Copyright search-related features request...

#65
by lathropa - opened

So, I was thinking that if people are intending to use any of the generated images commercially, the issue might come up as to how close the generated images are to something in the training set. (This might also be of interest to academic researchers for other reasons.) With that in mind, two feature ideas for the interface:

  1. For some reasonable n, and some available metric of similarity, show me the n closest images in the training set to the currently generated image

  2. For the currently generated image, click to reverse-image search the web (e.g. Google image search) to see if what you've generated is very close to something that might infringe copyright on an existing image online

Huggingface Projects org
edited Sep 13, 2022

https://lexica.art/ already does similarity search between generated images. I guess some other people can/should extend https://lexica.art/ to also include the trained data in the similarity search. (in Stable Diffusion case, it was laion-5b)

Sign up or log in to comment