ohayonguy
commited on
Commit
•
a00800e
1
Parent(s):
94bce76
improved interface and added examples
Browse files- app.py +110 -67
- examples/00000055.png +0 -0
- examples/00000085.png +0 -0
- examples/00000113.png +0 -0
- examples/00000137.png +0 -0
- examples/01.png +0 -0
- examples/03.jpg +0 -0
app.py
CHANGED
@@ -1,25 +1,27 @@
|
|
|
|
|
|
1 |
import os
|
2 |
|
|
|
|
|
3 |
if os.getenv('SPACES_ZERO_GPU') == "true":
|
4 |
os.environ['SPACES_ZERO_GPU'] = "1"
|
5 |
os.environ['K_DIFFUSION_USE_COMPILE'] = "0"
|
|
|
6 |
import spaces
|
7 |
import cv2
|
8 |
from tqdm import tqdm
|
9 |
import gradio as gr
|
10 |
import random
|
11 |
import torch
|
12 |
-
from basicsr.archs.
|
13 |
from basicsr.utils import img2tensor, tensor2img
|
14 |
-
from gradio_imageslider import ImageSlider
|
15 |
from facexlib.utils.face_restoration_helper import FaceRestoreHelper
|
16 |
from realesrgan.utils import RealESRGANer
|
17 |
|
18 |
from lightning_models.mmse_rectified_flow import MMSERectifiedFlow
|
19 |
|
20 |
-
|
21 |
-
|
22 |
-
MAX_SEED = 1000000
|
23 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
24 |
|
25 |
os.makedirs('pretrained_models', exist_ok=True)
|
@@ -28,25 +30,42 @@ if not os.path.exists(realesr_model_path):
|
|
28 |
os.system(
|
29 |
"wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth -O pretrained_models/RealESRGAN_x4plus.pth")
|
30 |
|
31 |
-
# background enhancer with RealESRGAN
|
32 |
-
model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu')
|
33 |
-
half = True if torch.cuda.is_available() else False
|
34 |
-
upsampler = RealESRGANer(scale=4, model_path=realesr_model_path, model=model, tile=400, tile_pad=10, pre_pad=0,
|
35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
|
|
37 |
pmrf = MMSERectifiedFlow.from_pretrained('ohayonguy/PMRF_blind_face_image_restoration').to(device=device)
|
38 |
|
39 |
-
face_helper_dummy = FaceRestoreHelper(
|
40 |
-
1,
|
41 |
-
face_size=512,
|
42 |
-
crop_ratio=(1, 1),
|
43 |
-
det_model='retinaface_resnet50',
|
44 |
-
save_ext='png',
|
45 |
-
use_parse=True,
|
46 |
-
device=device,
|
47 |
-
model_rootpath=None)
|
48 |
-
|
49 |
-
|
50 |
def generate_reconstructions(pmrf_model, x, y, non_noisy_z0, num_flow_steps, device):
|
51 |
source_dist_samples = pmrf_model.create_source_distribution_samples(x, y, non_noisy_z0)
|
52 |
dt = (1.0 / num_flow_steps) * (1.0 - pmrf_model.hparams.eps)
|
@@ -57,58 +76,61 @@ def generate_reconstructions(pmrf_model, x, y, non_noisy_z0, num_flow_steps, dev
|
|
57 |
v_t_next = pmrf_model(x_t=x_t_next, t=t_one * num_t, y=y).to(x_t_next.dtype)
|
58 |
x_t_next = x_t_next.clone() + v_t_next * dt
|
59 |
|
60 |
-
return x_t_next.clip(0, 1)
|
61 |
|
62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
@torch.inference_mode()
|
64 |
@spaces.GPU()
|
65 |
-
def enhance_face(img, face_helper, has_aligned, num_flow_steps,
|
66 |
face_helper.clean_all()
|
67 |
-
if has_aligned: #
|
68 |
img = cv2.resize(img, (512, 512), interpolation=cv2.INTER_LINEAR)
|
69 |
face_helper.cropped_faces = [img]
|
70 |
else:
|
71 |
face_helper.read_image(img)
|
72 |
-
face_helper.
|
73 |
-
|
74 |
-
# TODO: even with eye_dist_threshold, it will still introduce wrong detections and restorations.
|
75 |
-
# align and warp each face
|
76 |
face_helper.align_warp_face()
|
77 |
if len(face_helper.cropped_faces) == 0:
|
78 |
raise gr.Error("Could not identify any face in the image.")
|
79 |
if len(face_helper.cropped_faces) > 1:
|
80 |
-
gr.Info(f"Identified {len(face_helper.cropped_faces)}
|
|
|
81 |
else:
|
82 |
gr.Info(f"Identified one face in the image.")
|
83 |
|
84 |
# face restoration
|
85 |
for i, cropped_face in tqdm(enumerate(face_helper.cropped_faces)):
|
86 |
-
# prepare data
|
87 |
-
h, w = cropped_face.shape[0], cropped_face.shape[1]
|
88 |
-
cropped_face = cv2.resize(cropped_face, (512, 512), interpolation=cv2.INTER_LINEAR)
|
89 |
-
# face_helper.cropped_faces[i] = cropped_face
|
90 |
cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
|
91 |
cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
|
92 |
|
93 |
-
|
94 |
-
|
|
|
|
|
|
|
|
|
95 |
restored_face = tensor2img(output.to(torch.float32).squeeze(0), rgb2bgr=True, min_max=(0, 1))
|
96 |
-
restored_face =
|
97 |
-
|
98 |
-
restored_face = restored_face.astype('uint8')
|
99 |
face_helper.add_restored_face(restored_face)
|
100 |
|
101 |
-
if not has_aligned
|
102 |
# upsample the background
|
103 |
-
|
104 |
-
|
105 |
-
bg_img = upsampler.enhance(img, outscale=scale)[0]
|
106 |
-
else:
|
107 |
-
bg_img = None
|
108 |
-
|
109 |
face_helper.get_inverse_affine(None)
|
110 |
# paste each restored face to the input image
|
111 |
restored_img = face_helper.paste_faces_to_input_image(upsample_img=bg_img)
|
|
|
|
|
112 |
return face_helper.cropped_faces, face_helper.restored_faces, restored_img
|
113 |
else:
|
114 |
return face_helper.cropped_faces, face_helper.restored_faces, None
|
@@ -123,12 +145,7 @@ def inference(seed, randomize_seed, img, aligned, scale, num_flow_steps,
|
|
123 |
if randomize_seed:
|
124 |
seed = random.randint(0, MAX_SEED)
|
125 |
torch.manual_seed(seed)
|
126 |
-
|
127 |
-
scale = 4 # avoid too large scale value
|
128 |
-
img = cv2.imread(img, cv2.IMREAD_UNCHANGED)
|
129 |
-
if len(img.shape) == 2: # for gray inputs
|
130 |
-
img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
|
131 |
-
|
132 |
h, w = img.shape[0:2]
|
133 |
if h > 4500 or w > 4500:
|
134 |
raise gr.Error('Image size too large.')
|
@@ -143,22 +160,22 @@ def inference(seed, randomize_seed, img, aligned, scale, num_flow_steps,
|
|
143 |
device=device,
|
144 |
model_rootpath=None)
|
145 |
|
146 |
-
has_aligned =
|
147 |
-
cropped_face,
|
148 |
-
|
|
|
|
|
149 |
scale=scale)
|
150 |
if has_aligned:
|
151 |
-
output =
|
152 |
-
# input = cropped_face[0].astype('uint8')
|
153 |
else:
|
154 |
output = restored_img
|
155 |
-
# input = img
|
156 |
|
157 |
output = cv2.cvtColor(output, cv2.COLOR_BGR2RGB)
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
return output
|
162 |
|
163 |
|
164 |
intro = """
|
@@ -177,8 +194,9 @@ Please refer to our project's page for more details: https://pmrf-ml.github.io/.
|
|
177 |
|
178 |
*Notes* :
|
179 |
|
180 |
-
1. Our model is designed to restore aligned face images, where there is *only one* face in the image, and the face is centered. Here, however, we incorporate mechanisms that allow restoring the quality of *any* image that contains *any* number of faces. Thus, the resulting quality of such general images is not guaranteed.
|
181 |
-
2.
|
|
|
182 |
|
183 |
---
|
184 |
"""
|
@@ -216,6 +234,7 @@ css = """
|
|
216 |
}
|
217 |
"""
|
218 |
|
|
|
219 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
220 |
gr.HTML(intro)
|
221 |
gr.Markdown(markdown_top)
|
@@ -232,7 +251,7 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
|
232 |
value=25,
|
233 |
)
|
234 |
upscale_factor = gr.Slider(
|
235 |
-
label="Scale factor
|
236 |
minimum=1,
|
237 |
maximum=4,
|
238 |
step=0.1,
|
@@ -247,13 +266,37 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
|
247 |
)
|
248 |
|
249 |
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
250 |
-
aligned = gr.Checkbox(label="The input is an aligned face image", value=False)
|
251 |
|
252 |
with gr.Row():
|
253 |
run_button = gr.Button(value="Submit", variant="primary")
|
254 |
|
255 |
with gr.Row():
|
256 |
result = gr.Image(label="Output", type="numpy", show_label=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
|
258 |
gr.Markdown(article)
|
259 |
gr.on(
|
@@ -267,8 +310,8 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
|
267 |
upscale_factor,
|
268 |
num_inference_steps,
|
269 |
],
|
270 |
-
outputs=result,
|
271 |
-
show_api=False,
|
272 |
# show_progress="minimal",
|
273 |
)
|
274 |
|
|
|
1 |
+
# Some of the implementations below are adopted from
|
2 |
+
# https://huggingface.co/spaces/sczhou/CodeFormer and https://huggingface.co/spaces/wzhouxiff/RestoreFormerPlusPlus
|
3 |
import os
|
4 |
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
|
7 |
if os.getenv('SPACES_ZERO_GPU') == "true":
|
8 |
os.environ['SPACES_ZERO_GPU'] = "1"
|
9 |
os.environ['K_DIFFUSION_USE_COMPILE'] = "0"
|
10 |
+
|
11 |
import spaces
|
12 |
import cv2
|
13 |
from tqdm import tqdm
|
14 |
import gradio as gr
|
15 |
import random
|
16 |
import torch
|
17 |
+
from basicsr.archs.rrdbnet_arch import RRDBNet
|
18 |
from basicsr.utils import img2tensor, tensor2img
|
|
|
19 |
from facexlib.utils.face_restoration_helper import FaceRestoreHelper
|
20 |
from realesrgan.utils import RealESRGANer
|
21 |
|
22 |
from lightning_models.mmse_rectified_flow import MMSERectifiedFlow
|
23 |
|
24 |
+
MAX_SEED = 10000
|
|
|
|
|
25 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
26 |
|
27 |
os.makedirs('pretrained_models', exist_ok=True)
|
|
|
30 |
os.system(
|
31 |
"wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth -O pretrained_models/RealESRGAN_x4plus.pth")
|
32 |
|
33 |
+
# # background enhancer with RealESRGAN
|
34 |
+
# model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu')
|
35 |
+
# half = True if torch.cuda.is_available() else False
|
36 |
+
# upsampler = RealESRGANer(scale=4, model_path=realesr_model_path, model=model, tile=400, tile_pad=10, pre_pad=0,
|
37 |
+
# half=half)
|
38 |
+
|
39 |
+
|
40 |
+
def set_realesrgan():
|
41 |
+
use_half = False
|
42 |
+
if torch.cuda.is_available(): # set False in CPU/MPS mode
|
43 |
+
no_half_gpu_list = ['1650', '1660'] # set False for GPUs that don't support f16
|
44 |
+
if not True in [gpu in torch.cuda.get_device_name(0) for gpu in no_half_gpu_list]:
|
45 |
+
use_half = True
|
46 |
+
|
47 |
+
model = RRDBNet(
|
48 |
+
num_in_ch=3,
|
49 |
+
num_out_ch=3,
|
50 |
+
num_feat=64,
|
51 |
+
num_block=23,
|
52 |
+
num_grow_ch=32,
|
53 |
+
scale=2,
|
54 |
+
)
|
55 |
+
upsampler = RealESRGANer(
|
56 |
+
scale=2,
|
57 |
+
model_path="https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/RealESRGAN_x2plus.pth",
|
58 |
+
model=model,
|
59 |
+
tile=400,
|
60 |
+
tile_pad=40,
|
61 |
+
pre_pad=0,
|
62 |
+
half=use_half
|
63 |
+
)
|
64 |
+
return upsampler
|
65 |
|
66 |
+
upsampler = set_realesrgan()
|
67 |
pmrf = MMSERectifiedFlow.from_pretrained('ohayonguy/PMRF_blind_face_image_restoration').to(device=device)
|
68 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
def generate_reconstructions(pmrf_model, x, y, non_noisy_z0, num_flow_steps, device):
|
70 |
source_dist_samples = pmrf_model.create_source_distribution_samples(x, y, non_noisy_z0)
|
71 |
dt = (1.0 / num_flow_steps) * (1.0 - pmrf_model.hparams.eps)
|
|
|
76 |
v_t_next = pmrf_model(x_t=x_t_next, t=t_one * num_t, y=y).to(x_t_next.dtype)
|
77 |
x_t_next = x_t_next.clone() + v_t_next * dt
|
78 |
|
79 |
+
return x_t_next.clip(0, 1)
|
80 |
|
81 |
|
82 |
+
def resize(img, size):
|
83 |
+
# From https://github.com/sczhou/CodeFormer/blob/master/facelib/utils/face_restoration_helper.py
|
84 |
+
h, w = img.shape[0:2]
|
85 |
+
scale = size / min(h, w)
|
86 |
+
h, w = int(h * scale), int(w * scale)
|
87 |
+
interp = cv2.INTER_AREA if scale < 1 else cv2.INTER_LINEAR
|
88 |
+
return cv2.resize(img, (w, h), interpolation=interp)
|
89 |
+
|
90 |
@torch.inference_mode()
|
91 |
@spaces.GPU()
|
92 |
+
def enhance_face(img, face_helper, has_aligned, num_flow_steps, scale=2):
|
93 |
face_helper.clean_all()
|
94 |
+
if has_aligned: # The inputs are already aligned
|
95 |
img = cv2.resize(img, (512, 512), interpolation=cv2.INTER_LINEAR)
|
96 |
face_helper.cropped_faces = [img]
|
97 |
else:
|
98 |
face_helper.read_image(img)
|
99 |
+
face_helper.input_img = resize(face_helper.input_img, 640)
|
100 |
+
face_helper.get_face_landmarks_5(only_center_face=False, eye_dist_threshold=5)
|
|
|
|
|
101 |
face_helper.align_warp_face()
|
102 |
if len(face_helper.cropped_faces) == 0:
|
103 |
raise gr.Error("Could not identify any face in the image.")
|
104 |
if len(face_helper.cropped_faces) > 1:
|
105 |
+
gr.Info(f"Identified {len(face_helper.cropped_faces)} "
|
106 |
+
f"faces in the image. The algorithm will enhance the quality of each face.")
|
107 |
else:
|
108 |
gr.Info(f"Identified one face in the image.")
|
109 |
|
110 |
# face restoration
|
111 |
for i, cropped_face in tqdm(enumerate(face_helper.cropped_faces)):
|
|
|
|
|
|
|
|
|
112 |
cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
|
113 |
cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
|
114 |
|
115 |
+
output = generate_reconstructions(pmrf,
|
116 |
+
torch.zeros_like(cropped_face_t),
|
117 |
+
cropped_face_t,
|
118 |
+
None,
|
119 |
+
num_flow_steps,
|
120 |
+
device)
|
121 |
restored_face = tensor2img(output.to(torch.float32).squeeze(0), rgb2bgr=True, min_max=(0, 1))
|
122 |
+
restored_face = restored_face.astype("uint8")
|
|
|
|
|
123 |
face_helper.add_restored_face(restored_face)
|
124 |
|
125 |
+
if not has_aligned:
|
126 |
# upsample the background
|
127 |
+
# Now only support RealESRGAN for upsampling background
|
128 |
+
bg_img = upsampler.enhance(img, outscale=scale)[0]
|
|
|
|
|
|
|
|
|
129 |
face_helper.get_inverse_affine(None)
|
130 |
# paste each restored face to the input image
|
131 |
restored_img = face_helper.paste_faces_to_input_image(upsample_img=bg_img)
|
132 |
+
print(bg_img.shape, img.shape,restored_img.shape)
|
133 |
+
|
134 |
return face_helper.cropped_faces, face_helper.restored_faces, restored_img
|
135 |
else:
|
136 |
return face_helper.cropped_faces, face_helper.restored_faces, None
|
|
|
145 |
if randomize_seed:
|
146 |
seed = random.randint(0, MAX_SEED)
|
147 |
torch.manual_seed(seed)
|
148 |
+
img = cv2.imread(img, cv2.IMREAD_COLOR)
|
|
|
|
|
|
|
|
|
|
|
149 |
h, w = img.shape[0:2]
|
150 |
if h > 4500 or w > 4500:
|
151 |
raise gr.Error('Image size too large.')
|
|
|
160 |
device=device,
|
161 |
model_rootpath=None)
|
162 |
|
163 |
+
has_aligned = aligned
|
164 |
+
cropped_face, restored_faces, restored_img = enhance_face(img,
|
165 |
+
face_helper,
|
166 |
+
has_aligned,
|
167 |
+
num_flow_steps=num_flow_steps,
|
168 |
scale=scale)
|
169 |
if has_aligned:
|
170 |
+
output = restored_faces[0]
|
|
|
171 |
else:
|
172 |
output = restored_img
|
|
|
173 |
|
174 |
output = cv2.cvtColor(output, cv2.COLOR_BGR2RGB)
|
175 |
+
for i, restored_face in enumerate(restored_faces):
|
176 |
+
restored_faces[i] = cv2.cvtColor(restored_face, cv2.COLOR_BGR2RGB)
|
177 |
+
torch.cuda.empty_cache()
|
178 |
+
return output, restored_faces
|
179 |
|
180 |
|
181 |
intro = """
|
|
|
194 |
|
195 |
*Notes* :
|
196 |
|
197 |
+
1. Our model is designed to restore aligned face images, where there is *only one* face in the image, and the face is centered and aligned. Here, however, we incorporate mechanisms that allow restoring the quality of *any* image that contains *any* number of faces. Thus, the resulting quality of such general images is not guaranteed.
|
198 |
+
2. If the faces in your image are not aligned, make sure that the checkbox "The input is an aligned face image" in *not* marked.
|
199 |
+
3. Too large images may result in out-of-memory error.
|
200 |
|
201 |
---
|
202 |
"""
|
|
|
234 |
}
|
235 |
"""
|
236 |
|
237 |
+
|
238 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
239 |
gr.HTML(intro)
|
240 |
gr.Markdown(markdown_top)
|
|
|
251 |
value=25,
|
252 |
)
|
253 |
upscale_factor = gr.Slider(
|
254 |
+
label="Scale factor. Applicable only to non-aligned face images. This will upscale the entire image.",
|
255 |
minimum=1,
|
256 |
maximum=4,
|
257 |
step=0.1,
|
|
|
266 |
)
|
267 |
|
268 |
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
269 |
+
aligned = gr.Checkbox(label="The input is an aligned face image.", value=False)
|
270 |
|
271 |
with gr.Row():
|
272 |
run_button = gr.Button(value="Submit", variant="primary")
|
273 |
|
274 |
with gr.Row():
|
275 |
result = gr.Image(label="Output", type="numpy", show_label=True)
|
276 |
+
with gr.Row():
|
277 |
+
gallery = gr.Gallery(label="Restored faces gallery", type="numpy", show_label=True)
|
278 |
+
|
279 |
+
examples = gr.Examples(
|
280 |
+
examples=[
|
281 |
+
[42, False, "examples/01.png", False, 1, 25],
|
282 |
+
[42, False, "examples/03.jpg", False, 2, 25],
|
283 |
+
[42, False, "examples/00000055.png", True, 1, 25],
|
284 |
+
[42, False, "examples/00000085.png", True, 1, 25],
|
285 |
+
[42, False, "examples/00000113.png", True, 1, 25],
|
286 |
+
[42, False, "examples/00000137.png", True, 1, 25],
|
287 |
+
],
|
288 |
+
fn=inference,
|
289 |
+
inputs=[
|
290 |
+
seed,
|
291 |
+
randomize_seed,
|
292 |
+
input_im,
|
293 |
+
aligned,
|
294 |
+
upscale_factor,
|
295 |
+
num_inference_steps,
|
296 |
+
],
|
297 |
+
outputs=[result, gallery],
|
298 |
+
cache_examples="lazy",
|
299 |
+
)
|
300 |
|
301 |
gr.Markdown(article)
|
302 |
gr.on(
|
|
|
310 |
upscale_factor,
|
311 |
num_inference_steps,
|
312 |
],
|
313 |
+
outputs=[result, gallery],
|
314 |
+
# show_api=False,
|
315 |
# show_progress="minimal",
|
316 |
)
|
317 |
|
examples/00000055.png
ADDED
examples/00000085.png
ADDED
examples/00000113.png
ADDED
examples/00000137.png
ADDED
examples/01.png
ADDED
examples/03.jpg
ADDED