Spaces:
Running
on
Zero
Running
on
Zero
Update webui/runner.py
Browse files- webui/runner.py +161 -161
webui/runner.py
CHANGED
@@ -1,161 +1,161 @@
|
|
1 |
-
import torch
|
2 |
-
from PIL import Image
|
3 |
-
from diffusers import DDIMScheduler
|
4 |
-
from accelerate.utils import set_seed
|
5 |
-
from torchvision.transforms.functional import to_pil_image, to_tensor
|
6 |
-
|
7 |
-
from pipeline_sd import ADPipeline
|
8 |
-
from pipeline_sdxl import ADPipeline as ADXLPipeline
|
9 |
-
from utils import Controller
|
10 |
-
|
11 |
-
import os
|
12 |
-
import spaces
|
13 |
-
|
14 |
-
|
15 |
-
class Runner:
|
16 |
-
def __init__(self):
|
17 |
-
self.sd15 = None
|
18 |
-
self.sdxl = None
|
19 |
-
self.loss_fn = torch.nn.L1Loss(reduction="mean")
|
20 |
-
|
21 |
-
def load_pipeline(self, model_path_or_name):
|
22 |
-
|
23 |
-
if 'xl' in model_path_or_name and self.sdxl is None:
|
24 |
-
scheduler = DDIMScheduler.from_pretrained(
|
25 |
-
self.sdxl = ADXLPipeline.from_pretrained(
|
26 |
-
self.sdxl.classifier = self.sdxl.unet
|
27 |
-
elif self.sd15 is None:
|
28 |
-
scheduler = DDIMScheduler.from_pretrained(
|
29 |
-
self.sd15 = ADPipeline.from_pretrained(
|
30 |
-
self.sd15.classifier = self.sd15.unet
|
31 |
-
|
32 |
-
def preprocecss(self, image: Image.Image, height=None, width=None):
|
33 |
-
if width is None or height is None:
|
34 |
-
width, height = image.size
|
35 |
-
new_width = (width // 64) * 64
|
36 |
-
new_height = (height // 64) * 64
|
37 |
-
size = (new_width, new_height)
|
38 |
-
image = image.resize(size, Image.BICUBIC)
|
39 |
-
return to_tensor(image).unsqueeze(0)
|
40 |
-
|
41 |
-
@spaces.GPU
|
42 |
-
def run_style_transfer(self, content_image, style_image, seed, num_steps, lr, content_weight, mixed_precision, model, **kwargs):
|
43 |
-
self.load_pipeline(model)
|
44 |
-
|
45 |
-
content_image = self.preprocecss(content_image)
|
46 |
-
style_image = self.preprocecss(style_image, height=512, width=512)
|
47 |
-
|
48 |
-
height, width = content_image.shape[-2:]
|
49 |
-
set_seed(seed)
|
50 |
-
controller = Controller(self_layers=(10, 16))
|
51 |
-
result = self.sd15.optimize(
|
52 |
-
lr=lr,
|
53 |
-
batch_size=1,
|
54 |
-
iters=1,
|
55 |
-
width=width,
|
56 |
-
height=height,
|
57 |
-
weight=content_weight,
|
58 |
-
controller=controller,
|
59 |
-
style_image=style_image,
|
60 |
-
content_image=content_image,
|
61 |
-
mixed_precision=mixed_precision,
|
62 |
-
num_inference_steps=num_steps,
|
63 |
-
enable_gradient_checkpoint=False,
|
64 |
-
)
|
65 |
-
output_image = to_pil_image(result[0])
|
66 |
-
del result
|
67 |
-
torch.cuda.empty_cache()
|
68 |
-
return [output_image]
|
69 |
-
|
70 |
-
@spaces.GPU
|
71 |
-
def run_style_t2i_generation(self, style_image, prompt, negative_prompt, guidance_scale, height, width, seed, num_steps, iterations, lr, num_images_per_prompt, mixed_precision, is_adain, model):
|
72 |
-
self.load_pipeline(model)
|
73 |
-
|
74 |
-
use_xl = 'xl' in model
|
75 |
-
height, width = (1024, 1024) if 'xl' in model else (512, 512)
|
76 |
-
style_image = self.preprocecss(style_image, height=height, width=width)
|
77 |
-
|
78 |
-
set_seed(seed)
|
79 |
-
self_layers = (64, 70) if use_xl else (10, 16)
|
80 |
-
|
81 |
-
controller = Controller(self_layers=self_layers)
|
82 |
-
|
83 |
-
pipeline = self.sdxl if use_xl else self.sd15
|
84 |
-
images = pipeline.sample(
|
85 |
-
controller=controller,
|
86 |
-
iters=iterations,
|
87 |
-
lr=lr,
|
88 |
-
adain=is_adain,
|
89 |
-
height=height,
|
90 |
-
width=width,
|
91 |
-
mixed_precision=mixed_precision,
|
92 |
-
style_image=style_image,
|
93 |
-
prompt=prompt,
|
94 |
-
negative_prompt=negative_prompt,
|
95 |
-
guidance_scale=guidance_scale,
|
96 |
-
num_inference_steps=num_steps,
|
97 |
-
num_images_per_prompt=num_images_per_prompt,
|
98 |
-
enable_gradient_checkpoint=False
|
99 |
-
)
|
100 |
-
output_images = [to_pil_image(image) for image in images]
|
101 |
-
|
102 |
-
del images
|
103 |
-
torch.cuda.empty_cache()
|
104 |
-
return output_images
|
105 |
-
|
106 |
-
@spaces.GPU
|
107 |
-
def run_texture_synthesis(self, texture_image, height, width, seed, num_steps, iterations, lr, mixed_precision, num_images_per_prompt, synthesis_way,model):
|
108 |
-
self.load_pipeline(model)
|
109 |
-
|
110 |
-
texture_image = self.preprocecss(texture_image, height=512, width=512)
|
111 |
-
|
112 |
-
set_seed(seed)
|
113 |
-
controller = Controller(self_layers=(10, 16))
|
114 |
-
|
115 |
-
if synthesis_way == 'Sampling':
|
116 |
-
results = self.sd15.sample(
|
117 |
-
lr=lr,
|
118 |
-
adain=False,
|
119 |
-
iters=iterations,
|
120 |
-
width=width,
|
121 |
-
height=height,
|
122 |
-
weight=0.,
|
123 |
-
controller=controller,
|
124 |
-
style_image=texture_image,
|
125 |
-
content_image=None,
|
126 |
-
prompt="",
|
127 |
-
negative_prompt="",
|
128 |
-
mixed_precision=mixed_precision,
|
129 |
-
num_inference_steps=num_steps,
|
130 |
-
guidance_scale=1.,
|
131 |
-
num_images_per_prompt=num_images_per_prompt,
|
132 |
-
enable_gradient_checkpoint=False,
|
133 |
-
)
|
134 |
-
elif synthesis_way == 'MultiDiffusion':
|
135 |
-
results = self.sd15.panorama(
|
136 |
-
lr=lr,
|
137 |
-
iters=iterations,
|
138 |
-
width=width,
|
139 |
-
height=height,
|
140 |
-
weight=0.,
|
141 |
-
controller=controller,
|
142 |
-
style_image=texture_image,
|
143 |
-
content_image=None,
|
144 |
-
prompt="",
|
145 |
-
negative_prompt="",
|
146 |
-
stride=8,
|
147 |
-
view_batch_size=8,
|
148 |
-
mixed_precision=mixed_precision,
|
149 |
-
num_inference_steps=num_steps,
|
150 |
-
guidance_scale=1.,
|
151 |
-
num_images_per_prompt=num_images_per_prompt,
|
152 |
-
enable_gradient_checkpoint=False,
|
153 |
-
)
|
154 |
-
else:
|
155 |
-
raise ValueError
|
156 |
-
|
157 |
-
output_images = [to_pil_image(image) for image in results]
|
158 |
-
del results
|
159 |
-
torch.cuda.empty_cache()
|
160 |
-
return output_images
|
161 |
-
|
|
|
1 |
+
import torch
|
2 |
+
from PIL import Image
|
3 |
+
from diffusers import DDIMScheduler
|
4 |
+
from accelerate.utils import set_seed
|
5 |
+
from torchvision.transforms.functional import to_pil_image, to_tensor
|
6 |
+
|
7 |
+
from pipeline_sd import ADPipeline
|
8 |
+
from pipeline_sdxl import ADPipeline as ADXLPipeline
|
9 |
+
from utils import Controller
|
10 |
+
|
11 |
+
import os
|
12 |
+
import spaces
|
13 |
+
|
14 |
+
|
15 |
+
class Runner:
|
16 |
+
def __init__(self):
|
17 |
+
self.sd15 = None
|
18 |
+
self.sdxl = None
|
19 |
+
self.loss_fn = torch.nn.L1Loss(reduction="mean")
|
20 |
+
|
21 |
+
def load_pipeline(self, model_path_or_name):
|
22 |
+
|
23 |
+
if 'xl' in model_path_or_name and self.sdxl is None:
|
24 |
+
scheduler = DDIMScheduler.from_pretrained(model_path_or_name, subfolder="scheduler")
|
25 |
+
self.sdxl = ADXLPipeline.from_pretrained(model_path_or_name, scheduler=scheduler, safety_checker=None)
|
26 |
+
self.sdxl.classifier = self.sdxl.unet
|
27 |
+
elif self.sd15 is None:
|
28 |
+
scheduler = DDIMScheduler.from_pretrained(model_path_or_name, subfolder="scheduler")
|
29 |
+
self.sd15 = ADPipeline.from_pretrained(model_path_or_name, scheduler=scheduler, safety_checker=None)
|
30 |
+
self.sd15.classifier = self.sd15.unet
|
31 |
+
|
32 |
+
def preprocecss(self, image: Image.Image, height=None, width=None):
|
33 |
+
if width is None or height is None:
|
34 |
+
width, height = image.size
|
35 |
+
new_width = (width // 64) * 64
|
36 |
+
new_height = (height // 64) * 64
|
37 |
+
size = (new_width, new_height)
|
38 |
+
image = image.resize(size, Image.BICUBIC)
|
39 |
+
return to_tensor(image).unsqueeze(0)
|
40 |
+
|
41 |
+
@spaces.GPU
|
42 |
+
def run_style_transfer(self, content_image, style_image, seed, num_steps, lr, content_weight, mixed_precision, model, **kwargs):
|
43 |
+
self.load_pipeline(model)
|
44 |
+
|
45 |
+
content_image = self.preprocecss(content_image)
|
46 |
+
style_image = self.preprocecss(style_image, height=512, width=512)
|
47 |
+
|
48 |
+
height, width = content_image.shape[-2:]
|
49 |
+
set_seed(seed)
|
50 |
+
controller = Controller(self_layers=(10, 16))
|
51 |
+
result = self.sd15.optimize(
|
52 |
+
lr=lr,
|
53 |
+
batch_size=1,
|
54 |
+
iters=1,
|
55 |
+
width=width,
|
56 |
+
height=height,
|
57 |
+
weight=content_weight,
|
58 |
+
controller=controller,
|
59 |
+
style_image=style_image,
|
60 |
+
content_image=content_image,
|
61 |
+
mixed_precision=mixed_precision,
|
62 |
+
num_inference_steps=num_steps,
|
63 |
+
enable_gradient_checkpoint=False,
|
64 |
+
)
|
65 |
+
output_image = to_pil_image(result[0])
|
66 |
+
del result
|
67 |
+
torch.cuda.empty_cache()
|
68 |
+
return [output_image]
|
69 |
+
|
70 |
+
@spaces.GPU
|
71 |
+
def run_style_t2i_generation(self, style_image, prompt, negative_prompt, guidance_scale, height, width, seed, num_steps, iterations, lr, num_images_per_prompt, mixed_precision, is_adain, model):
|
72 |
+
self.load_pipeline(model)
|
73 |
+
|
74 |
+
use_xl = 'xl' in model
|
75 |
+
height, width = (1024, 1024) if 'xl' in model else (512, 512)
|
76 |
+
style_image = self.preprocecss(style_image, height=height, width=width)
|
77 |
+
|
78 |
+
set_seed(seed)
|
79 |
+
self_layers = (64, 70) if use_xl else (10, 16)
|
80 |
+
|
81 |
+
controller = Controller(self_layers=self_layers)
|
82 |
+
|
83 |
+
pipeline = self.sdxl if use_xl else self.sd15
|
84 |
+
images = pipeline.sample(
|
85 |
+
controller=controller,
|
86 |
+
iters=iterations,
|
87 |
+
lr=lr,
|
88 |
+
adain=is_adain,
|
89 |
+
height=height,
|
90 |
+
width=width,
|
91 |
+
mixed_precision=mixed_precision,
|
92 |
+
style_image=style_image,
|
93 |
+
prompt=prompt,
|
94 |
+
negative_prompt=negative_prompt,
|
95 |
+
guidance_scale=guidance_scale,
|
96 |
+
num_inference_steps=num_steps,
|
97 |
+
num_images_per_prompt=num_images_per_prompt,
|
98 |
+
enable_gradient_checkpoint=False
|
99 |
+
)
|
100 |
+
output_images = [to_pil_image(image) for image in images]
|
101 |
+
|
102 |
+
del images
|
103 |
+
torch.cuda.empty_cache()
|
104 |
+
return output_images
|
105 |
+
|
106 |
+
@spaces.GPU
|
107 |
+
def run_texture_synthesis(self, texture_image, height, width, seed, num_steps, iterations, lr, mixed_precision, num_images_per_prompt, synthesis_way,model):
|
108 |
+
self.load_pipeline(model)
|
109 |
+
|
110 |
+
texture_image = self.preprocecss(texture_image, height=512, width=512)
|
111 |
+
|
112 |
+
set_seed(seed)
|
113 |
+
controller = Controller(self_layers=(10, 16))
|
114 |
+
|
115 |
+
if synthesis_way == 'Sampling':
|
116 |
+
results = self.sd15.sample(
|
117 |
+
lr=lr,
|
118 |
+
adain=False,
|
119 |
+
iters=iterations,
|
120 |
+
width=width,
|
121 |
+
height=height,
|
122 |
+
weight=0.,
|
123 |
+
controller=controller,
|
124 |
+
style_image=texture_image,
|
125 |
+
content_image=None,
|
126 |
+
prompt="",
|
127 |
+
negative_prompt="",
|
128 |
+
mixed_precision=mixed_precision,
|
129 |
+
num_inference_steps=num_steps,
|
130 |
+
guidance_scale=1.,
|
131 |
+
num_images_per_prompt=num_images_per_prompt,
|
132 |
+
enable_gradient_checkpoint=False,
|
133 |
+
)
|
134 |
+
elif synthesis_way == 'MultiDiffusion':
|
135 |
+
results = self.sd15.panorama(
|
136 |
+
lr=lr,
|
137 |
+
iters=iterations,
|
138 |
+
width=width,
|
139 |
+
height=height,
|
140 |
+
weight=0.,
|
141 |
+
controller=controller,
|
142 |
+
style_image=texture_image,
|
143 |
+
content_image=None,
|
144 |
+
prompt="",
|
145 |
+
negative_prompt="",
|
146 |
+
stride=8,
|
147 |
+
view_batch_size=8,
|
148 |
+
mixed_precision=mixed_precision,
|
149 |
+
num_inference_steps=num_steps,
|
150 |
+
guidance_scale=1.,
|
151 |
+
num_images_per_prompt=num_images_per_prompt,
|
152 |
+
enable_gradient_checkpoint=False,
|
153 |
+
)
|
154 |
+
else:
|
155 |
+
raise ValueError
|
156 |
+
|
157 |
+
output_images = [to_pil_image(image) for image in results]
|
158 |
+
del results
|
159 |
+
torch.cuda.empty_cache()
|
160 |
+
return output_images
|
161 |
+
|