Existing evaluations of video generation models do not test whether repeated generations recover the correct distribution of outcomes

existing evaluations largely assess individual-video plausibility and do not test whether repeated generations recover the correct distribution

http://arxiv.org/abs/2608.27345v1