Comment by minimaxir

A few months ago I made a proof-of-concept on how finetuning Stable Diffusion XL on known bad/incoherent images can actually allow it to output "better" images if those images are used as a negative prompt, i.e. specifying a high-dimensional area of the latent space that model generation should stay away from: https://news.ycombinator.com/item?id=37211519

There's a nonzero chance that encouraging the creation of a large dataset of known tampered data can ironically improve generative AI art models by allowing the model to recognize tampered data and allow the training process to work around it.