One would be training content censorship models.

Another would be training general image generation models, because the trick to getting human details like fingers right turned out to be training on erotica then attempting to censor the model (if at all)