File size: 2,276 Bytes
cfccdc2
d77a781
596d353
d77a781
0006a92
586ddad
 
 
cfccdc2
e6d1b4c
 
 
fdb1186
0eaa580
c2305a3
 
 
 
 
36c3ba1
 
c2305a3
 
 
 
 
 
 
 
 
36c3ba1
c2305a3
 
 
 
 
 
 
 
b2f39c8
e6d1b4c
b2f39c8
2370152
b2f39c8
2652bae
fdb1186
dfa6c0a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
import gradio as gr
import numpy as np
# from edict_functions import EDICT_editing
from PIL import Image
from utils import Endpoint, get_token
from io import BytesIO
import requests


def f(x):
    return x

description = '*This demo is temporarily suspended for internal reasons* A gradio demo for [EDICT](https://arxiv.org/abs/2211.12446) (CVPR23)'
# description = gr.Markdown(description)

article = """

### Prompting Style

As with many text-to-image methods, the prompting style of EDICT can make a big difference. When in doubt, experiment! Some guidance:
* Parallel *Original Description* and *Edit Description* construction as much as possible. Inserting/editing single words often is enough to affect a change while maintaining a lot of the original structure
* Words that will affect the entire setting (e.g. "A photo of " vs. "A painting of") can make a big difference. Playing around with them can help a lot

### Parameters
Both `edit_strength` and `guidance_scale` have similar properties qualitatively: the higher the value the more the image will change. We suggest
* Increasing/decreasing `edit_strength` first, particularly to alter/preserve more of the original structure/content
* Then changing `guidance_scale` to make the change in the edited region more or less pronounced.

Usually we find changing `edit_strength` to be enough, but feel free to play around (and report any interesting results)!

### Misc.

Having difficulty coming up with a caption? Try [BLIP](https://huggingface.co/spaces/Salesforce/BLIP2) to automatically generate one!

As with most StableDiffusion approaches, faces/text are often problematic to render, especially if they're small. Having these in the foreground will help keep them cleaner.

A returned black image means that the [Safety Checker](https://huggingface.co/CompVis/stable-diffusion-safety-checker) triggered on the photo. This happens in odd cases sometimes (it often rejects
the huggingface logo or variations), but we need to keep it in for obvious reasons.
"""

iface = gr.Interface(fn=f,
                     inputs=['image'],
                     # examples = examples,
                     outputs="image",
                     description=description,
                     article=article)
iface.launch()