Dear Twitter, there was recently a paper showing that different diffusion models generate very similar images when starting with the same initial noise. I did not note the reference. Somebody, please, has it?
@zdeborova You might be thinking about this one? arxiv.org/abs/2310.05264 (though this phenomenon was known before this paper) See also arxiv.org/abs/2310.02557, where we show that it happens even for diffusion models trained on different samples from the same training distribution
@FlorentinGuth Thanks. But if it was known before, who should be credited in your opinion?
@zdeborova I remembered reading it in one of the earlier papers, and I finally found it: it's mentioned in arxiv.org/abs/2011.13456 when introducing the probability flow ODE (bottom of page 7 and appendix D.5)