File size: 2,096 Bytes
cfccdc2
d77a781
596d353
d77a781
0006a92
586ddad
 
 
cfccdc2
e6d1b4c
 
 
91e9183
0eaa580
c2305a3
 
 
 
 
36c3ba1
 
c2305a3
 
 
 
 
 
 
 
 
36c3ba1
c2305a3
 
 
 
 
 
 
 
b2f39c8
a94928e
 
 
 
 
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
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 out of order** \n 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.
"""


with demo:
    gr.Markdown(description)
    gr.Markdown(article)
demo.launch()