Re:Stable Diffusion 1.5
by tarsinge on Friday October 21, 2022 @01:29AM
(#33284821)
Is it possible to tell from the weights that a model was trained on copyrighted material?
[ Parent
| on HN ]
possibly but you'd need to do a couple of transformations first.
https://youtu.be/0_BBRNYInx8?t=85 This video (released yesterday) talks about how SD takes an image, converts it to latents.
You'd need to decode those latents back to an image representation and scan. (possibly other ways but that's the most straightforward I can come up with although time intensive).