GFPGAN

1
import argparse
2
import cv2
3
import glob
4
import numpy as np
5
import os
6
import torch
7
from basicsr.utils import imwrite
8

9
from gfpgan import GFPGANer
10

11

12
def main():
13
    """Inference demo for GFPGAN (for users).
14
    """
15
    parser = argparse.ArgumentParser()
16
    parser.add_argument(
17
        '-i',
18
        '--input',
19
        type=str,
20
        default='inputs/whole_imgs',
21
        help='Input image or folder. Default: inputs/whole_imgs')
22
    parser.add_argument('-o', '--output', type=str, default='results', help='Output folder. Default: results')
23
    # we use version to select models, which is more user-friendly
24
    parser.add_argument(
25
        '-v', '--version', type=str, default='1.3', help='GFPGAN model version. Option: 1 | 1.2 | 1.3. Default: 1.3')
26
    parser.add_argument(
27
        '-s', '--upscale', type=int, default=2, help='The final upsampling scale of the image. Default: 2')
28

29
    parser.add_argument(
30
        '--bg_upsampler', type=str, default='realesrgan', help='background upsampler. Default: realesrgan')
31
    parser.add_argument(
32
        '--bg_tile',
33
        type=int,
34
        default=400,
35
        help='Tile size for background sampler, 0 for no tile during testing. Default: 400')
36
    parser.add_argument('--suffix', type=str, default=None, help='Suffix of the restored faces')
37
    parser.add_argument('--only_center_face', action='store_true', help='Only restore the center face')
38
    parser.add_argument('--aligned', action='store_true', help='Input are aligned faces')
39
    parser.add_argument(
40
        '--ext',
41
        type=str,
42
        default='auto',
43
        help='Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto')
44
    parser.add_argument('-w', '--weight', type=float, default=0.5, help='Adjustable weights.')
45
    args = parser.parse_args()
46

47
    args = parser.parse_args()
48

49
    # ------------------------ input & output ------------------------
50
    if args.input.endswith('/'):
51
        args.input = args.input[:-1]
52
    if os.path.isfile(args.input):
53
        img_list = [args.input]
54
    else:
55
        img_list = sorted(glob.glob(os.path.join(args.input, '*')))
56

57
    os.makedirs(args.output, exist_ok=True)
58

59
    # ------------------------ set up background upsampler ------------------------
60
    if args.bg_upsampler == 'realesrgan':
61
        if not torch.cuda.is_available():  # CPU
62
            import warnings
63
            warnings.warn('The unoptimized RealESRGAN is slow on CPU. We do not use it. '
64
                          'If you really want to use it, please modify the corresponding codes.')
65
            bg_upsampler = None
66
        else:
67
            from basicsr.archs.rrdbnet_arch import RRDBNet
68
            from realesrgan import RealESRGANer
69
            model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
70
            bg_upsampler = RealESRGANer(
71
                scale=2,
72
                model_path='https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth',
73
                model=model,
74
                tile=args.bg_tile,
75
                tile_pad=10,
76
                pre_pad=0,
77
                half=True)  # need to set False in CPU mode
78
    else:
79
        bg_upsampler = None
80

81
    # ------------------------ set up GFPGAN restorer ------------------------
82
    if args.version == '1':
83
        arch = 'original'
84
        channel_multiplier = 1
85
        model_name = 'GFPGANv1'
86
        url = 'https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth'
87
    elif args.version == '1.2':
88
        arch = 'clean'
89
        channel_multiplier = 2
90
        model_name = 'GFPGANCleanv1-NoCE-C2'
91
        url = 'https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth'
92
    elif args.version == '1.3':
93
        arch = 'clean'
94
        channel_multiplier = 2
95
        model_name = 'GFPGANv1.3'
96
        url = 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth'
97
    elif args.version == '1.4':
98
        arch = 'clean'
99
        channel_multiplier = 2
100
        model_name = 'GFPGANv1.4'
101
        url = 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth'
102
    elif args.version == 'RestoreFormer':
103
        arch = 'RestoreFormer'
104
        channel_multiplier = 2
105
        model_name = 'RestoreFormer'
106
        url = 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.4/RestoreFormer.pth'
107
    else:
108
        raise ValueError(f'Wrong model version {args.version}.')
109

110
    # determine model paths
111
    model_path = os.path.join('experiments/pretrained_models', model_name + '.pth')
112
    if not os.path.isfile(model_path):
113
        model_path = os.path.join('gfpgan/weights', model_name + '.pth')
114
    if not os.path.isfile(model_path):
115
        # download pre-trained models from url
116
        model_path = url
117

118
    restorer = GFPGANer(
119
        model_path=model_path,
120
        upscale=args.upscale,
121
        arch=arch,
122
        channel_multiplier=channel_multiplier,
123
        bg_upsampler=bg_upsampler)
124

125
    # ------------------------ restore ------------------------
126
    for img_path in img_list:
127
        # read image
128
        img_name = os.path.basename(img_path)
129
        print(f'Processing {img_name} ...')
130
        basename, ext = os.path.splitext(img_name)
131
        input_img = cv2.imread(img_path, cv2.IMREAD_COLOR)
132

133
        # restore faces and background if necessary
134
        cropped_faces, restored_faces, restored_img = restorer.enhance(
135
            input_img,
136
            has_aligned=args.aligned,
137
            only_center_face=args.only_center_face,
138
            paste_back=True,
139
            weight=args.weight)
140

141
        # save faces
142
        for idx, (cropped_face, restored_face) in enumerate(zip(cropped_faces, restored_faces)):
143
            # save cropped face
144
            save_crop_path = os.path.join(args.output, 'cropped_faces', f'{basename}_{idx:02d}.png')
145
            imwrite(cropped_face, save_crop_path)
146
            # save restored face
147
            if args.suffix is not None:
148
                save_face_name = f'{basename}_{idx:02d}_{args.suffix}.png'
149
            else:
150
                save_face_name = f'{basename}_{idx:02d}.png'
151
            save_restore_path = os.path.join(args.output, 'restored_faces', save_face_name)
152
            imwrite(restored_face, save_restore_path)
153
            # save comparison image
154
            cmp_img = np.concatenate((cropped_face, restored_face), axis=1)
155
            imwrite(cmp_img, os.path.join(args.output, 'cmp', f'{basename}_{idx:02d}.png'))
156

157
        # save restored img
158
        if restored_img is not None:
159
            if args.ext == 'auto':
160
                extension = ext[1:]
161
            else:
162
                extension = args.ext
163

164
            if args.suffix is not None:
165
                save_restore_path = os.path.join(args.output, 'restored_imgs', f'{basename}_{args.suffix}.{extension}')
166
            else:
167
                save_restore_path = os.path.join(args.output, 'restored_imgs', f'{basename}.{extension}')
168
            imwrite(restored_img, save_restore_path)
169

170
    print(f'Results are in the [{args.output}] folder.')
171

172

173
if __name__ == '__main__':
174
    main()
175