Fréchet Forcing: Advancements in Long-Horizon Streaming World Models

Fréchet Forcing has been introduced as a method to train long-horizon streaming world models. It utilizes self-generated rollouts and statistical fine-tuning with real data to prevent long-term drift in streaming video world models, eliminating the need for teachers or long videos. The MiniWorld-1B model, pre-trained on 253-frame clips, demonstrates stability in rollouts exceeding 4000 frames after fine-tuning on 381-frame clips using up to 36 GPU hours. This development represents a significant advancement in the field of long-horizon streaming world models.

AI 根据报道生成 · 2 小时前更新

还没有足够的连续观测数据，暂不绘制趋势。