Spaces:
Running
on
Zero
Running
on
Zero
ohayonguy
commited on
Commit
•
b51eadf
1
Parent(s):
d18dfca
changed output format to png
Browse files
app.py
CHANGED
@@ -4,9 +4,9 @@ import os
|
|
4 |
|
5 |
import matplotlib.pyplot as plt
|
6 |
|
7 |
-
if os.getenv(
|
8 |
-
os.environ[
|
9 |
-
os.environ[
|
10 |
|
11 |
import spaces
|
12 |
import cv2
|
@@ -24,11 +24,13 @@ from lightning_models.mmse_rectified_flow import MMSERectifiedFlow
|
|
24 |
MAX_SEED = 10000
|
25 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
26 |
|
27 |
-
os.makedirs(
|
28 |
-
realesr_model_path =
|
29 |
if not os.path.exists(realesr_model_path):
|
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')
|
@@ -39,18 +41,15 @@ if not os.path.exists(realesr_model_path):
|
|
39 |
|
40 |
def set_realesrgan():
|
41 |
use_half = False
|
42 |
-
if torch.cuda.is_available():
|
43 |
-
no_half_gpu_list = [
|
44 |
-
if not True in [
|
|
|
|
|
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,
|
@@ -59,20 +58,28 @@ def set_realesrgan():
|
|
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(
|
|
|
|
|
|
|
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(
|
|
|
|
|
71 |
dt = (1.0 / num_flow_steps) * (1.0 - pmrf_model.hparams.eps)
|
72 |
x_t_next = source_dist_samples.clone()
|
73 |
t_one = torch.ones(x.shape[0], device=device)
|
74 |
for i in tqdm(range(num_flow_steps)):
|
75 |
-
num_t = (i / num_flow_steps) * (
|
|
|
|
|
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 |
|
@@ -87,6 +94,7 @@ def resize(img, size):
|
|
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):
|
@@ -102,20 +110,26 @@ def enhance_face(img, face_helper, has_aligned, num_flow_steps, scale=2):
|
|
102 |
if len(face_helper.cropped_faces) == 0:
|
103 |
raise gr.Error("Could not identify any face in the image.")
|
104 |
if has_aligned and len(face_helper.cropped_faces) > 1:
|
105 |
-
raise gr.Error(
|
|
|
|
|
106 |
|
107 |
# face restoration
|
108 |
for i, cropped_face in tqdm(enumerate(face_helper.cropped_faces)):
|
109 |
-
cropped_face_t = img2tensor(cropped_face / 255
|
110 |
cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
|
111 |
|
112 |
-
output = generate_reconstructions(
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
|
|
|
|
|
|
|
|
119 |
restored_face = restored_face.astype("uint8")
|
120 |
face_helper.add_restored_face(restored_face)
|
121 |
|
@@ -126,8 +140,6 @@ def enhance_face(img, face_helper, has_aligned, num_flow_steps, scale=2):
|
|
126 |
face_helper.get_inverse_affine(None)
|
127 |
# paste each restored face to the input image
|
128 |
restored_img = face_helper.paste_faces_to_input_image(upsample_img=bg_img)
|
129 |
-
print(bg_img.shape, img.shape,restored_img.shape)
|
130 |
-
|
131 |
return face_helper.cropped_faces, face_helper.restored_faces, restored_img
|
132 |
else:
|
133 |
return face_helper.cropped_faces, face_helper.restored_faces, None
|
@@ -135,8 +147,15 @@ def enhance_face(img, face_helper, has_aligned, num_flow_steps, scale=2):
|
|
135 |
|
136 |
@torch.inference_mode()
|
137 |
@spaces.GPU()
|
138 |
-
def inference(
|
139 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
140 |
if img is None:
|
141 |
raise gr.Error("Please upload an image before submitting.")
|
142 |
if randomize_seed:
|
@@ -145,24 +164,23 @@ def inference(seed, randomize_seed, img, aligned, scale, num_flow_steps,
|
|
145 |
img = cv2.imread(img, cv2.IMREAD_COLOR)
|
146 |
h, w = img.shape[0:2]
|
147 |
if h > 4500 or w > 4500:
|
148 |
-
raise gr.Error(
|
149 |
|
150 |
face_helper = FaceRestoreHelper(
|
151 |
scale,
|
152 |
face_size=512,
|
153 |
crop_ratio=(1, 1),
|
154 |
-
det_model=
|
155 |
-
save_ext=
|
156 |
use_parse=True,
|
157 |
device=device,
|
158 |
-
model_rootpath=None
|
|
|
159 |
|
160 |
has_aligned = aligned
|
161 |
-
cropped_face, restored_faces, restored_img = enhance_face(
|
162 |
-
|
163 |
-
|
164 |
-
num_flow_steps=num_flow_steps,
|
165 |
-
scale=scale)
|
166 |
if has_aligned:
|
167 |
output = restored_faces[0]
|
168 |
else:
|
@@ -231,7 +249,6 @@ css = """
|
|
231 |
}
|
232 |
"""
|
233 |
|
234 |
-
|
235 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
236 |
gr.HTML(intro)
|
237 |
gr.Markdown(markdown_top)
|
@@ -255,15 +272,13 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
|
255 |
value=1,
|
256 |
)
|
257 |
seed = gr.Slider(
|
258 |
-
label="Seed",
|
259 |
-
minimum=0,
|
260 |
-
maximum=MAX_SEED,
|
261 |
-
step=1,
|
262 |
-
value=42,
|
263 |
)
|
264 |
|
265 |
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
266 |
-
aligned = gr.Checkbox(
|
|
|
|
|
267 |
|
268 |
with gr.Row():
|
269 |
with gr.Column(scale=1):
|
@@ -272,9 +287,9 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
|
272 |
clear_button = gr.ClearButton(value="Clear")
|
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 |
clear_button.add(input_im)
|
280 |
clear_button.add(result)
|
|
|
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
|
|
|
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)
|
28 |
+
realesr_model_path = "pretrained_models/RealESRGAN_x4plus.pth"
|
29 |
if not os.path.exists(realesr_model_path):
|
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 |
+
|
34 |
|
35 |
# # background enhancer with RealESRGAN
|
36 |
# model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu')
|
|
|
41 |
|
42 |
def set_realesrgan():
|
43 |
use_half = False
|
44 |
+
if torch.cuda.is_available(): # set False in CPU/MPS mode
|
45 |
+
no_half_gpu_list = ["1650", "1660"] # set False for GPUs that don't support f16
|
46 |
+
if not True in [
|
47 |
+
gpu in torch.cuda.get_device_name(0) for gpu in no_half_gpu_list
|
48 |
+
]:
|
49 |
use_half = True
|
50 |
|
51 |
model = RRDBNet(
|
52 |
+
num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2,
|
|
|
|
|
|
|
|
|
|
|
53 |
)
|
54 |
upsampler = RealESRGANer(
|
55 |
scale=2,
|
|
|
58 |
tile=400,
|
59 |
tile_pad=40,
|
60 |
pre_pad=0,
|
61 |
+
half=use_half,
|
62 |
)
|
63 |
return upsampler
|
64 |
|
65 |
+
|
66 |
upsampler = set_realesrgan()
|
67 |
+
pmrf = MMSERectifiedFlow.from_pretrained(
|
68 |
+
"ohayonguy/PMRF_blind_face_image_restoration"
|
69 |
+
).to(device=device)
|
70 |
+
|
71 |
|
72 |
def generate_reconstructions(pmrf_model, x, y, non_noisy_z0, num_flow_steps, device):
|
73 |
+
source_dist_samples = pmrf_model.create_source_distribution_samples(
|
74 |
+
x, y, non_noisy_z0
|
75 |
+
)
|
76 |
dt = (1.0 / num_flow_steps) * (1.0 - pmrf_model.hparams.eps)
|
77 |
x_t_next = source_dist_samples.clone()
|
78 |
t_one = torch.ones(x.shape[0], device=device)
|
79 |
for i in tqdm(range(num_flow_steps)):
|
80 |
+
num_t = (i / num_flow_steps) * (
|
81 |
+
1.0 - pmrf_model.hparams.eps
|
82 |
+
) + pmrf_model.hparams.eps
|
83 |
v_t_next = pmrf_model(x_t=x_t_next, t=t_one * num_t, y=y).to(x_t_next.dtype)
|
84 |
x_t_next = x_t_next.clone() + v_t_next * dt
|
85 |
|
|
|
94 |
interp = cv2.INTER_AREA if scale < 1 else cv2.INTER_LINEAR
|
95 |
return cv2.resize(img, (w, h), interpolation=interp)
|
96 |
|
97 |
+
|
98 |
@torch.inference_mode()
|
99 |
@spaces.GPU()
|
100 |
def enhance_face(img, face_helper, has_aligned, num_flow_steps, scale=2):
|
|
|
110 |
if len(face_helper.cropped_faces) == 0:
|
111 |
raise gr.Error("Could not identify any face in the image.")
|
112 |
if has_aligned and len(face_helper.cropped_faces) > 1:
|
113 |
+
raise gr.Error(
|
114 |
+
"You marked that the input image is aligned, but multiple faces were detected."
|
115 |
+
)
|
116 |
|
117 |
# face restoration
|
118 |
for i, cropped_face in tqdm(enumerate(face_helper.cropped_faces)):
|
119 |
+
cropped_face_t = img2tensor(cropped_face / 255.0, bgr2rgb=True, float32=True)
|
120 |
cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
|
121 |
|
122 |
+
output = generate_reconstructions(
|
123 |
+
pmrf,
|
124 |
+
torch.zeros_like(cropped_face_t),
|
125 |
+
cropped_face_t,
|
126 |
+
None,
|
127 |
+
num_flow_steps,
|
128 |
+
device,
|
129 |
+
)
|
130 |
+
restored_face = tensor2img(
|
131 |
+
output.to(torch.float32).squeeze(0), rgb2bgr=True, min_max=(0, 1)
|
132 |
+
)
|
133 |
restored_face = restored_face.astype("uint8")
|
134 |
face_helper.add_restored_face(restored_face)
|
135 |
|
|
|
140 |
face_helper.get_inverse_affine(None)
|
141 |
# paste each restored face to the input image
|
142 |
restored_img = face_helper.paste_faces_to_input_image(upsample_img=bg_img)
|
|
|
|
|
143 |
return face_helper.cropped_faces, face_helper.restored_faces, restored_img
|
144 |
else:
|
145 |
return face_helper.cropped_faces, face_helper.restored_faces, None
|
|
|
147 |
|
148 |
@torch.inference_mode()
|
149 |
@spaces.GPU()
|
150 |
+
def inference(
|
151 |
+
seed,
|
152 |
+
randomize_seed,
|
153 |
+
img,
|
154 |
+
aligned,
|
155 |
+
scale,
|
156 |
+
num_flow_steps,
|
157 |
+
progress=gr.Progress(track_tqdm=True),
|
158 |
+
):
|
159 |
if img is None:
|
160 |
raise gr.Error("Please upload an image before submitting.")
|
161 |
if randomize_seed:
|
|
|
164 |
img = cv2.imread(img, cv2.IMREAD_COLOR)
|
165 |
h, w = img.shape[0:2]
|
166 |
if h > 4500 or w > 4500:
|
167 |
+
raise gr.Error("Image size too large.")
|
168 |
|
169 |
face_helper = FaceRestoreHelper(
|
170 |
scale,
|
171 |
face_size=512,
|
172 |
crop_ratio=(1, 1),
|
173 |
+
det_model="retinaface_resnet50",
|
174 |
+
save_ext="png",
|
175 |
use_parse=True,
|
176 |
device=device,
|
177 |
+
model_rootpath=None,
|
178 |
+
)
|
179 |
|
180 |
has_aligned = aligned
|
181 |
+
cropped_face, restored_faces, restored_img = enhance_face(
|
182 |
+
img, face_helper, has_aligned, num_flow_steps=num_flow_steps, scale=scale
|
183 |
+
)
|
|
|
|
|
184 |
if has_aligned:
|
185 |
output = restored_faces[0]
|
186 |
else:
|
|
|
249 |
}
|
250 |
"""
|
251 |
|
|
|
252 |
with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
|
253 |
gr.HTML(intro)
|
254 |
gr.Markdown(markdown_top)
|
|
|
272 |
value=1,
|
273 |
)
|
274 |
seed = gr.Slider(
|
275 |
+
label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42,
|
|
|
|
|
|
|
|
|
276 |
)
|
277 |
|
278 |
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
279 |
+
aligned = gr.Checkbox(
|
280 |
+
label="The input is an aligned face image.", value=False
|
281 |
+
)
|
282 |
|
283 |
with gr.Row():
|
284 |
with gr.Column(scale=1):
|
|
|
287 |
clear_button = gr.ClearButton(value="Clear")
|
288 |
|
289 |
with gr.Row():
|
290 |
+
result = gr.Image(label="Output", type="numpy", show_label=True, format="png")
|
291 |
with gr.Row():
|
292 |
+
gallery = gr.Gallery(label="Restored faces gallery", type="numpy", show_label=True, format="png")
|
293 |
|
294 |
clear_button.add(input_im)
|
295 |
clear_button.add(result)
|