Spaces:
Runtime error
Runtime error
Georgiy Grigorev
commited on
Commit
•
054082d
1
Parent(s):
c85ee6c
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import os
|
3 |
+
from torch.optim import AdamW
|
4 |
+
from diffusers import StableDiffusionPipeline
|
5 |
+
from torch import autocast, inference_mode
|
6 |
+
import torch
|
7 |
+
import numpy as np
|
8 |
+
|
9 |
+
from scheduling_ddim import DDIMScheduler
|
10 |
+
|
11 |
+
|
12 |
+
device = 'cuda'
|
13 |
+
# don't forget to add your token or comment if already logged in
|
14 |
+
pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5",
|
15 |
+
scheduler=DDIMScheduler(beta_end=0.012,
|
16 |
+
beta_schedule="scaled_linear",
|
17 |
+
beta_start=0.00085),
|
18 |
+
use_auth_token="").to(device)
|
19 |
+
_ = pipe.vae.requires_grad_(False)
|
20 |
+
_ = pipe.text_encoder.requires_grad_(False)
|
21 |
+
_ = pipe.unet.requires_grad_(False)
|
22 |
+
|
23 |
+
def preprocess(image):
|
24 |
+
w, h = image.size
|
25 |
+
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
|
26 |
+
image = image.resize((w, h), resample=Image.LANCZOS)
|
27 |
+
image = np.array(image).astype(np.float32) / 255.0
|
28 |
+
image = image[None].transpose(0, 3, 1, 2)
|
29 |
+
image = torch.from_numpy(image)
|
30 |
+
return 2.0 * image - 1.0
|
31 |
+
|
32 |
+
def im2latent(pipe, im, generator):
|
33 |
+
init_image = preprocess(im).to(pipe.device)
|
34 |
+
init_latent_dist = pipe.vae.encode(init_image).latent_dist
|
35 |
+
init_latents = init_latent_dist.sample(generator=generator)
|
36 |
+
|
37 |
+
return init_latents * 0.18215
|
38 |
+
|
39 |
+
|
40 |
+
def image_mod(init_image, source_prompt, prompt, scale, steps, seed):
|
41 |
+
# fix seed
|
42 |
+
g = torch.Generator(device=pipe.device).manual_seed(84)
|
43 |
+
|
44 |
+
image_latents = im2latent(pipe, init_image, g)
|
45 |
+
pipe.scheduler.set_timesteps(steps)
|
46 |
+
# use text describing an image
|
47 |
+
# source_prompt = "a photo of a woman"
|
48 |
+
context = pipe._encode_prompt(source_prompt, pipe.device, 1, False, "")
|
49 |
+
|
50 |
+
decoded_latents = image_latents.clone()
|
51 |
+
with autocast(device), inference_mode():
|
52 |
+
# we are pivoting timesteps as we are moving in opposite direction
|
53 |
+
timesteps = pipe.scheduler.timesteps.flip(0)
|
54 |
+
# this would be our targets for pivoting
|
55 |
+
init_trajectory = torch.empty(len(timesteps), *decoded_latents.size()[1:], device=decoded_latents.device, dtype=decoded_latents.dtype)
|
56 |
+
for i, t in enumerate(tqdm(timesteps)):
|
57 |
+
init_trajectory[i:i+1] = decoded_latents
|
58 |
+
noise_pred = pipe.unet(decoded_latents, t, encoder_hidden_states=context).sample
|
59 |
+
decoded_latents = pipe.scheduler.reverse_step(noise_pred, t, decoded_latents).next_sample
|
60 |
+
|
61 |
+
# we would need to flip trajectory values for pivoting in right direction
|
62 |
+
init_trajectory = init_trajectory.cpu().flip(0)
|
63 |
+
|
64 |
+
latents = decoded_latents.clone()
|
65 |
+
context_uncond = pipe._encode_prompt("", pipe.device, 1, False, "")
|
66 |
+
# we will be optimizing uncond text embedding
|
67 |
+
context_uncond.requires_grad_(True)
|
68 |
+
|
69 |
+
# use same text
|
70 |
+
# prompt = "a photo of a woman"
|
71 |
+
context_cond = pipe._encode_prompt(prompt, pipe.device, 1, False, "")
|
72 |
+
|
73 |
+
# default lr works
|
74 |
+
opt = AdamW([context_uncond])
|
75 |
+
|
76 |
+
# concat latents for classifier-free guidance
|
77 |
+
latents = torch.cat([latents, latents])
|
78 |
+
latents.requires_grad_(True)
|
79 |
+
context = torch.cat((context_uncond, context_cond))
|
80 |
+
|
81 |
+
with autocast(device):
|
82 |
+
for i, t in enumerate(tqdm(pipe.scheduler.timesteps)):
|
83 |
+
latents = pipe.scheduler.scale_model_input(latents, t)
|
84 |
+
uncond, cond = pipe.unet(latents, t, encoder_hidden_states=context).sample.chunk(2)
|
85 |
+
with torch.enable_grad():
|
86 |
+
latents = pipe.scheduler.step(uncond + scale * (cond - uncond), t, latents, generator=g).prev_sample
|
87 |
+
|
88 |
+
opt.zero_grad()
|
89 |
+
# optimize uncond text emb
|
90 |
+
pivot_value = init_trajectory[[i]].to(pipe.device)
|
91 |
+
(latents - pivot_value).mean().backward()
|
92 |
+
opt.step()
|
93 |
+
latents = latents.detach()
|
94 |
+
|
95 |
+
images = pipe.decode_latents(latents)
|
96 |
+
im = pipe.numpy_to_pil(images)[0]
|
97 |
+
return im
|
98 |
+
|
99 |
+
|
100 |
+
demo = gr.Interface(
|
101 |
+
image_mod,
|
102 |
+
inputs=[gr.Image(type="pil"), gr.Textbox("a photo of a person"), gr.Textbox("a photo of a person"), gr.Slider(0, 10, 0.5, 0.1), gr.Slider(0, 100, 51, 1), gr.Number(42)],
|
103 |
+
outputs="image",
|
104 |
+
flagging_options=["blurry", "incorrect", "other"], examples=[
|
105 |
+
os.path.join(os.path.dirname(__file__), "images/00001.jpg"),
|
106 |
+
])
|
107 |
+
|
108 |
+
if __name__ == "__main__":
|
109 |
+
demo.launch()
|