whereas anime more broadly is probably pulling in a lot of poorer-quality anime art...(Also, sometimes if asked for “anime” it gives me content that either looks like 3D rendered video game cutscenes, or occasionally what I assume is meant to be people at an anime con in cosplay.)
That’s how you know it’s not a problem of pulling in lots of poorer-quality anime art. First, poorer-quality doesn’t impede learning that much; remember, you just prompt for high-quality. Compute allowing, more n is always better. And second, if it was a master of poorer-quality anime drawings, it wouldn’t be desperately ‘sliding away’, if you will, like squeezing a balloon, from rendering true anime, as opposed to CGI of anime or Western fanart of anime or photographs of physical objects related to anime. It would just do it (perhaps generating poorer-quality anime), not generate high-quality samples of everything but anime. (See my comment there for more examples.)
The problem is it’s somehow not trained on anime. Everything it knows about anime seems to come primarily from adjacent images and the CLIP guidance (which does know plenty about anime, but we also know that pixel generation from CLIP guidance never works as well).
Thread of all known anime examples.
That’s how you know it’s not a problem of pulling in lots of poorer-quality anime art. First, poorer-quality doesn’t impede learning that much; remember, you just prompt for high-quality. Compute allowing, more n is always better. And second, if it was a master of poorer-quality anime drawings, it wouldn’t be desperately ‘sliding away’, if you will, like squeezing a balloon, from rendering true anime, as opposed to CGI of anime or Western fanart of anime or photographs of physical objects related to anime. It would just do it (perhaps generating poorer-quality anime), not generate high-quality samples of everything but anime. (See my comment there for more examples.)
The problem is it’s somehow not trained on anime. Everything it knows about anime seems to come primarily from adjacent images and the CLIP guidance (which does know plenty about anime, but we also know that pixel generation from CLIP guidance never works as well).