There is this interesting paper in which training data was extracted from a diffusion model, so images were able to be reproduced with great detail. It's beyond my knowledge, but I think it's only reproducible if the images appear multiple times in the training data. https://arxiv.org/abs/2301.13188

Comments