Spaces:
Running
Running
enhancer
Browse files- image_enhancer.oy +124 -0
image_enhancer.oy
ADDED
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
from gfpgan import GFPGANer
|
4 |
+
from tqdm import tqdm
|
5 |
+
import cv2
|
6 |
+
from enum import Enum
|
7 |
+
|
8 |
+
class EnhancementMethod(str, Enum):
|
9 |
+
gfpgan = "gfpgan"
|
10 |
+
RestoreFormer = "RestoreFormer"
|
11 |
+
codeformer = "codeformer"
|
12 |
+
realesrgan = "realesrgan"
|
13 |
+
|
14 |
+
|
15 |
+
class Enhancer:
|
16 |
+
def __init__(self, method=EnhancementMethod, background_enhancement=True, upscale=2):
|
17 |
+
# Set up RealESRGAN for background enhancement
|
18 |
+
if background_enhancement:
|
19 |
+
if upscale == 2:
|
20 |
+
if not torch.cuda.is_available(): # CPU
|
21 |
+
import warnings
|
22 |
+
warnings.warn('The unoptimized RealESRGAN is slow on CPU. We do not use it. '
|
23 |
+
'If you really want to use it, please modify the corresponding codes.')
|
24 |
+
self.bg_upsampler = None
|
25 |
+
else:
|
26 |
+
from basicsr.archs.rrdbnet_arch import RRDBNet
|
27 |
+
from realesrgan import RealESRGANer
|
28 |
+
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
|
29 |
+
self.bg_upsampler = RealESRGANer(
|
30 |
+
scale=2,
|
31 |
+
model_path='https://huggingface.co/dtarnow/UPscaler/resolve/main/RealESRGAN_x2plus.pth',
|
32 |
+
model=model,
|
33 |
+
tile=400,
|
34 |
+
tile_pad=10,
|
35 |
+
pre_pad=0,
|
36 |
+
half=True) # need to set False in CPU mode
|
37 |
+
elif upscale == 4:
|
38 |
+
if not torch.cuda.is_available(): # CPU
|
39 |
+
import warnings
|
40 |
+
warnings.warn('The unoptimized RealESRGAN is slow on CPU. We do not use it. '
|
41 |
+
'If you really want to use it, please modify the corresponding codes.')
|
42 |
+
self.bg_upsampler = None
|
43 |
+
else:
|
44 |
+
from basicsr.archs.rrdbnet_arch import RRDBNet
|
45 |
+
from realesrgan import RealESRGANer
|
46 |
+
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
|
47 |
+
self.bg_upsampler = RealESRGANer(
|
48 |
+
scale=4,
|
49 |
+
model_path='https://huggingface.co/lllyasviel/Annotators/resolve/main/RealESRGAN_x4plus.pth',
|
50 |
+
model=model,
|
51 |
+
tile=400,
|
52 |
+
tile_pad=10,
|
53 |
+
pre_pad=0,
|
54 |
+
half=True) # need to set False in CPU mode
|
55 |
+
else:
|
56 |
+
raise ValueError(f'Wrong upscale constant {upscale}.')
|
57 |
+
else:
|
58 |
+
self.bg_upsampler = None
|
59 |
+
|
60 |
+
# Set up GPFGAN for face enhancement
|
61 |
+
if method == 'gfpgan':
|
62 |
+
self.arch = 'clean'
|
63 |
+
self.channel_multiplier = 2
|
64 |
+
self.model_name = 'GFPGANv1.4'
|
65 |
+
self.url = 'https://huggingface.co/gmk123/GFPGAN/resolve/main/GFPGANv1.4.pth'
|
66 |
+
elif method == 'RestoreFormer':
|
67 |
+
self.arch = 'RestoreFormer'
|
68 |
+
self.channel_multiplier = 2
|
69 |
+
self.model_name = 'RestoreFormer'
|
70 |
+
self.url = 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.4/RestoreFormer.pth'
|
71 |
+
elif method == 'codeformer': # TODO:
|
72 |
+
self.arch = 'CodeFormer'
|
73 |
+
self.channel_multiplier = 2
|
74 |
+
self.model_name = 'CodeFormer'
|
75 |
+
self.url = 'https://huggingface.co/sinadi/aar/resolve/main/codeformer.pth'
|
76 |
+
else:
|
77 |
+
raise ValueError(f'Wrong model version {method}.')
|
78 |
+
|
79 |
+
# Determine the model path and if the model is not available, download it
|
80 |
+
model_path = os.path.join('gfpgan/weights', self.model_name + '.pth')
|
81 |
+
|
82 |
+
if not os.path.isfile(model_path):
|
83 |
+
model_path = os.path.join('checkpoints', self.model_name + '.pth')
|
84 |
+
|
85 |
+
if not os.path.isfile(model_path):
|
86 |
+
# Download pre-trained models from url
|
87 |
+
model_path = self.url
|
88 |
+
|
89 |
+
self.restorer = GFPGANer(
|
90 |
+
model_path=model_path,
|
91 |
+
upscale=upscale,
|
92 |
+
arch=self.arch,
|
93 |
+
channel_multiplier=self.channel_multiplier,
|
94 |
+
bg_upsampler=self.bg_upsampler)
|
95 |
+
|
96 |
+
|
97 |
+
def check_image_dimensions(self, image):
|
98 |
+
# Get the dimensions of the image
|
99 |
+
height, width, _ = image.shape
|
100 |
+
return True
|
101 |
+
|
102 |
+
# Check if either dimension exceeds 2048 pixels :Todo
|
103 |
+
# if width > 2048 or height > 2048:
|
104 |
+
# return True
|
105 |
+
|
106 |
+
# else:
|
107 |
+
# print("Image dimensions are within the limit.")
|
108 |
+
# return True
|
109 |
+
|
110 |
+
|
111 |
+
def enhance(self, image):
|
112 |
+
img = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
|
113 |
+
if self.check_image_dimensions(img):
|
114 |
+
cropped_faces, restored_faces, r_img = self.restorer.enhance(
|
115 |
+
img,
|
116 |
+
has_aligned=False,
|
117 |
+
only_center_face=False,
|
118 |
+
paste_back=True)
|
119 |
+
else:
|
120 |
+
r_img = img
|
121 |
+
|
122 |
+
r_img = cv2.cvtColor(r_img, cv2.COLOR_BGR2RGB)
|
123 |
+
|
124 |
+
return r_img
|