Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,462 +1,466 @@
|
|
1 |
-
import os
|
2 |
-
import sys
|
3 |
-
|
4 |
-
sys.path.append(os.path.abspath(os.path.dirname(os.getcwd())))
|
5 |
-
# os.chdir("../")
|
6 |
-
import cv2
|
7 |
-
import gradio as gr
|
8 |
-
import numpy as np
|
9 |
-
from pathlib import Path
|
10 |
-
from matplotlib import pyplot as plt
|
11 |
-
import torch
|
12 |
-
import tempfile
|
13 |
-
|
14 |
-
from stable_diffusion_inpaint import fill_img_with_sd, replace_img_with_sd
|
15 |
-
from lama_inpaint import (
|
16 |
-
inpaint_img_with_lama,
|
17 |
-
build_lama_model,
|
18 |
-
inpaint_img_with_builded_lama,
|
19 |
-
)
|
20 |
-
from utils import (
|
21 |
-
load_img_to_array,
|
22 |
-
save_array_to_img,
|
23 |
-
dilate_mask,
|
24 |
-
show_mask,
|
25 |
-
show_points,
|
26 |
-
)
|
27 |
-
from PIL import Image
|
28 |
-
from segment_anything import SamPredictor, sam_model_registry
|
29 |
-
import argparse
|
30 |
-
|
31 |
-
|
32 |
-
def setup_args(parser):
|
33 |
-
parser.add_argument(
|
34 |
-
"--lama_config",
|
35 |
-
type=str,
|
36 |
-
default="./lama/configs/prediction/default.yaml",
|
37 |
-
help="The path to the config file of lama model. "
|
38 |
-
"Default: the config of big-lama",
|
39 |
-
)
|
40 |
-
parser.add_argument(
|
41 |
-
"--lama_ckpt",
|
42 |
-
type=str,
|
43 |
-
default="pretrained_models/big-lama",
|
44 |
-
help="The path to the lama checkpoint.",
|
45 |
-
)
|
46 |
-
parser.add_argument(
|
47 |
-
"--sam_ckpt",
|
48 |
-
type=str,
|
49 |
-
default="./pretrained_models/sam_vit_h_4b8939.pth",
|
50 |
-
help="The path to the SAM checkpoint to use for mask generation.",
|
51 |
-
)
|
52 |
-
|
53 |
-
|
54 |
-
def mkstemp(suffix, dir=None):
|
55 |
-
fd, path = tempfile.mkstemp(suffix=f"{suffix}", dir=dir)
|
56 |
-
os.close(fd)
|
57 |
-
return Path(path)
|
58 |
-
|
59 |
-
|
60 |
-
def get_sam_feat(img):
|
61 |
-
model["sam"].set_image(img)
|
62 |
-
features = model["sam"].features
|
63 |
-
orig_h = model["sam"].orig_h
|
64 |
-
orig_w = model["sam"].orig_w
|
65 |
-
input_h = model["sam"].input_h
|
66 |
-
input_w = model["sam"].input_w
|
67 |
-
model["sam"].reset_image()
|
68 |
-
return features, orig_h, orig_w, input_h, input_w
|
69 |
-
|
70 |
-
|
71 |
-
def get_fill_img_with_sd(image, mask, image_resolution, text_prompt):
|
72 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
73 |
-
if len(mask.shape) == 3:
|
74 |
-
mask = mask[:, :, 0]
|
75 |
-
np_image = np.array(image, dtype=np.uint8)
|
76 |
-
H, W, C = np_image.shape
|
77 |
-
np_image = HWC3(np_image)
|
78 |
-
np_image = resize_image(np_image, image_resolution)
|
79 |
-
mask = cv2.resize(
|
80 |
-
mask, (np_image.shape[1], np_image.shape[0]), interpolation=cv2.INTER_NEAREST
|
81 |
-
)
|
82 |
-
|
83 |
-
img_fill = fill_img_with_sd(np_image, mask, text_prompt, device=device)
|
84 |
-
img_fill = img_fill.astype(np.uint8)
|
85 |
-
return img_fill
|
86 |
-
|
87 |
-
|
88 |
-
def get_replace_img_with_sd(image, mask, image_resolution, text_prompt):
|
89 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
90 |
-
if len(mask.shape) == 3:
|
91 |
-
mask = mask[:, :, 0]
|
92 |
-
np_image = np.array(image, dtype=np.uint8)
|
93 |
-
H, W, C = np_image.shape
|
94 |
-
np_image = HWC3(np_image)
|
95 |
-
np_image = resize_image(np_image, image_resolution)
|
96 |
-
mask = cv2.resize(
|
97 |
-
mask, (np_image.shape[1], np_image.shape[0]), interpolation=cv2.INTER_NEAREST
|
98 |
-
)
|
99 |
-
|
100 |
-
img_replaced = replace_img_with_sd(np_image, mask, text_prompt, device=device)
|
101 |
-
img_replaced = img_replaced.astype(np.uint8)
|
102 |
-
return img_replaced
|
103 |
-
|
104 |
-
|
105 |
-
def HWC3(x):
|
106 |
-
assert x.dtype == np.uint8
|
107 |
-
if x.ndim == 2:
|
108 |
-
x = x[:, :, None]
|
109 |
-
assert x.ndim == 3
|
110 |
-
H, W, C = x.shape
|
111 |
-
assert C == 1 or C == 3 or C == 4
|
112 |
-
if C == 3:
|
113 |
-
return x
|
114 |
-
if C == 1:
|
115 |
-
return np.concatenate([x, x, x], axis=2)
|
116 |
-
if C == 4:
|
117 |
-
color = x[:, :, 0:3].astype(np.float32)
|
118 |
-
alpha = x[:, :, 3:4].astype(np.float32) / 255.0
|
119 |
-
y = color * alpha + 255.0 * (1.0 - alpha)
|
120 |
-
y = y.clip(0, 255).astype(np.uint8)
|
121 |
-
return y
|
122 |
-
|
123 |
-
|
124 |
-
def resize_image(input_image, resolution):
|
125 |
-
H, W, C = input_image.shape
|
126 |
-
k = float(resolution) / min(H, W)
|
127 |
-
H = int(np.round(H * k / 64.0)) * 64
|
128 |
-
W = int(np.round(W * k / 64.0)) * 64
|
129 |
-
img = cv2.resize(
|
130 |
-
input_image,
|
131 |
-
(W, H),
|
132 |
-
interpolation=cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA,
|
133 |
-
)
|
134 |
-
return img
|
135 |
-
|
136 |
-
|
137 |
-
def resize_points(clicked_points, original_shape, resolution):
|
138 |
-
original_height, original_width, _ = original_shape
|
139 |
-
original_height = float(original_height)
|
140 |
-
original_width = float(original_width)
|
141 |
-
|
142 |
-
scale_factor = float(resolution) / min(original_height, original_width)
|
143 |
-
resized_points = []
|
144 |
-
|
145 |
-
for point in clicked_points:
|
146 |
-
x, y, lab = point
|
147 |
-
resized_x = int(round(x * scale_factor))
|
148 |
-
resized_y = int(round(y * scale_factor))
|
149 |
-
resized_point = (resized_x, resized_y, lab)
|
150 |
-
resized_points.append(resized_point)
|
151 |
-
|
152 |
-
return resized_points
|
153 |
-
|
154 |
-
|
155 |
-
def get_click_mask(
|
156 |
-
clicked_points, features, orig_h, orig_w, input_h, input_w, dilate_kernel_size
|
157 |
-
):
|
158 |
-
# model['sam'].set_image(image)
|
159 |
-
model["sam"].is_image_set = True
|
160 |
-
model["sam"].features = features
|
161 |
-
model["sam"].orig_h = orig_h
|
162 |
-
model["sam"].orig_w = orig_w
|
163 |
-
model["sam"].input_h = input_h
|
164 |
-
model["sam"].input_w = input_w
|
165 |
-
|
166 |
-
# Separate the points and labels
|
167 |
-
points, labels = zip(*[(point[:2], point[2]) for point in clicked_points])
|
168 |
-
|
169 |
-
# Convert the points and labels to numpy arrays
|
170 |
-
input_point = np.array(points)
|
171 |
-
input_label = np.array(labels)
|
172 |
-
|
173 |
-
masks, _, _ = model["sam"].predict(
|
174 |
-
point_coords=input_point,
|
175 |
-
point_labels=input_label,
|
176 |
-
multimask_output=False,
|
177 |
-
)
|
178 |
-
if dilate_kernel_size is not None:
|
179 |
-
masks = [dilate_mask(mask, dilate_kernel_size) for mask in masks]
|
180 |
-
else:
|
181 |
-
masks = [mask for mask in masks]
|
182 |
-
|
183 |
-
return masks
|
184 |
-
|
185 |
-
|
186 |
-
def process_image_click(
|
187 |
-
original_image,
|
188 |
-
point_prompt,
|
189 |
-
clicked_points,
|
190 |
-
image_resolution,
|
191 |
-
features,
|
192 |
-
orig_h,
|
193 |
-
orig_w,
|
194 |
-
input_h,
|
195 |
-
input_w,
|
196 |
-
dilate_kernel_size,
|
197 |
-
evt: gr.SelectData,
|
198 |
-
):
|
199 |
-
if clicked_points is None:
|
200 |
-
clicked_points = []
|
201 |
-
|
202 |
-
# print("Received click event:", evt)
|
203 |
-
if original_image is None:
|
204 |
-
# print("No image loaded.")
|
205 |
-
return None, clicked_points, None
|
206 |
-
|
207 |
-
clicked_coords = evt.index
|
208 |
-
if clicked_coords is None:
|
209 |
-
# print("No valid coordinates received.")
|
210 |
-
return None, clicked_points, None
|
211 |
-
|
212 |
-
x, y = clicked_coords
|
213 |
-
label = point_prompt
|
214 |
-
lab = 1 if label == "Foreground Point" else 0
|
215 |
-
clicked_points.append((x, y, lab))
|
216 |
-
# print("Updated points list:", clicked_points)
|
217 |
-
|
218 |
-
input_image = np.array(original_image, dtype=np.uint8)
|
219 |
-
H, W, C = input_image.shape
|
220 |
-
input_image = HWC3(input_image)
|
221 |
-
img = resize_image(input_image, image_resolution)
|
222 |
-
# print("Processed image size:", img.shape)
|
223 |
-
|
224 |
-
resized_points = resize_points(clicked_points, input_image.shape, image_resolution)
|
225 |
-
mask_click_np = get_click_mask(
|
226 |
-
resized_points, features, orig_h, orig_w, input_h, input_w, dilate_kernel_size
|
227 |
-
)
|
228 |
-
mask_click_np = np.transpose(mask_click_np, (1, 2, 0)) * 255.0
|
229 |
-
mask_image = HWC3(mask_click_np.astype(np.uint8))
|
230 |
-
mask_image = cv2.resize(mask_image, (W, H), interpolation=cv2.INTER_LINEAR)
|
231 |
-
# print("Mask image prepared.")
|
232 |
-
|
233 |
-
edited_image = input_image
|
234 |
-
for x, y, lab in clicked_points:
|
235 |
-
color = (255, 0, 0) if lab == 1 else (0, 0, 255)
|
236 |
-
edited_image = cv2.circle(edited_image, (x, y), 20, color, -1)
|
237 |
-
|
238 |
-
opacity_mask = 0.75
|
239 |
-
opacity_edited = 1.0
|
240 |
-
overlay_image = cv2.addWeighted(
|
241 |
-
edited_image,
|
242 |
-
opacity_edited,
|
243 |
-
(mask_image * np.array([0 / 255, 255 / 255, 0 / 255])).astype(np.uint8),
|
244 |
-
opacity_mask,
|
245 |
-
0,
|
246 |
-
)
|
247 |
-
|
248 |
-
no_mask_overlay = edited_image.copy()
|
249 |
-
|
250 |
-
return no_mask_overlay, overlay_image, clicked_points, mask_image
|
251 |
-
|
252 |
-
|
253 |
-
def image_upload(image, image_resolution):
|
254 |
-
if image is None:
|
255 |
-
return None, None, None, None, None, None
|
256 |
-
else:
|
257 |
-
np_image = np.array(image, dtype=np.uint8)
|
258 |
-
H, W, C = np_image.shape
|
259 |
-
np_image = HWC3(np_image)
|
260 |
-
np_image = resize_image(np_image, image_resolution)
|
261 |
-
features, orig_h, orig_w, input_h, input_w = get_sam_feat(np_image)
|
262 |
-
return image, features, orig_h, orig_w, input_h, input_w
|
263 |
-
|
264 |
-
|
265 |
-
def get_inpainted_img(image, mask, image_resolution):
|
266 |
-
lama_config = args.lama_config
|
267 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
268 |
-
if len(mask.shape) == 3:
|
269 |
-
mask = mask[:, :, 0]
|
270 |
-
img_inpainted = inpaint_img_with_builded_lama(
|
271 |
-
model["lama"], image, mask, lama_config, device=device
|
272 |
-
)
|
273 |
-
return img_inpainted
|
274 |
-
|
275 |
-
|
276 |
-
# get args
|
277 |
-
parser = argparse.ArgumentParser()
|
278 |
-
setup_args(parser)
|
279 |
-
args = parser.parse_args(sys.argv[1:])
|
280 |
-
# build models
|
281 |
-
model = {}
|
282 |
-
# build the sam model
|
283 |
-
model_type = "vit_h"
|
284 |
-
ckpt_p = args.sam_ckpt
|
285 |
-
model_sam = sam_model_registry[model_type](checkpoint=ckpt_p)
|
286 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
287 |
-
model_sam.to(device=device)
|
288 |
-
model["sam"] = SamPredictor(model_sam)
|
289 |
-
|
290 |
-
# build the lama model
|
291 |
-
lama_config = args.lama_config
|
292 |
-
lama_ckpt = args.lama_ckpt
|
293 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
294 |
-
model["lama"] = build_lama_model(lama_config, lama_ckpt, device=device)
|
295 |
-
|
296 |
-
button_size = (100, 50)
|
297 |
-
with gr.Blocks() as demo:
|
298 |
-
clicked_points = gr.State([])
|
299 |
-
# origin_image = gr.State(None)
|
300 |
-
click_mask = gr.State(None)
|
301 |
-
features = gr.State(None)
|
302 |
-
orig_h = gr.State(None)
|
303 |
-
orig_w = gr.State(None)
|
304 |
-
input_h = gr.State(None)
|
305 |
-
input_w = gr.State(None)
|
306 |
-
|
307 |
-
with gr.Row():
|
308 |
-
with gr.Column(variant="panel"):
|
309 |
-
with gr.Row():
|
310 |
-
gr.Markdown("## Upload an image and click the region you want to edit.")
|
311 |
-
with gr.Row():
|
312 |
-
source_image_click = gr.Image(
|
313 |
-
type="numpy",
|
314 |
-
interactive=True,
|
315 |
-
label="Upload and Edit Image",
|
316 |
-
)
|
317 |
-
|
318 |
-
image_edit_complete = gr.Image(
|
319 |
-
type="numpy",
|
320 |
-
interactive=False,
|
321 |
-
label="Editing Complete",
|
322 |
-
)
|
323 |
-
with gr.Row():
|
324 |
-
point_prompt = gr.Radio(
|
325 |
-
choices=["Foreground Point", "Background Point"],
|
326 |
-
value="Foreground Point",
|
327 |
-
label="Point Label",
|
328 |
-
interactive=True,
|
329 |
-
show_label=False,
|
330 |
-
)
|
331 |
-
image_resolution = gr.Slider(
|
332 |
-
label="Image Resolution",
|
333 |
-
minimum=256,
|
334 |
-
maximum=768,
|
335 |
-
value=512,
|
336 |
-
step=64,
|
337 |
-
)
|
338 |
-
dilate_kernel_size = gr.Slider(
|
339 |
-
label="Dilate Kernel Size", minimum=0, maximum=30, value=15, step=1
|
340 |
-
)
|
341 |
-
with gr.Column(variant="panel"):
|
342 |
-
with gr.Row():
|
343 |
-
gr.Markdown("## Control Panel")
|
344 |
-
text_prompt = gr.Textbox(label="Text Prompt")
|
345 |
-
lama = gr.Button("Inpaint Image", variant="primary")
|
346 |
-
fill_sd = gr.Button("Fill Anything with SD", variant="primary")
|
347 |
-
replace_sd = gr.Button("Replace Anything with SD", variant="primary")
|
348 |
-
clear_button_image = gr.Button(value="Reset", variant="secondary")
|
349 |
-
|
350 |
-
# todo: maybe we can delete this row, for it's unnecessary to show the original mask for customers
|
351 |
-
with gr.Row(variant="panel"):
|
352 |
-
with gr.Column():
|
353 |
-
with gr.Row():
|
354 |
-
gr.Markdown("## Mask")
|
355 |
-
with gr.Row():
|
356 |
-
click_mask = gr.Image(
|
357 |
-
type="numpy",
|
358 |
-
label="Click Mask",
|
359 |
-
interactive=False,
|
360 |
-
)
|
361 |
-
with gr.Column():
|
362 |
-
with gr.Row():
|
363 |
-
gr.Markdown("## Image Removed with Mask")
|
364 |
-
with gr.Row():
|
365 |
-
img_rm_with_mask = gr.Image(
|
366 |
-
type="numpy",
|
367 |
-
label="Image Removed with Mask",
|
368 |
-
interactive=False,
|
369 |
-
)
|
370 |
-
|
371 |
-
with gr.Column():
|
372 |
-
with gr.Row():
|
373 |
-
gr.Markdown("## Fill Anything with Mask")
|
374 |
-
with gr.Row():
|
375 |
-
img_fill_with_mask = gr.Image(
|
376 |
-
type="numpy",
|
377 |
-
label="Image Fill Anything with Mask",
|
378 |
-
interactive=False,
|
379 |
-
)
|
380 |
-
|
381 |
-
with gr.Column():
|
382 |
-
with gr.Row():
|
383 |
-
gr.Markdown("## Replace Anything with Mask")
|
384 |
-
with gr.Row():
|
385 |
-
img_replace_with_mask = gr.Image(
|
386 |
-
type="numpy",
|
387 |
-
label="Image Replace Anything with Mask",
|
388 |
-
interactive=False,
|
389 |
-
)
|
390 |
-
|
391 |
-
|
392 |
-
|
393 |
-
|
394 |
-
|
395 |
-
|
396 |
-
|
397 |
-
|
398 |
-
|
399 |
-
|
400 |
-
|
401 |
-
|
402 |
-
|
403 |
-
|
404 |
-
|
405 |
-
|
406 |
-
|
407 |
-
|
408 |
-
|
409 |
-
|
410 |
-
|
411 |
-
|
412 |
-
|
413 |
-
|
414 |
-
|
415 |
-
|
416 |
-
|
417 |
-
|
418 |
-
|
419 |
-
|
420 |
-
|
421 |
-
|
422 |
-
|
423 |
-
|
424 |
-
|
425 |
-
|
426 |
-
|
427 |
-
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
436 |
-
|
437 |
-
|
438 |
-
|
439 |
-
|
440 |
-
|
441 |
-
|
442 |
-
|
443 |
-
|
444 |
-
|
445 |
-
|
446 |
-
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
|
451 |
-
|
452 |
-
|
453 |
-
|
454 |
-
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
|
462 |
-
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
|
4 |
+
sys.path.append(os.path.abspath(os.path.dirname(os.getcwd())))
|
5 |
+
# os.chdir("../")
|
6 |
+
import cv2
|
7 |
+
import gradio as gr
|
8 |
+
import numpy as np
|
9 |
+
from pathlib import Path
|
10 |
+
from matplotlib import pyplot as plt
|
11 |
+
import torch
|
12 |
+
import tempfile
|
13 |
+
|
14 |
+
from stable_diffusion_inpaint import fill_img_with_sd, replace_img_with_sd
|
15 |
+
from lama_inpaint import (
|
16 |
+
inpaint_img_with_lama,
|
17 |
+
build_lama_model,
|
18 |
+
inpaint_img_with_builded_lama,
|
19 |
+
)
|
20 |
+
from utils import (
|
21 |
+
load_img_to_array,
|
22 |
+
save_array_to_img,
|
23 |
+
dilate_mask,
|
24 |
+
show_mask,
|
25 |
+
show_points,
|
26 |
+
)
|
27 |
+
from PIL import Image
|
28 |
+
from segment_anything import SamPredictor, sam_model_registry
|
29 |
+
import argparse
|
30 |
+
|
31 |
+
|
32 |
+
def setup_args(parser):
|
33 |
+
parser.add_argument(
|
34 |
+
"--lama_config",
|
35 |
+
type=str,
|
36 |
+
default="./lama/configs/prediction/default.yaml",
|
37 |
+
help="The path to the config file of lama model. "
|
38 |
+
"Default: the config of big-lama",
|
39 |
+
)
|
40 |
+
parser.add_argument(
|
41 |
+
"--lama_ckpt",
|
42 |
+
type=str,
|
43 |
+
default="./pretrained_models/big-lama",
|
44 |
+
help="The path to the lama checkpoint.",
|
45 |
+
)
|
46 |
+
parser.add_argument(
|
47 |
+
"--sam_ckpt",
|
48 |
+
type=str,
|
49 |
+
default="./pretrained_models/sam_vit_h_4b8939.pth",
|
50 |
+
help="The path to the SAM checkpoint to use for mask generation.",
|
51 |
+
)
|
52 |
+
|
53 |
+
|
54 |
+
def mkstemp(suffix, dir=None):
|
55 |
+
fd, path = tempfile.mkstemp(suffix=f"{suffix}", dir=dir)
|
56 |
+
os.close(fd)
|
57 |
+
return Path(path)
|
58 |
+
|
59 |
+
|
60 |
+
def get_sam_feat(img):
|
61 |
+
model["sam"].set_image(img)
|
62 |
+
features = model["sam"].features
|
63 |
+
orig_h = model["sam"].orig_h
|
64 |
+
orig_w = model["sam"].orig_w
|
65 |
+
input_h = model["sam"].input_h
|
66 |
+
input_w = model["sam"].input_w
|
67 |
+
model["sam"].reset_image()
|
68 |
+
return features, orig_h, orig_w, input_h, input_w
|
69 |
+
|
70 |
+
|
71 |
+
def get_fill_img_with_sd(image, mask, image_resolution, text_prompt):
|
72 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
73 |
+
if len(mask.shape) == 3:
|
74 |
+
mask = mask[:, :, 0]
|
75 |
+
np_image = np.array(image, dtype=np.uint8)
|
76 |
+
H, W, C = np_image.shape
|
77 |
+
np_image = HWC3(np_image)
|
78 |
+
np_image = resize_image(np_image, image_resolution)
|
79 |
+
mask = cv2.resize(
|
80 |
+
mask, (np_image.shape[1], np_image.shape[0]), interpolation=cv2.INTER_NEAREST
|
81 |
+
)
|
82 |
+
|
83 |
+
img_fill = fill_img_with_sd(np_image, mask, text_prompt, device=device)
|
84 |
+
img_fill = img_fill.astype(np.uint8)
|
85 |
+
return img_fill
|
86 |
+
|
87 |
+
|
88 |
+
def get_replace_img_with_sd(image, mask, image_resolution, text_prompt):
|
89 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
90 |
+
if len(mask.shape) == 3:
|
91 |
+
mask = mask[:, :, 0]
|
92 |
+
np_image = np.array(image, dtype=np.uint8)
|
93 |
+
H, W, C = np_image.shape
|
94 |
+
np_image = HWC3(np_image)
|
95 |
+
np_image = resize_image(np_image, image_resolution)
|
96 |
+
mask = cv2.resize(
|
97 |
+
mask, (np_image.shape[1], np_image.shape[0]), interpolation=cv2.INTER_NEAREST
|
98 |
+
)
|
99 |
+
|
100 |
+
img_replaced = replace_img_with_sd(np_image, mask, text_prompt, device=device)
|
101 |
+
img_replaced = img_replaced.astype(np.uint8)
|
102 |
+
return img_replaced
|
103 |
+
|
104 |
+
|
105 |
+
def HWC3(x):
|
106 |
+
assert x.dtype == np.uint8
|
107 |
+
if x.ndim == 2:
|
108 |
+
x = x[:, :, None]
|
109 |
+
assert x.ndim == 3
|
110 |
+
H, W, C = x.shape
|
111 |
+
assert C == 1 or C == 3 or C == 4
|
112 |
+
if C == 3:
|
113 |
+
return x
|
114 |
+
if C == 1:
|
115 |
+
return np.concatenate([x, x, x], axis=2)
|
116 |
+
if C == 4:
|
117 |
+
color = x[:, :, 0:3].astype(np.float32)
|
118 |
+
alpha = x[:, :, 3:4].astype(np.float32) / 255.0
|
119 |
+
y = color * alpha + 255.0 * (1.0 - alpha)
|
120 |
+
y = y.clip(0, 255).astype(np.uint8)
|
121 |
+
return y
|
122 |
+
|
123 |
+
|
124 |
+
def resize_image(input_image, resolution):
|
125 |
+
H, W, C = input_image.shape
|
126 |
+
k = float(resolution) / min(H, W)
|
127 |
+
H = int(np.round(H * k / 64.0)) * 64
|
128 |
+
W = int(np.round(W * k / 64.0)) * 64
|
129 |
+
img = cv2.resize(
|
130 |
+
input_image,
|
131 |
+
(W, H),
|
132 |
+
interpolation=cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA,
|
133 |
+
)
|
134 |
+
return img
|
135 |
+
|
136 |
+
|
137 |
+
def resize_points(clicked_points, original_shape, resolution):
|
138 |
+
original_height, original_width, _ = original_shape
|
139 |
+
original_height = float(original_height)
|
140 |
+
original_width = float(original_width)
|
141 |
+
|
142 |
+
scale_factor = float(resolution) / min(original_height, original_width)
|
143 |
+
resized_points = []
|
144 |
+
|
145 |
+
for point in clicked_points:
|
146 |
+
x, y, lab = point
|
147 |
+
resized_x = int(round(x * scale_factor))
|
148 |
+
resized_y = int(round(y * scale_factor))
|
149 |
+
resized_point = (resized_x, resized_y, lab)
|
150 |
+
resized_points.append(resized_point)
|
151 |
+
|
152 |
+
return resized_points
|
153 |
+
|
154 |
+
|
155 |
+
def get_click_mask(
|
156 |
+
clicked_points, features, orig_h, orig_w, input_h, input_w, dilate_kernel_size
|
157 |
+
):
|
158 |
+
# model['sam'].set_image(image)
|
159 |
+
model["sam"].is_image_set = True
|
160 |
+
model["sam"].features = features
|
161 |
+
model["sam"].orig_h = orig_h
|
162 |
+
model["sam"].orig_w = orig_w
|
163 |
+
model["sam"].input_h = input_h
|
164 |
+
model["sam"].input_w = input_w
|
165 |
+
|
166 |
+
# Separate the points and labels
|
167 |
+
points, labels = zip(*[(point[:2], point[2]) for point in clicked_points])
|
168 |
+
|
169 |
+
# Convert the points and labels to numpy arrays
|
170 |
+
input_point = np.array(points)
|
171 |
+
input_label = np.array(labels)
|
172 |
+
|
173 |
+
masks, _, _ = model["sam"].predict(
|
174 |
+
point_coords=input_point,
|
175 |
+
point_labels=input_label,
|
176 |
+
multimask_output=False,
|
177 |
+
)
|
178 |
+
if dilate_kernel_size is not None:
|
179 |
+
masks = [dilate_mask(mask, dilate_kernel_size) for mask in masks]
|
180 |
+
else:
|
181 |
+
masks = [mask for mask in masks]
|
182 |
+
|
183 |
+
return masks
|
184 |
+
|
185 |
+
|
186 |
+
def process_image_click(
|
187 |
+
original_image,
|
188 |
+
point_prompt,
|
189 |
+
clicked_points,
|
190 |
+
image_resolution,
|
191 |
+
features,
|
192 |
+
orig_h,
|
193 |
+
orig_w,
|
194 |
+
input_h,
|
195 |
+
input_w,
|
196 |
+
dilate_kernel_size,
|
197 |
+
evt: gr.SelectData,
|
198 |
+
):
|
199 |
+
if clicked_points is None:
|
200 |
+
clicked_points = []
|
201 |
+
|
202 |
+
# print("Received click event:", evt)
|
203 |
+
if original_image is None:
|
204 |
+
# print("No image loaded.")
|
205 |
+
return None, clicked_points, None
|
206 |
+
|
207 |
+
clicked_coords = evt.index
|
208 |
+
if clicked_coords is None:
|
209 |
+
# print("No valid coordinates received.")
|
210 |
+
return None, clicked_points, None
|
211 |
+
|
212 |
+
x, y = clicked_coords
|
213 |
+
label = point_prompt
|
214 |
+
lab = 1 if label == "Foreground Point" else 0
|
215 |
+
clicked_points.append((x, y, lab))
|
216 |
+
# print("Updated points list:", clicked_points)
|
217 |
+
|
218 |
+
input_image = np.array(original_image, dtype=np.uint8)
|
219 |
+
H, W, C = input_image.shape
|
220 |
+
input_image = HWC3(input_image)
|
221 |
+
img = resize_image(input_image, image_resolution)
|
222 |
+
# print("Processed image size:", img.shape)
|
223 |
+
|
224 |
+
resized_points = resize_points(clicked_points, input_image.shape, image_resolution)
|
225 |
+
mask_click_np = get_click_mask(
|
226 |
+
resized_points, features, orig_h, orig_w, input_h, input_w, dilate_kernel_size
|
227 |
+
)
|
228 |
+
mask_click_np = np.transpose(mask_click_np, (1, 2, 0)) * 255.0
|
229 |
+
mask_image = HWC3(mask_click_np.astype(np.uint8))
|
230 |
+
mask_image = cv2.resize(mask_image, (W, H), interpolation=cv2.INTER_LINEAR)
|
231 |
+
# print("Mask image prepared.")
|
232 |
+
|
233 |
+
edited_image = input_image
|
234 |
+
for x, y, lab in clicked_points:
|
235 |
+
color = (255, 0, 0) if lab == 1 else (0, 0, 255)
|
236 |
+
edited_image = cv2.circle(edited_image, (x, y), 20, color, -1)
|
237 |
+
|
238 |
+
opacity_mask = 0.75
|
239 |
+
opacity_edited = 1.0
|
240 |
+
overlay_image = cv2.addWeighted(
|
241 |
+
edited_image,
|
242 |
+
opacity_edited,
|
243 |
+
(mask_image * np.array([0 / 255, 255 / 255, 0 / 255])).astype(np.uint8),
|
244 |
+
opacity_mask,
|
245 |
+
0,
|
246 |
+
)
|
247 |
+
|
248 |
+
no_mask_overlay = edited_image.copy()
|
249 |
+
|
250 |
+
return no_mask_overlay, overlay_image, clicked_points, mask_image
|
251 |
+
|
252 |
+
|
253 |
+
def image_upload(image, image_resolution):
|
254 |
+
if image is None:
|
255 |
+
return None, None, None, None, None, None
|
256 |
+
else:
|
257 |
+
np_image = np.array(image, dtype=np.uint8)
|
258 |
+
H, W, C = np_image.shape
|
259 |
+
np_image = HWC3(np_image)
|
260 |
+
np_image = resize_image(np_image, image_resolution)
|
261 |
+
features, orig_h, orig_w, input_h, input_w = get_sam_feat(np_image)
|
262 |
+
return image, features, orig_h, orig_w, input_h, input_w
|
263 |
+
|
264 |
+
|
265 |
+
def get_inpainted_img(image, mask, image_resolution):
|
266 |
+
lama_config = args.lama_config
|
267 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
268 |
+
if len(mask.shape) == 3:
|
269 |
+
mask = mask[:, :, 0]
|
270 |
+
img_inpainted = inpaint_img_with_builded_lama(
|
271 |
+
model["lama"], image, mask, lama_config, device=device
|
272 |
+
)
|
273 |
+
return img_inpainted
|
274 |
+
|
275 |
+
|
276 |
+
# get args
|
277 |
+
parser = argparse.ArgumentParser()
|
278 |
+
setup_args(parser)
|
279 |
+
args = parser.parse_args(sys.argv[1:])
|
280 |
+
# build models
|
281 |
+
model = {}
|
282 |
+
# build the sam model
|
283 |
+
model_type = "vit_h"
|
284 |
+
ckpt_p = args.sam_ckpt
|
285 |
+
model_sam = sam_model_registry[model_type](checkpoint=ckpt_p)
|
286 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
287 |
+
model_sam.to(device=device)
|
288 |
+
model["sam"] = SamPredictor(model_sam)
|
289 |
+
|
290 |
+
# build the lama model
|
291 |
+
lama_config = args.lama_config
|
292 |
+
lama_ckpt = args.lama_ckpt
|
293 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
294 |
+
model["lama"] = build_lama_model(lama_config, lama_ckpt, device=device)
|
295 |
+
|
296 |
+
button_size = (100, 50)
|
297 |
+
with gr.Blocks() as demo:
|
298 |
+
clicked_points = gr.State([])
|
299 |
+
# origin_image = gr.State(None)
|
300 |
+
click_mask = gr.State(None)
|
301 |
+
features = gr.State(None)
|
302 |
+
orig_h = gr.State(None)
|
303 |
+
orig_w = gr.State(None)
|
304 |
+
input_h = gr.State(None)
|
305 |
+
input_w = gr.State(None)
|
306 |
+
|
307 |
+
with gr.Row():
|
308 |
+
with gr.Column(variant="panel"):
|
309 |
+
with gr.Row():
|
310 |
+
gr.Markdown("## Upload an image and click the region you want to edit.")
|
311 |
+
with gr.Row():
|
312 |
+
source_image_click = gr.Image(
|
313 |
+
type="numpy",
|
314 |
+
interactive=True,
|
315 |
+
label="Upload and Edit Image",
|
316 |
+
)
|
317 |
+
|
318 |
+
image_edit_complete = gr.Image(
|
319 |
+
type="numpy",
|
320 |
+
interactive=False,
|
321 |
+
label="Editing Complete",
|
322 |
+
)
|
323 |
+
with gr.Row():
|
324 |
+
point_prompt = gr.Radio(
|
325 |
+
choices=["Foreground Point", "Background Point"],
|
326 |
+
value="Foreground Point",
|
327 |
+
label="Point Label",
|
328 |
+
interactive=True,
|
329 |
+
show_label=False,
|
330 |
+
)
|
331 |
+
image_resolution = gr.Slider(
|
332 |
+
label="Image Resolution",
|
333 |
+
minimum=256,
|
334 |
+
maximum=768,
|
335 |
+
value=512,
|
336 |
+
step=64,
|
337 |
+
)
|
338 |
+
dilate_kernel_size = gr.Slider(
|
339 |
+
label="Dilate Kernel Size", minimum=0, maximum=30, value=15, step=1
|
340 |
+
)
|
341 |
+
with gr.Column(variant="panel"):
|
342 |
+
with gr.Row():
|
343 |
+
gr.Markdown("## Control Panel")
|
344 |
+
text_prompt = gr.Textbox(label="Text Prompt")
|
345 |
+
lama = gr.Button("Inpaint Image", variant="primary")
|
346 |
+
fill_sd = gr.Button("Fill Anything with SD", variant="primary")
|
347 |
+
replace_sd = gr.Button("Replace Anything with SD", variant="primary")
|
348 |
+
clear_button_image = gr.Button(value="Reset", variant="secondary")
|
349 |
+
|
350 |
+
# todo: maybe we can delete this row, for it's unnecessary to show the original mask for customers
|
351 |
+
with gr.Row(variant="panel"):
|
352 |
+
with gr.Column():
|
353 |
+
with gr.Row():
|
354 |
+
gr.Markdown("## Mask")
|
355 |
+
with gr.Row():
|
356 |
+
click_mask = gr.Image(
|
357 |
+
type="numpy",
|
358 |
+
label="Click Mask",
|
359 |
+
interactive=False,
|
360 |
+
)
|
361 |
+
with gr.Column():
|
362 |
+
with gr.Row():
|
363 |
+
gr.Markdown("## Image Removed with Mask")
|
364 |
+
with gr.Row():
|
365 |
+
img_rm_with_mask = gr.Image(
|
366 |
+
type="numpy",
|
367 |
+
label="Image Removed with Mask",
|
368 |
+
interactive=False,
|
369 |
+
)
|
370 |
+
|
371 |
+
with gr.Column():
|
372 |
+
with gr.Row():
|
373 |
+
gr.Markdown("## Fill Anything with Mask")
|
374 |
+
with gr.Row():
|
375 |
+
img_fill_with_mask = gr.Image(
|
376 |
+
type="numpy",
|
377 |
+
label="Image Fill Anything with Mask",
|
378 |
+
interactive=False,
|
379 |
+
)
|
380 |
+
|
381 |
+
with gr.Column():
|
382 |
+
with gr.Row():
|
383 |
+
gr.Markdown("## Replace Anything with Mask")
|
384 |
+
with gr.Row():
|
385 |
+
img_replace_with_mask = gr.Image(
|
386 |
+
type="numpy",
|
387 |
+
label="Image Replace Anything with Mask",
|
388 |
+
interactive=False,
|
389 |
+
)
|
390 |
+
|
391 |
+
gr.Markdown(
|
392 |
+
"Github Source Code: [Link](https://github.com/pg56714/Inpaint-Anything-Gradio)"
|
393 |
+
)
|
394 |
+
|
395 |
+
source_image_click.upload(
|
396 |
+
image_upload,
|
397 |
+
inputs=[source_image_click, image_resolution],
|
398 |
+
outputs=[source_image_click, features, orig_h, orig_w, input_h, input_w],
|
399 |
+
)
|
400 |
+
|
401 |
+
source_image_click.select(
|
402 |
+
process_image_click,
|
403 |
+
inputs=[
|
404 |
+
source_image_click,
|
405 |
+
point_prompt,
|
406 |
+
clicked_points,
|
407 |
+
image_resolution,
|
408 |
+
features,
|
409 |
+
orig_h,
|
410 |
+
orig_w,
|
411 |
+
input_h,
|
412 |
+
input_w,
|
413 |
+
dilate_kernel_size,
|
414 |
+
],
|
415 |
+
outputs=[source_image_click, image_edit_complete, clicked_points, click_mask],
|
416 |
+
show_progress=True,
|
417 |
+
queue=True,
|
418 |
+
)
|
419 |
+
|
420 |
+
lama.click(
|
421 |
+
get_inpainted_img,
|
422 |
+
inputs=[source_image_click, click_mask, image_resolution],
|
423 |
+
outputs=[img_rm_with_mask],
|
424 |
+
)
|
425 |
+
|
426 |
+
fill_sd.click(
|
427 |
+
get_fill_img_with_sd,
|
428 |
+
inputs=[source_image_click, click_mask, image_resolution, text_prompt],
|
429 |
+
outputs=[img_fill_with_mask],
|
430 |
+
)
|
431 |
+
|
432 |
+
replace_sd.click(
|
433 |
+
get_replace_img_with_sd,
|
434 |
+
inputs=[source_image_click, click_mask, image_resolution, text_prompt],
|
435 |
+
outputs=[img_replace_with_mask],
|
436 |
+
)
|
437 |
+
|
438 |
+
def reset(*args):
|
439 |
+
return [None for _ in args]
|
440 |
+
|
441 |
+
clear_button_image.click(
|
442 |
+
reset,
|
443 |
+
inputs=[
|
444 |
+
source_image_click,
|
445 |
+
image_edit_complete,
|
446 |
+
clicked_points,
|
447 |
+
click_mask,
|
448 |
+
features,
|
449 |
+
img_rm_with_mask,
|
450 |
+
img_fill_with_mask,
|
451 |
+
img_replace_with_mask,
|
452 |
+
],
|
453 |
+
outputs=[
|
454 |
+
source_image_click,
|
455 |
+
image_edit_complete,
|
456 |
+
clicked_points,
|
457 |
+
click_mask,
|
458 |
+
features,
|
459 |
+
img_rm_with_mask,
|
460 |
+
img_fill_with_mask,
|
461 |
+
img_replace_with_mask,
|
462 |
+
],
|
463 |
+
)
|
464 |
+
|
465 |
+
if __name__ == "__main__":
|
466 |
+
demo.launch(debug=False, show_error=True)
|