Update app.py
Browse files
app.py
CHANGED
@@ -1,86 +1,180 @@
|
|
1 |
-
import cv2
|
2 |
-
import numpy as np
|
3 |
import gradio as gr
|
4 |
-
|
5 |
-
from
|
6 |
import torch
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
-
|
9 |
-
def load_sam_model():
|
10 |
-
# Tải checkpoint từ Hugging Face với map_location=torch.device('cpu')
|
11 |
-
checkpoint_path = hf_hub_download(repo_id="facebook/sam-vit-huge", filename="pytorch_model.bin")
|
12 |
-
|
13 |
-
# Load checkpoint với map_location=torch.device('cpu')
|
14 |
-
checkpoint = torch.load(checkpoint_path, map_location=torch.device('cpu'))
|
15 |
-
|
16 |
-
# Khởi tạo mô hình SAM
|
17 |
-
model_type = "vit_h"
|
18 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
19 |
-
|
20 |
-
# Truyền checkpoint vào mô hình
|
21 |
-
sam = sam_model_registry[model_type]()
|
22 |
-
sam.load_state_dict(checkpoint)
|
23 |
-
sam.to(device=device)
|
24 |
-
predictor = SamPredictor(sam)
|
25 |
-
return predictor
|
26 |
-
|
27 |
-
predictor = load_sam_model()
|
28 |
-
|
29 |
-
def generate_mask(image, event: gr.SelectData):
|
30 |
"""
|
31 |
-
|
32 |
-
:param image: The input image (numpy array).
|
33 |
-
:param event: Gradio SelectData containing the click coordinates.
|
34 |
-
:return: A binary mask where the selected object is black, and the rest is white.
|
35 |
"""
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
input_point = np.array([[x, y]])
|
43 |
-
input_label = np.array([1]) # 1 indicates foreground
|
44 |
-
|
45 |
-
# Generate masks
|
46 |
-
masks, scores, logits = predictor.predict(
|
47 |
-
point_coords=input_point,
|
48 |
-
point_labels=input_label,
|
49 |
-
multimask_output=True,
|
50 |
-
)
|
51 |
|
52 |
-
|
53 |
-
best_mask = masks[np.argmax(scores)]
|
54 |
|
55 |
-
|
56 |
-
|
57 |
-
binary_mask = cv2.bitwise_not(binary_mask) # Invert colors (black for object)
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
return binary_mask
|
60 |
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
"""
|
63 |
-
|
64 |
"""
|
65 |
-
|
66 |
-
|
67 |
-
gr.Markdown("Upload an image, click on an object to select it, and generate a binary mask.")
|
68 |
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
-
|
75 |
-
|
76 |
-
gr.Markdown("1. Upload an image.")
|
77 |
-
gr.Markdown("2. Click on the object you want to change.")
|
78 |
-
gr.Markdown("3. The mask will be generated automatically.")
|
79 |
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
|
84 |
-
|
85 |
-
demo = app()
|
86 |
-
demo.launch()
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import numpy as np
|
3 |
+
from PIL import Image, ImageDraw
|
4 |
import torch
|
5 |
+
from transformers import SamModel, SamProcessor
|
6 |
+
from diffusers import StableDiffusionInpaintPipeline
|
7 |
+
|
8 |
+
# Constants
|
9 |
+
IMG_SIZE = 512
|
10 |
|
11 |
+
def generate_mask(image, points):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
"""
|
13 |
+
Generates a mask using SAM based on input points.
|
|
|
|
|
|
|
14 |
"""
|
15 |
+
if not points:
|
16 |
+
return None
|
17 |
+
|
18 |
+
# Initialize SAM model and processor on CPU
|
19 |
+
sam_model = SamModel.from_pretrained("facebook/sam-vit-huge", torch_dtype=torch.float32).to("cpu")
|
20 |
+
sam_processor = SamProcessor.from_pretrained("facebook/sam-vit-huge")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
+
inputs = sam_processor(image, points=points, return_tensors="pt").to("cpu")
|
|
|
23 |
|
24 |
+
with torch.no_grad():
|
25 |
+
outputs = sam_model(**inputs)
|
|
|
26 |
|
27 |
+
masks = sam_processor.image_processor.post_process_masks(
|
28 |
+
outputs.pred_masks.cpu(),
|
29 |
+
inputs["original_sizes"].cpu(),
|
30 |
+
inputs["reshaped_input_sizes"].cpu()
|
31 |
+
)
|
32 |
+
|
33 |
+
if len(masks) == 0:
|
34 |
+
return None
|
35 |
+
|
36 |
+
best_mask = masks[0][0][outputs.iou_scores.argmax()]
|
37 |
+
binary_mask = ~best_mask.numpy().astype(bool).astype(int)
|
38 |
return binary_mask
|
39 |
|
40 |
+
|
41 |
+
def replace_object(image, mask, prompt, negative_prompt, seed, guidance_scale):
|
42 |
+
"""
|
43 |
+
Replaces the object in the image based on the mask and prompt.
|
44 |
+
"""
|
45 |
+
if mask is None:
|
46 |
+
return image
|
47 |
+
|
48 |
+
# Initialize Inpainting pipeline on CPU with a compatible model
|
49 |
+
inpaint_pipeline = StableDiffusionInpaintPipeline.from_pretrained(
|
50 |
+
"stabilityai/stable-diffusion-2-inpainting",
|
51 |
+
torch_dtype=torch.float32
|
52 |
+
).to("cpu")
|
53 |
+
|
54 |
+
mask_image = Image.fromarray((mask * 255).astype(np.uint8))
|
55 |
+
|
56 |
+
generator = torch.Generator("cpu").manual_seed(seed)
|
57 |
+
|
58 |
+
try:
|
59 |
+
result = inpaint_pipeline(
|
60 |
+
prompt=prompt,
|
61 |
+
image=image,
|
62 |
+
mask_image=mask_image,
|
63 |
+
negative_prompt=negative_prompt if negative_prompt else None,
|
64 |
+
generator=generator,
|
65 |
+
guidance_scale=guidance_scale
|
66 |
+
).images[0]
|
67 |
+
return result
|
68 |
+
except Exception as e:
|
69 |
+
print(f"Inpainting error: {e}")
|
70 |
+
return image
|
71 |
+
|
72 |
+
|
73 |
+
def visualize_mask(image, mask):
|
74 |
"""
|
75 |
+
Overlays the mask on the image for visualization.
|
76 |
"""
|
77 |
+
if mask is None:
|
78 |
+
return image
|
|
|
79 |
|
80 |
+
bg_transparent = np.zeros(mask.shape + (4,), dtype=np.uint8)
|
81 |
+
bg_transparent[mask == 1] = [0, 255, 0, 127] # Green with transparency
|
82 |
+
mask_rgba = Image.fromarray(bg_transparent)
|
83 |
+
|
84 |
+
overlay = Image.alpha_composite(image.convert("RGBA"), mask_rgba)
|
85 |
+
return overlay.convert("RGB")
|
86 |
+
|
87 |
+
|
88 |
+
def get_points(img, evt: gr.SelectData, input_points):
|
89 |
+
"""
|
90 |
+
Captures points selected by the user on the image.
|
91 |
+
"""
|
92 |
+
x, y = evt.index
|
93 |
+
input_points.append([x, y])
|
94 |
|
95 |
+
# Generate mask based on selected points
|
96 |
+
mask = generate_mask(img, input_points)
|
|
|
|
|
|
|
97 |
|
98 |
+
# Mark selected points with a green crossmark
|
99 |
+
draw = ImageDraw.Draw(img)
|
100 |
+
size = 10
|
101 |
+
for point in input_points:
|
102 |
+
px, py = point
|
103 |
+
draw.line((px - size, py, px + size, py), fill="green", width=5)
|
104 |
+
draw.line((px, py - size, px, py + size), fill="green", width=5)
|
105 |
|
106 |
+
# Visualize the mask overlay
|
107 |
+
masked_image = visualize_mask(img, mask)
|
108 |
+
|
109 |
+
return masked_image, input_points
|
110 |
+
|
111 |
+
|
112 |
+
def run_inpaint(prompt, negative_prompt, cfg, seed, invert, input_image, input_points):
|
113 |
+
"""
|
114 |
+
Runs the inpainting process based on user inputs.
|
115 |
+
"""
|
116 |
+
if input_image is None or len(input_points) == 0:
|
117 |
+
raise gr.Error("No points provided. Click on the image to select the object to segment with SAM.")
|
118 |
+
|
119 |
+
mask = generate_mask(input_image, input_points)
|
120 |
+
|
121 |
+
if invert:
|
122 |
+
mask = ~mask
|
123 |
+
|
124 |
+
try:
|
125 |
+
inpainted = replace_object(input_image, mask, prompt, negative_prompt, seed, cfg)
|
126 |
+
except Exception as e:
|
127 |
+
raise gr.Error(str(e))
|
128 |
+
|
129 |
+
return inpainted.resize((IMG_SIZE, IMG_SIZE))
|
130 |
+
|
131 |
+
|
132 |
+
def preprocess(input_img):
|
133 |
+
"""
|
134 |
+
Preprocesses the uploaded image to ensure it is square and resized.
|
135 |
+
"""
|
136 |
+
if input_img is None:
|
137 |
+
return None
|
138 |
+
|
139 |
+
width, height = input_img.size
|
140 |
+
if width != height:
|
141 |
+
# Add white padding to make the image square
|
142 |
+
new_size = max(width, height)
|
143 |
+
new_image = Image.new("RGB", (new_size, new_size), 'white')
|
144 |
+
left = (new_size - width) // 2
|
145 |
+
top = (new_size - height) // 2
|
146 |
+
new_image.paste(input_img, (left, top))
|
147 |
+
input_img = new_image
|
148 |
+
|
149 |
+
return input_img.resize((IMG_SIZE, IMG_SIZE))
|
150 |
+
|
151 |
+
|
152 |
+
# Gradio Interface
|
153 |
+
with gr.Blocks() as demo:
|
154 |
+
gr.Markdown("# Object Replacement with SAM and Stable Diffusion Inpainting")
|
155 |
+
gr.Markdown("Upload an image, click on the object you want to replace, and generate a new image.")
|
156 |
+
|
157 |
+
with gr.Row():
|
158 |
+
with gr.Column():
|
159 |
+
input_image = gr.Image(label="Upload Image", type="pil")
|
160 |
+
output_image = gr.Image(label="Generated Image", type="pil")
|
161 |
+
input_points = gr.State([]) # Store selected points
|
162 |
+
|
163 |
+
with gr.Column():
|
164 |
+
prompt = gr.Textbox(label="Prompt for Inpainting")
|
165 |
+
negative_prompt = gr.Textbox(label="Negative Prompt (Optional)")
|
166 |
+
cfg = gr.Slider(1, 20, value=7.5, label="Guidance Scale")
|
167 |
+
seed = gr.Number(value=42, label="Seed")
|
168 |
+
invert = gr.Checkbox(label="Invert Mask")
|
169 |
+
run_button = gr.Button("Run Inpainting")
|
170 |
+
reset_button = gr.Button("Reset Points")
|
171 |
+
|
172 |
+
input_image.select(get_points, inputs=[input_image, input_points], outputs=[output_image, input_points])
|
173 |
+
run_button.click(
|
174 |
+
run_inpaint,
|
175 |
+
inputs=[prompt, negative_prompt, cfg, seed, invert, input_image, input_points],
|
176 |
+
outputs=output_image
|
177 |
+
)
|
178 |
+
reset_button.click(lambda: (None, []), outputs=[output_image, input_points])
|
179 |
|
180 |
+
demo.launch()
|
|
|
|