-
Notifications
You must be signed in to change notification settings - Fork 2
/
waifu2x_transparent_2.py
139 lines (107 loc) · 3.71 KB
/
waifu2x_transparent_2.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
#!/usr/bin/env python3
import cv2
import numpy as np
import skimage.transform
from utils import do_imgs, floor_even, read_img, trim_img, untrim_img, write_img
from waifu2x import pad_size, run_img
model_filename = "./models/waifu2x/noise0_scale2x.onnx"
in_filenames = [
"./in.png",
]
out_suffix = "_waifu2x"
pre_blur = 0.5
pre_darken = False
pre_lighten = False
trim_eps = 1e-3
waifu2x_strength = 0.7
waifu2x_alpha = False
alpha_blur_scale = 0.7
alpha_blur_gamma = 1
alpha_blur_strength = 0.7
use_gpu = True
if use_gpu:
import cupy as cp
else:
cp = np
cp.asnumpy = lambda x: x
def cross_sum(a, out):
a = cp.pad(a, ((1, 1), (1, 1), (0, 0)))
out[:] = a[2:, 1:-1] + a[:-2, 1:-1] + a[1:-1, 2:] + a[1:-1, :-2]
def bleed_alpha(img, alpha, eps, max_iter=10**3, tol=1e-4):
img = cp.asarray(img)
alpha = cp.asarray(alpha)
assert alpha.ndim == 2
alpha = alpha[:, :, None]
mask_0 = alpha > eps
mask = mask_0.astype(img.dtype)
# TODO: np.maximum breaks type stability in numba
confidence = cp.maximum(alpha, eps).astype(img.dtype)
del alpha
img_new = cp.zeros_like(img)
for iter_count in range(max_iter):
cross_sum(mask * confidence * img, img_new)
cross_sum(mask * confidence, mask)
img_new /= cp.maximum(mask, 1e-7).astype(mask.dtype)
img_new = cp.where(mask_0, img, img_new)
norm = ((img_new - img) ** 2).max()
if norm < tol:
break
img, img_new = img_new, img
mask = (mask > 0).astype(mask.dtype)
img_new = cp.asnumpy(img_new)
return iter_count + 1, img_new
def do_blur(img):
if pre_blur:
img_blur = cv2.GaussianBlur(img, (0, 0), pre_blur)
else:
img_blur = img
if pre_darken:
assert pre_blur
assert not pre_lighten
img = np.minimum(img, img_blur)
elif pre_lighten:
assert pre_blur
img = np.maximum(img, img_blur)
else:
img = img_blur
return img
def convert_img(sess, in_filename, out_filename):
# Network input is BGR
img, alpha = read_img(in_filename, swap_rb=False, signed=False, return_alpha=True)
original_shape, (trim_t, trim_b, trim_l, trim_r) = trim_img(
img, alpha, trim_eps, pad=pad_size
)
trim_b = trim_t + floor_even(trim_b - trim_t)
trim_r = trim_l + floor_even(trim_r - trim_l)
img = img[trim_t:trim_b, trim_l:trim_r, :]
alpha = alpha[trim_t:trim_b, trim_l:trim_r]
iter_count, img = bleed_alpha(img, alpha, trim_eps)
print("Bleed alpha iter", iter_count)
img_out = do_blur(img)
img_out = skimage.transform.resize(img_out, (img.shape[0] // 2, img.shape[1] // 2))
img_out = run_img(sess, img_out)
img = (1 - waifu2x_strength) * img + waifu2x_strength * img_out
del img_out
alpha_blur = do_blur(alpha)
if waifu2x_alpha:
alpha_blur = skimage.transform.resize(
alpha_blur, (alpha.shape[0] // 2, alpha.shape[1] // 2)
)
alpha_blur = np.repeat(alpha_blur[:, :, None], 3, axis=2)
alpha_blur = run_img(sess, alpha_blur)
alpha_blur = alpha_blur.mean(axis=2)
else:
alpha_blur = skimage.transform.resize(
alpha_blur, tuple(int(alpha_blur_scale * x) for x in alpha.shape)
)
alpha_blur = skimage.transform.resize(alpha_blur, alpha.shape)
alpha_blur **= alpha_blur_gamma
alpha = (1 - alpha_blur_strength) * alpha + alpha_blur_strength * alpha_blur
del alpha_blur
img, alpha = untrim_img(
img, alpha, original_shape, (trim_t, trim_b, trim_l, trim_r)
)
# Network output is BGR
write_img(out_filename, img, alpha, swap_rb=False, signed=False)
if __name__ == "__main__":
do_imgs(convert_img, model_filename, in_filenames, out_suffix)