentation/networks/u2net.pth"
|
11 |
model = U2NET(3, 1)
|
12 |
+
state_dict = torch.load(model_path, map_location=torch.device('cpu'))
|
13 |
+
state_dict = {k.replace('module.', ''): v for k, v in state_dict.items()} # Remove 'module.' prefix
|
14 |
model.load_state_dict(state_dict)
|
15 |
model.eval()
|
16 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
def segment_dress(image_np):
|
18 |
+
"""Segment the dress using U²-Net & refine with Lab color space."""
|
19 |
+
|
20 |
+
# Convert to Lab space
|
21 |
+
img_lab = cv2.cvtColor(image_np, cv2.COLOR_RGB2LAB)
|
22 |
+
L, A, B = cv2.split(img_lab)
|
23 |
+
|
24 |
+
# Use K-means clustering to detect dominant dress region
|
25 |
+
pixel_values = img_lab.reshape((-1, 3)).astype(np.float32)
|
26 |
+
k = 3 # Three clusters: background, skin, dress
|
27 |
+
_, labels, centers = cv2.kmeans(pixel_values, k, None, (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 10, 1.0), 10, cv2.KMEANS_RANDOM_CENTERS)
|
28 |
+
labels = labels.reshape(image_np.shape[:2])
|
29 |
+
|
30 |
+
# Assume dress is the largest non-background cluster
|
31 |
+
unique_labels, counts = np.unique(labels, return_counts=True)
|
32 |
+
dress_label = unique_labels[np.argmax(counts[1:]) + 1] # Avoid background
|
33 |
+
|
34 |
+
# Create dress mask
|
35 |
+
mask = (labels == dress_label).astype(np.uint8) * 255
|
36 |
+
|
37 |
+
# Use U²-Net prediction to refine segmentation
|
38 |
transform_pipeline = transforms.Compose([
|
39 |
transforms.ToTensor(),
|
40 |
transforms.Resize((320, 320))
|
41 |
])
|
42 |
+
|
43 |
image = Image.fromarray(image_np).convert("RGB")
|
44 |
input_tensor = transform_pipeline(image).unsqueeze(0)
|
45 |
|
46 |
with torch.no_grad():
|
47 |
output = model(input_tensor)[0][0].squeeze().cpu().numpy()
|
48 |
|
49 |
+
u2net_mask = (output > 0.5).astype(np.uint8) * 255
|
50 |
+
u2net_mask = cv2.resize(u2net_mask, (image_np.shape[1], image_np.shape[0]), interpolation=cv2.INTER_NEAREST)
|
51 |
+
|
52 |
+
# Combine K-means and U²-Net masks
|
53 |
+
refined_mask = cv2.bitwise_and(mask, u2net_mask)
|
54 |
+
|
55 |
+
return refined_mask
|
56 |
+
|
57 |
+
def detect_design(image_np, dress_mask):
|
58 |
+
"""Detect the design part of the dress and separate it from fabric."""
|
59 |
+
gray = cv2.cvtColor(image_np, cv2.COLOR_RGB2GRAY)
|
60 |
+
edges = cv2.Canny(gray, 50, 150)
|
61 |
+
|
62 |
+
# Expand detected edges to mask the design area
|
63 |
+
kernel = np.ones((5, 5), np.uint8)
|
64 |
+
design_mask = cv2.dilate(edges, kernel, iterations=2)
|
65 |
+
|
66 |
+
# Keep only the design within the dress area
|
67 |
+
design_mask = cv2.bitwise_and(design_mask, dress_mask)
|
68 |
+
return design_mask
|
69 |
+
|
70 |
+
def recolor_dress(image_np, mask, design_mask, target_color):
|
71 |
+
"""Change dress color while preserving texture and design."""
|
72 |
img_lab = cv2.cvtColor(image_np, cv2.COLOR_RGB2LAB)
|
73 |
+
target_color_lab = cv2.cvtColor(np.uint8([[target_color]]), cv2.COLOR_BGR2LAB)[0][0]
|
74 |
+
|
75 |
+
# Preserve lightness (L) and change only chromatic channels (A & B)
|
76 |
+
blend_factor = 0.7
|
77 |
+
img_lab[..., 1] = np.where((mask > 128) & (design_mask == 0), img_lab[..., 1] * (1 - blend_factor) + target_color_lab[1] * blend_factor, img_lab[..., 1])
|
78 |
+
img_lab[..., 2] = np.where((mask > 128) & (design_mask == 0), img_lab[..., 2] * (1 - blend_factor) + target_color_lab[2] * blend_factor, img_lab[..., 2])
|
79 |
+
|
80 |
+
img_recolored = cv2.cvtColor(img_lab, cv2.COLOR_LAB2RGB)
|
81 |
+
return img_recolored
|
82 |
+
|
83 |
+
def change_dress_color(image_path, color):
|
84 |
+
"""Change the dress color naturally while keeping textures and design."""
|
85 |
+
if image_path is None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
return None
|
|
|
87 |
|
88 |
+
img = Image.open(image_path).convert("RGB")
|
89 |
+
img_np = np.array(img)
|
90 |
+
dress_mask = segment_dress(img_np)
|
91 |
+
design_mask = detect_design(img_np, dress_mask)
|
92 |
+
|
93 |
+
if dress_mask is None:
|
94 |
+
return img # No dress detected
|
95 |
+
|
96 |
+
# Convert the selected color to BGR
|
97 |
+
color_map = {
|
98 |
+
"Red": (0, 0, 255), "Blue": (255, 0, 0), "Green": (0, 255, 0), "Yellow": (0, 255, 255),
|
99 |
+
"Purple": (128, 0, 128), "Orange": (0, 165, 255), "Cyan": (255, 255, 0), "Magenta": (255, 0, 255),
|
100 |
+
"White": (255, 255, 255), "Black": (0, 0, 0)
|
101 |
+
}
|
102 |
+
new_color_bgr = np.array(color_map.get(color, (0, 0, 255)), dtype=np.uint8) # Default to Red
|
103 |
+
|
104 |
+
# Recolor the dress naturally
|
105 |
+
img_recolored = recolor_dress(img_np, dress_mask, design_mask, new_color_bgr)
|
106 |
+
|
107 |
+
return Image.fromarray(img_recolored)
|
108 |
+
|
109 |
+
# Gradio Interface
|
110 |
+
demo = gr.Interface(
|
111 |
+
fn=change_dress_color,
|
112 |
inputs=[
|
113 |
+
gr.Image(type="filepath", label="Upload Dress Image"),
|
114 |
+
gr.Radio(["Red", "Blue", "Green", "Yellow", "Purple", "Orange", "Cyan", "Magenta", "White", "Black"], label="Choose New Dress Color")
|
115 |
],
|
116 |
+
outputs=gr.Image(type="pil", label="Color Changed Dress"),
|
117 |
+
title="Dress Color Changer",
|
118 |
+
description="Upload an image of a dress and select a new color to change its appearance naturally while preserving the design."
|
|
|
|
|
|
|
|
|
119 |
)
|
120 |
|
|
|
121 |
if __name__ == "__main__":
|
122 |
+
demo.launch()
|