Spaces:
Running
on
Zero
Running
on
Zero
Update hf_demo.py
Browse files- hf_demo.py +149 -149
hf_demo.py
CHANGED
@@ -1,150 +1,150 @@
|
|
1 |
-
# Authors: Hui Ren (rhfeiyang.github.io)
|
2 |
-
import
|
3 |
-
import
|
4 |
-
import gradio as gr
|
5 |
-
from diffusers import DiffusionPipeline
|
6 |
-
import matplotlib.pyplot as plt
|
7 |
-
import torch
|
8 |
-
from PIL import Image
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
13 |
-
dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float16
|
14 |
-
pipe = DiffusionPipeline.from_pretrained("rhfeiyang/art-free-diffusion-v1",
|
15 |
-
dtype=dtype).to(device)
|
16 |
-
|
17 |
-
from inference import get_lora_network, inference, get_validation_dataloader
|
18 |
-
lora_map = {
|
19 |
-
"None": "None",
|
20 |
-
"Andre Derain": "andre-derain_subset1",
|
21 |
-
"Vincent van Gogh": "van_gogh_subset1",
|
22 |
-
"Andy Warhol": "andy_subset1",
|
23 |
-
"Walter Battiss": "walter-battiss_subset2",
|
24 |
-
"Camille Corot": "camille-corot_subset1",
|
25 |
-
"Claude Monet": "monet_subset2",
|
26 |
-
"Pablo Picasso": "picasso_subset1",
|
27 |
-
"Jackson Pollock": "jackson-pollock_subset1",
|
28 |
-
"Gerhard Richter": "gerhard-richter_subset1",
|
29 |
-
"M.C. Escher": "m.c.-escher_subset1",
|
30 |
-
"Albert Gleizes": "albert-gleizes_subset1",
|
31 |
-
"Hokusai": "katsushika-hokusai_subset1",
|
32 |
-
"Wassily Kandinsky": "kandinsky_subset1",
|
33 |
-
"Gustav Klimt": "klimt_subset3",
|
34 |
-
"Roy Lichtenstein": "roy-lichtenstein_subset1",
|
35 |
-
"Henri Matisse": "henri-matisse_subset1",
|
36 |
-
"Joan Miro": "joan-miro_subset2",
|
37 |
-
}
|
38 |
-
@spaces.GPU
|
39 |
-
def demo_inference_gen(adapter_choice:str, prompt:str, samples:int=1,seed:int=0, steps=50, guidance_scale=7.5):
|
40 |
-
adapter_path = lora_map[adapter_choice]
|
41 |
-
if adapter_path not in [None, "None"]:
|
42 |
-
adapter_path = f"data/Art_adapters/{adapter_path}/adapter_alpha1.0_rank1_all_up_1000steps.pt"
|
43 |
-
|
44 |
-
prompts = [prompt]*samples
|
45 |
-
infer_loader = get_validation_dataloader(prompts)
|
46 |
-
network = get_lora_network(pipe.unet, adapter_path)["network"]
|
47 |
-
pred_images = inference(network, pipe.tokenizer, pipe.text_encoder, pipe.vae, pipe.unet, pipe.scheduler, infer_loader,
|
48 |
-
height=512, width=512, scales=[1.0],
|
49 |
-
save_dir=None, seed=seed,steps=steps, guidance_scale=guidance_scale,
|
50 |
-
start_noise=-1, show=False, style_prompt="sks art", no_load=True,
|
51 |
-
from_scratch=True, device=device)[0][1.0]
|
52 |
-
return pred_images
|
53 |
-
@spaces.GPU
|
54 |
-
def demo_inference_stylization(adapter_path:str, prompts:list, image:list, start_noise=800,seed:int=0):
|
55 |
-
infer_loader = get_validation_dataloader(prompts, image)
|
56 |
-
network = get_lora_network(pipe.unet, adapter_path,"all_up")["network"]
|
57 |
-
pred_images = inference(network, pipe.tokenizer, pipe.text_encoder, pipe.vae, pipe.unet, pipe.scheduler, infer_loader,
|
58 |
-
height=512, width=512, scales=[0.,1.],
|
59 |
-
save_dir=None, seed=seed,steps=20, guidance_scale=7.5,
|
60 |
-
start_noise=start_noise, show=True, style_prompt="sks art", no_load=True,
|
61 |
-
from_scratch=False, device=device)
|
62 |
-
return pred_images
|
63 |
-
|
64 |
-
# def infer(prompt, samples, steps, scale, seed):
|
65 |
-
# generator = torch.Generator(device=device).manual_seed(seed)
|
66 |
-
# images_list = pipe( # type: ignore
|
67 |
-
# [prompt] * samples,
|
68 |
-
# num_inference_steps=steps,
|
69 |
-
# guidance_scale=scale,
|
70 |
-
# generator=generator,
|
71 |
-
# )
|
72 |
-
# images = []
|
73 |
-
# safe_image = Image.open(r"data/unsafe.png")
|
74 |
-
# print(images_list)
|
75 |
-
# for i, image in enumerate(images_list["images"]): # type: ignore
|
76 |
-
# if images_list["nsfw_content_detected"][i]: # type: ignore
|
77 |
-
# images.append(safe_image)
|
78 |
-
# else:
|
79 |
-
# images.append(image)
|
80 |
-
# return images
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
block = gr.Blocks()
|
86 |
-
# Direct infer
|
87 |
-
with block:
|
88 |
-
with gr.Group():
|
89 |
-
gr.Markdown(" # Art-Free Diffusion Demo")
|
90 |
-
with gr.Row():
|
91 |
-
text = gr.Textbox(
|
92 |
-
label="Enter your prompt",
|
93 |
-
max_lines=2,
|
94 |
-
placeholder="Enter your prompt",
|
95 |
-
container=False,
|
96 |
-
value="Park with cherry blossom trees, picnicker’s and a clear blue pond.",
|
97 |
-
)
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
btn = gr.Button("Run", scale=0)
|
102 |
-
gallery = gr.Gallery(
|
103 |
-
label="Generated images",
|
104 |
-
show_label=False,
|
105 |
-
elem_id="gallery",
|
106 |
-
columns=[2],
|
107 |
-
)
|
108 |
-
|
109 |
-
advanced_button = gr.Button("Advanced options", elem_id="advanced-btn")
|
110 |
-
|
111 |
-
with gr.Row(elem_id="advanced-options"):
|
112 |
-
adapter_choice = gr.Dropdown(
|
113 |
-
label="Choose adapter",
|
114 |
-
choices=["None", "Andre Derain","Vincent van Gogh","Andy Warhol", "Walter Battiss",
|
115 |
-
"Camille Corot", "Claude Monet", "Pablo Picasso",
|
116 |
-
"Jackson Pollock", "Gerhard Richter", "M.C. Escher",
|
117 |
-
"Albert Gleizes", "Hokusai", "Wassily Kandinsky", "Gustav Klimt", "Roy Lichtenstein",
|
118 |
-
"Henri Matisse", "Joan Miro"
|
119 |
-
],
|
120 |
-
value="None"
|
121 |
-
)
|
122 |
-
# print(adapter_choice[0])
|
123 |
-
# lora_path = lora_map[adapter_choice.value]
|
124 |
-
# if lora_path is not None:
|
125 |
-
# lora_path = f"data/Art_adapters/{lora_path}/adapter_alpha1.0_rank1_all_up_1000steps.pt"
|
126 |
-
|
127 |
-
samples = gr.Slider(label="Images", minimum=1, maximum=4, value=1, step=1)
|
128 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=20, step=1)
|
129 |
-
scale = gr.Slider(
|
130 |
-
label="Guidance Scale", minimum=0, maximum=50, value=7.5, step=0.1
|
131 |
-
)
|
132 |
-
print(scale)
|
133 |
-
seed = gr.Slider(
|
134 |
-
label="Seed",
|
135 |
-
minimum=0,
|
136 |
-
maximum=2147483647,
|
137 |
-
step=1,
|
138 |
-
randomize=True,
|
139 |
-
)
|
140 |
-
|
141 |
-
gr.on([text.submit, btn.click], demo_inference_gen, inputs=[adapter_choice, text, samples, seed, steps, scale], outputs=gallery)
|
142 |
-
advanced_button.click(
|
143 |
-
None,
|
144 |
-
[],
|
145 |
-
text,
|
146 |
-
)
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
block.launch()
|
|
|
1 |
+
# Authors: Hui Ren (rhfeiyang.github.io)
|
2 |
+
import spaces
|
3 |
+
import os
|
4 |
+
import gradio as gr
|
5 |
+
from diffusers import DiffusionPipeline
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
import torch
|
8 |
+
from PIL import Image
|
9 |
+
|
10 |
+
|
11 |
+
|
12 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
13 |
+
dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float16
|
14 |
+
pipe = DiffusionPipeline.from_pretrained("rhfeiyang/art-free-diffusion-v1",
|
15 |
+
dtype=dtype).to(device)
|
16 |
+
|
17 |
+
from inference import get_lora_network, inference, get_validation_dataloader
|
18 |
+
lora_map = {
|
19 |
+
"None": "None",
|
20 |
+
"Andre Derain": "andre-derain_subset1",
|
21 |
+
"Vincent van Gogh": "van_gogh_subset1",
|
22 |
+
"Andy Warhol": "andy_subset1",
|
23 |
+
"Walter Battiss": "walter-battiss_subset2",
|
24 |
+
"Camille Corot": "camille-corot_subset1",
|
25 |
+
"Claude Monet": "monet_subset2",
|
26 |
+
"Pablo Picasso": "picasso_subset1",
|
27 |
+
"Jackson Pollock": "jackson-pollock_subset1",
|
28 |
+
"Gerhard Richter": "gerhard-richter_subset1",
|
29 |
+
"M.C. Escher": "m.c.-escher_subset1",
|
30 |
+
"Albert Gleizes": "albert-gleizes_subset1",
|
31 |
+
"Hokusai": "katsushika-hokusai_subset1",
|
32 |
+
"Wassily Kandinsky": "kandinsky_subset1",
|
33 |
+
"Gustav Klimt": "klimt_subset3",
|
34 |
+
"Roy Lichtenstein": "roy-lichtenstein_subset1",
|
35 |
+
"Henri Matisse": "henri-matisse_subset1",
|
36 |
+
"Joan Miro": "joan-miro_subset2",
|
37 |
+
}
|
38 |
+
@spaces.GPU
|
39 |
+
def demo_inference_gen(adapter_choice:str, prompt:str, samples:int=1,seed:int=0, steps=50, guidance_scale=7.5):
|
40 |
+
adapter_path = lora_map[adapter_choice]
|
41 |
+
if adapter_path not in [None, "None"]:
|
42 |
+
adapter_path = f"data/Art_adapters/{adapter_path}/adapter_alpha1.0_rank1_all_up_1000steps.pt"
|
43 |
+
|
44 |
+
prompts = [prompt]*samples
|
45 |
+
infer_loader = get_validation_dataloader(prompts)
|
46 |
+
network = get_lora_network(pipe.unet, adapter_path)["network"]
|
47 |
+
pred_images = inference(network, pipe.tokenizer, pipe.text_encoder, pipe.vae, pipe.unet, pipe.scheduler, infer_loader,
|
48 |
+
height=512, width=512, scales=[1.0],
|
49 |
+
save_dir=None, seed=seed,steps=steps, guidance_scale=guidance_scale,
|
50 |
+
start_noise=-1, show=False, style_prompt="sks art", no_load=True,
|
51 |
+
from_scratch=True, device=device)[0][1.0]
|
52 |
+
return pred_images
|
53 |
+
@spaces.GPU
|
54 |
+
def demo_inference_stylization(adapter_path:str, prompts:list, image:list, start_noise=800,seed:int=0):
|
55 |
+
infer_loader = get_validation_dataloader(prompts, image)
|
56 |
+
network = get_lora_network(pipe.unet, adapter_path,"all_up")["network"]
|
57 |
+
pred_images = inference(network, pipe.tokenizer, pipe.text_encoder, pipe.vae, pipe.unet, pipe.scheduler, infer_loader,
|
58 |
+
height=512, width=512, scales=[0.,1.],
|
59 |
+
save_dir=None, seed=seed,steps=20, guidance_scale=7.5,
|
60 |
+
start_noise=start_noise, show=True, style_prompt="sks art", no_load=True,
|
61 |
+
from_scratch=False, device=device)
|
62 |
+
return pred_images
|
63 |
+
|
64 |
+
# def infer(prompt, samples, steps, scale, seed):
|
65 |
+
# generator = torch.Generator(device=device).manual_seed(seed)
|
66 |
+
# images_list = pipe( # type: ignore
|
67 |
+
# [prompt] * samples,
|
68 |
+
# num_inference_steps=steps,
|
69 |
+
# guidance_scale=scale,
|
70 |
+
# generator=generator,
|
71 |
+
# )
|
72 |
+
# images = []
|
73 |
+
# safe_image = Image.open(r"data/unsafe.png")
|
74 |
+
# print(images_list)
|
75 |
+
# for i, image in enumerate(images_list["images"]): # type: ignore
|
76 |
+
# if images_list["nsfw_content_detected"][i]: # type: ignore
|
77 |
+
# images.append(safe_image)
|
78 |
+
# else:
|
79 |
+
# images.append(image)
|
80 |
+
# return images
|
81 |
+
|
82 |
+
|
83 |
+
|
84 |
+
|
85 |
+
block = gr.Blocks()
|
86 |
+
# Direct infer
|
87 |
+
with block:
|
88 |
+
with gr.Group():
|
89 |
+
gr.Markdown(" # Art-Free Diffusion Demo")
|
90 |
+
with gr.Row():
|
91 |
+
text = gr.Textbox(
|
92 |
+
label="Enter your prompt",
|
93 |
+
max_lines=2,
|
94 |
+
placeholder="Enter your prompt",
|
95 |
+
container=False,
|
96 |
+
value="Park with cherry blossom trees, picnicker’s and a clear blue pond.",
|
97 |
+
)
|
98 |
+
|
99 |
+
|
100 |
+
|
101 |
+
btn = gr.Button("Run", scale=0)
|
102 |
+
gallery = gr.Gallery(
|
103 |
+
label="Generated images",
|
104 |
+
show_label=False,
|
105 |
+
elem_id="gallery",
|
106 |
+
columns=[2],
|
107 |
+
)
|
108 |
+
|
109 |
+
advanced_button = gr.Button("Advanced options", elem_id="advanced-btn")
|
110 |
+
|
111 |
+
with gr.Row(elem_id="advanced-options"):
|
112 |
+
adapter_choice = gr.Dropdown(
|
113 |
+
label="Choose adapter",
|
114 |
+
choices=["None", "Andre Derain","Vincent van Gogh","Andy Warhol", "Walter Battiss",
|
115 |
+
"Camille Corot", "Claude Monet", "Pablo Picasso",
|
116 |
+
"Jackson Pollock", "Gerhard Richter", "M.C. Escher",
|
117 |
+
"Albert Gleizes", "Hokusai", "Wassily Kandinsky", "Gustav Klimt", "Roy Lichtenstein",
|
118 |
+
"Henri Matisse", "Joan Miro"
|
119 |
+
],
|
120 |
+
value="None"
|
121 |
+
)
|
122 |
+
# print(adapter_choice[0])
|
123 |
+
# lora_path = lora_map[adapter_choice.value]
|
124 |
+
# if lora_path is not None:
|
125 |
+
# lora_path = f"data/Art_adapters/{lora_path}/adapter_alpha1.0_rank1_all_up_1000steps.pt"
|
126 |
+
|
127 |
+
samples = gr.Slider(label="Images", minimum=1, maximum=4, value=1, step=1)
|
128 |
+
steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=20, step=1)
|
129 |
+
scale = gr.Slider(
|
130 |
+
label="Guidance Scale", minimum=0, maximum=50, value=7.5, step=0.1
|
131 |
+
)
|
132 |
+
print(scale)
|
133 |
+
seed = gr.Slider(
|
134 |
+
label="Seed",
|
135 |
+
minimum=0,
|
136 |
+
maximum=2147483647,
|
137 |
+
step=1,
|
138 |
+
randomize=True,
|
139 |
+
)
|
140 |
+
|
141 |
+
gr.on([text.submit, btn.click], demo_inference_gen, inputs=[adapter_choice, text, samples, seed, steps, scale], outputs=gallery)
|
142 |
+
advanced_button.click(
|
143 |
+
None,
|
144 |
+
[],
|
145 |
+
text,
|
146 |
+
)
|
147 |
+
|
148 |
+
|
149 |
+
|
150 |
block.launch()
|