DreamBooth really should have destroyed all doubt about this point point, as the AI can generated highly specific images of subjects that aren't even in the original training set.
I haven't seen any concrete examples of this yet.
This is a common misunderstanding. Image AIs don’t copy/paste anything, or read from a database. Stable Diffusion is only 2GB. It’s little more than a huge array of numbers, which relate to different tokens.
Finding the “midpoint” between a set of tokens is latent space. What emerges is nothing short of miraculous, but it comes from random noise, not copy/pasting anything.
One obvious counterexample is stylegan interpolations. If you interpolate between two images, the midpoint is usually unique — it’s often not obvious what the source material was. (E.g. Gwern’s anime interpolations; sure, they’re anime faces, but from where? “All of danbooru” might as well be “all styles of anime ever created.”)
Maybe someone else can argue the point further.
There's a lot of good discussion to be had around the ethics of AI art, training on copyrighted materials, etc. But it is equivocally not just copy pasting from a database of images.
> But it is equivocally not just copy pasting from a database of images.
Just because you put a lossy compression step in the middle doesn't mean it's not copy-paste anymore.
You can use it to make procedurally-generated art, but it's still obvious what the source material was.