I think training on output is one method that people use to create customized Stable Diffusion models. IIRC, it doesn't work well unless you manually label it,

I think training on output is one method that people use to create customized Stable Diffusion models. IIRC, it doesn't work well unless you manually label it, in which case it works fine. Letting existing AI label it doesn't do very well... but maybe GPT-4 will be good enough at that sort of task?