Comment by KTibow
3 years ago
I might be missing something because I don't know much about the architecture of either Nightshade or AI art generators, but I wonder if you could try to have a GAN-like architecture (an extra model trying to trick the model) for the part of the generator that labels images to build resistance to Nightshade-like filters.
It doesn't even have to be a full GAN, you only need to train the discriminator side to filter out the data. Clean reference images + Nightshade would be the generator side.