3 ms·
They're not necessarily plagiarized anymore than a human artist plagiarizes from looking at an existing image. If you accepted this, Disney would basically own
by astrange 4y ago
They're not necessarily plagiarized anymore than a human artist plagiarizes from looking at an existing image. If you accepted this, Disney would basically own every animation ever made.
StableDiffusion is a 10GB model made from 100TB of images, so there probably isn't even room to have memorized inputs - there's also some built-in defenses against that.
It's possible to make an AI that does explicitly make collages out of inputs (that goes under "fine tuning") and such a thing would be useful. OpenAI has a similar paper about a text AI that cites its sources called WebGPT.