Enables temporally coherent video generation with decentralized training, doubling FVD improvement and boosting CLIP similarity and aesthetics under equal compute.

arxiv.org/abs/2605.26064