ArtBench

Nature paper: diffusion outputs often trace back to no single image, and no single creator

The study reports that any one sample or creator can usually be removed from a large training set without changing a given output. Per-item credit schemes need a mechanism this result denies them.