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()