What if someone trains a model on public domain works only? It hasn't been done, because the law doesn't require it. LAION-5B is a big mess and its caption qua

What if someone trains a model on public domain works only? It hasn't been done, because the law doesn't require it. LAION-5B is a big mess and its caption quality is all over the place. Stability.ai basically chose a brute-force approach and threw a lot of data at the problem. But now imagine if someone took all the public domain images that they could find, properly labelled them and then trained a base model on all of that. I'm pretty sure that its output would also be very good. There are a ton of public domain photos out there so photorealism shouldn't be a problem. So - what's your argument then?