Frechet Forcing: Training Long-Horizon Streaming World Models from Short Clips

One-line takeaway

Frechet Forcing trains long-horizon streaming video world models from short clips: instead of distillation from an external teacher that ignores actions, it trains long rollouts to remain realistic, action-following continuations of their own scene, reducing drift beyond the training horizon.

Frechet Forcing 解决只用短片段训练的流式视频世界模型在长时间滚动时漂移的问题：不依赖对动作不敏感的外部教师模型，而是要求长滚动结果仍是本场景、遵循动作的真实延续来训练。