Re:Stable Diffusion Frivolous Litigation
by godelski on Sunday January 15, 2023 @04:21PM
(#34393554)
> the person coaxing copyrighted material from the latent space
I'd be careful with this terminology. These models are able to generate more diversity in data than they are able to memorize. So I just want to make sure readers do not read this as "copying and combining things to make an image" in a copy/paste manner rather than the same manner that humans do (very different mechanisms) of having a broader understanding of shapes and fuzzy concepts. But I'm also a generative modeling researcher and so I'm biased to be pedantic.
Re:Stable Diffusion Frivolous Litigation
by williamcotton on Sunday January 15, 2023 @05:15PM
(#34394169)
Thanks for the clarification!
Now, a tool that burdens an individual with a torrent of copy-written material might not be worth using, but that is separate from the data embedded in the latent space of the model itself.