Stochastic Video Generation with a Learned Prior.

Generating video frames that accurately predict future world states ischallenging. Existing approaches either fail to capture the full distributionof outcomes, or yield blurry generations, or both. In this paper we introducean unsupervised video generation model that learns a prior model of uncertaintyin a given environment. Video frames are generated by drawing samples from thisprior and combining them with a deterministic estimate of the future frame. Theapproach is simple and easily trained end-to-end on a variety of datasets.Sample generations are both varied and sharp, even many frames into the future,and compare favorably to those from existing approaches.