Spaces:
Running
on
A10G
Running
on
A10G
Linoy Tsaban
commited on
Commit
·
4697625
1
Parent(s):
acc80f0
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,259 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
import requests
|
4 |
+
from io import BytesIO
|
5 |
+
from diffusers import StableDiffusionPipeline
|
6 |
+
from diffusers import DDIMScheduler
|
7 |
+
from utils import *
|
8 |
+
from inversion_utils import *
|
9 |
+
from torch import autocast, inference_mode
|
10 |
+
import re
|
11 |
+
|
12 |
+
def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):
|
13 |
+
|
14 |
+
# inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf,
|
15 |
+
# based on the code in https://github.com/inbarhub/DDPM_inversion
|
16 |
+
|
17 |
+
# returns wt, zs, wts:
|
18 |
+
# wt - inverted latent
|
19 |
+
# wts - intermediate inverted latents
|
20 |
+
# zs - noise maps
|
21 |
+
|
22 |
+
sd_pipe.scheduler.set_timesteps(num_diffusion_steps)
|
23 |
+
|
24 |
+
# vae encode image
|
25 |
+
with autocast("cuda"), inference_mode():
|
26 |
+
w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215).float()
|
27 |
+
|
28 |
+
# find Zs and wts - forward process
|
29 |
+
wt, zs, wts = inversion_forward_process(sd_pipe, w0, etas=eta, prompt=prompt_src, cfg_scale=cfg_scale_src, prog_bar=True, num_inference_steps=num_diffusion_steps)
|
30 |
+
return wt, zs, wts
|
31 |
+
|
32 |
+
|
33 |
+
|
34 |
+
def sample(wt, zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
|
35 |
+
|
36 |
+
# reverse process (via Zs and wT)
|
37 |
+
w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=True, zs=zs[skip:])
|
38 |
+
|
39 |
+
# vae decode image
|
40 |
+
with autocast("cuda"), inference_mode():
|
41 |
+
x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample
|
42 |
+
if x0_dec.dim()<4:
|
43 |
+
x0_dec = x0_dec[None,:,:,:]
|
44 |
+
img = image_grid(x0_dec)
|
45 |
+
return img
|
46 |
+
|
47 |
+
# load pipelines
|
48 |
+
# sd_model_id = "runwayml/stable-diffusion-v1-5"
|
49 |
+
sd_model_id = "CompVis/stable-diffusion-v1-4"
|
50 |
+
# sd_model_id = "stabilityai/stable-diffusion-2-base"
|
51 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
52 |
+
sd_pipe = StableDiffusionPipeline.from_pretrained(sd_model_id).to(device)
|
53 |
+
sd_pipe.scheduler = DDIMScheduler.from_config(sd_model_id, subfolder = "scheduler")
|
54 |
+
|
55 |
+
|
56 |
+
|
57 |
+
def get_example():
|
58 |
+
case = [
|
59 |
+
[
|
60 |
+
'examples/source_a_man_wearing_a_brown_hoodie_in_a_crowded_street.jpeg',
|
61 |
+
'a man wearing a brown hoodie in a crowded street',
|
62 |
+
'a robot wearing a brown hoodie in a crowded street',
|
63 |
+
100,
|
64 |
+
36,
|
65 |
+
15,
|
66 |
+
'+painting',
|
67 |
+
10,
|
68 |
+
1,
|
69 |
+
'examples/ddpm_a_robot_wearing_a_brown_hoodie_in_a_crowded_street.png',
|
70 |
+
'examples/ddpm_sega_painting_of_a_robot_wearing_a_brown_hoodie_in_a_crowded_street.png'
|
71 |
+
],
|
72 |
+
[
|
73 |
+
'examples/source_wall_with_framed_photos.jpeg',
|
74 |
+
'',
|
75 |
+
'',
|
76 |
+
100,
|
77 |
+
36,
|
78 |
+
15,
|
79 |
+
'+pink drawings of muffins',
|
80 |
+
10,
|
81 |
+
1,
|
82 |
+
'examples/ddpm_wall_with_framed_photos.png',
|
83 |
+
'examples/ddpm_sega_plus_pink_drawings_of_muffins.png'
|
84 |
+
],
|
85 |
+
[
|
86 |
+
'examples/source_an_empty_room_with_concrete_walls.jpg',
|
87 |
+
'an empty room with concrete walls',
|
88 |
+
'glass walls',
|
89 |
+
100,
|
90 |
+
36,
|
91 |
+
17,
|
92 |
+
'+giant elephant',
|
93 |
+
10,
|
94 |
+
1,
|
95 |
+
'examples/ddpm_glass_walls.png',
|
96 |
+
'examples/ddpm_sega_glass_walls_gian_elephant.png'
|
97 |
+
]]
|
98 |
+
return case
|
99 |
+
|
100 |
+
inversion_map = dict()
|
101 |
+
|
102 |
+
def invert(input_image,
|
103 |
+
src_prompt ="",
|
104 |
+
steps=100,
|
105 |
+
src_cfg_scale = 3.5,
|
106 |
+
left = 0,
|
107 |
+
right = 0,
|
108 |
+
top = 0,
|
109 |
+
bottom = 0
|
110 |
+
):
|
111 |
+
# offsets=(0,0,0,0)
|
112 |
+
x0 = load_512(input_image, left,right, top, bottom, device)
|
113 |
+
|
114 |
+
|
115 |
+
# invert
|
116 |
+
wt, zs, wts = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale)
|
117 |
+
|
118 |
+
latnets = wts[skip].expand(1, -1, -1, -1)
|
119 |
+
inversion_map['latnets'] = latnets
|
120 |
+
inversion_map['zs'] = zs
|
121 |
+
inversion_map['wts'] = wts
|
122 |
+
|
123 |
+
|
124 |
+
|
125 |
+
|
126 |
+
return
|
127 |
+
|
128 |
+
def edit(tar_prompt="",
|
129 |
+
steps=100,
|
130 |
+
skip=36,
|
131 |
+
tar_cfg_scale=15,
|
132 |
+
|
133 |
+
):
|
134 |
+
outputs = []
|
135 |
+
num_generations = 1
|
136 |
+
for i in range(num_generations):
|
137 |
+
out = sample(wt, zs, wts, prompt_tar=tar_prompt,
|
138 |
+
cfg_scale_tar=tar_cfg_scale, skip=skip)
|
139 |
+
outputs.append(out)
|
140 |
+
|
141 |
+
return outputs
|
142 |
+
|
143 |
+
def reset():
|
144 |
+
inversion_map.clear()
|
145 |
+
|
146 |
+
|
147 |
+
########
|
148 |
+
# demo #
|
149 |
+
########
|
150 |
+
|
151 |
+
intro = """
|
152 |
+
<h1 style="font-weight: 1400; text-align: center; margin-bottom: 7px;">
|
153 |
+
Edit Friendly DDPM Inversion
|
154 |
+
</h1>
|
155 |
+
<p style="font-size: 0.9rem; text-align: center; margin: 0rem; line-height: 1.2em; margin-top:1em">
|
156 |
+
<a href="https://arxiv.org/abs/2301.12247" style="text-decoration: underline;" target="_blank">An Edit Friendly DDPM Noise Space:
|
157 |
+
Inversion and Manipulations </a>
|
158 |
+
<p/>
|
159 |
+
<p style="font-size: 0.9rem; margin: 0rem; line-height: 1.2em; margin-top:1em">
|
160 |
+
For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
|
161 |
+
<a href="https://huggingface.co/spaces/LinoyTsaban/ddpm_sega?duplicate=true">
|
162 |
+
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
|
163 |
+
<p/>"""
|
164 |
+
with gr.Blocks() as demo:
|
165 |
+
gr.HTML(intro)
|
166 |
+
with gr.Row():
|
167 |
+
src_prompt = gr.Textbox(lines=1, label="Source Prompt", interactive=True, placeholder="optional: describe the original image")
|
168 |
+
tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True, placeholder="optional: describe the target image to edit with DDPM")
|
169 |
+
|
170 |
+
with gr.Row():
|
171 |
+
input_image = gr.Image(label="Input Image", interactive=True)
|
172 |
+
input_image.style(height=512, width=512)
|
173 |
+
output_image = gr.Image(label=f"Edited Image", interactive=False)
|
174 |
+
output_image.style(height=512, width=512)
|
175 |
+
|
176 |
+
|
177 |
+
with gr.Row():
|
178 |
+
with gr.Column(scale=1, min_width=100):
|
179 |
+
invert_button = gr.Button("Load & Invert")
|
180 |
+
with gr.Column(scale=1, min_width=100):
|
181 |
+
edit_button = gr.Button("Sample & Edit")
|
182 |
+
|
183 |
+
|
184 |
+
with gr.Accordion("Advanced Options", open=False):
|
185 |
+
with gr.Row():
|
186 |
+
with gr.Column():
|
187 |
+
#inversion
|
188 |
+
steps = gr.Number(value=100, precision=0, label="Num Diffusion Steps", interactive=True)
|
189 |
+
src_cfg_scale = gr.Slider(minimum=1, maximum=15, value=3.5, label=f"Source Guidance Scale", interactive=True)
|
190 |
+
|
191 |
+
# reconstruction
|
192 |
+
skip = gr.Slider(minimum=0, maximum=40, value=36, precision=0, label="Skip Steps", interactive=True)
|
193 |
+
tar_cfg_scale = gr.Slider(minimum=7, maximum=18,value=15, label=f"Target Guidance Scale", interactive=True)
|
194 |
+
|
195 |
+
#shift
|
196 |
+
with gr.Column():
|
197 |
+
left = gr.Number(value=0, precision=0, label="Left Shift", interactive=True)
|
198 |
+
right = gr.Number(value=0, precision=0, label="Right Shift", interactive=True)
|
199 |
+
top = gr.Number(value=0, precision=0, label="Top Shift", interactive=True)
|
200 |
+
bottom = gr.Number(value=0, precision=0, label="Bottom Shift", interactive=True)
|
201 |
+
|
202 |
+
|
203 |
+
|
204 |
+
|
205 |
+
# gr.Markdown(help_text)
|
206 |
+
|
207 |
+
invert_button.click(
|
208 |
+
fn=invert,
|
209 |
+
inputs=[input_image,
|
210 |
+
src_prompt,
|
211 |
+
steps,
|
212 |
+
src_cfg_scale,
|
213 |
+
left,
|
214 |
+
right,
|
215 |
+
top,
|
216 |
+
bottom
|
217 |
+
],
|
218 |
+
outputs = [],
|
219 |
+
)
|
220 |
+
|
221 |
+
edit_button.click(
|
222 |
+
fn=edit,
|
223 |
+
inputs=[tar_prompt,
|
224 |
+
steps,
|
225 |
+
skip,
|
226 |
+
tar_cfg_scale,
|
227 |
+
],
|
228 |
+
outputs=[output_image],
|
229 |
+
)
|
230 |
+
|
231 |
+
|
232 |
+
|
233 |
+
|
234 |
+
input_image.change(
|
235 |
+
fn = reset
|
236 |
+
)
|
237 |
+
|
238 |
+
# gr.Examples(
|
239 |
+
# label='Examples',
|
240 |
+
# examples=get_example(),
|
241 |
+
# inputs=[input_image, src_prompt, tar_prompt, steps,
|
242 |
+
# # src_cfg_scale,
|
243 |
+
# skip,
|
244 |
+
# tar_cfg_scale,
|
245 |
+
# edit_concept,
|
246 |
+
# sega_edit_guidance,
|
247 |
+
# warm_up,
|
248 |
+
# # neg_guidance,
|
249 |
+
# ddpm_edited_image, sega_edited_image
|
250 |
+
# ],
|
251 |
+
# outputs=[ddpm_edited_image, sega_edited_image],
|
252 |
+
# # fn=edit,
|
253 |
+
# # cache_examples=True
|
254 |
+
# )
|
255 |
+
|
256 |
+
|
257 |
+
|
258 |
+
demo.queue()
|
259 |
+
demo.launch(share=False)
|