Autoregressive (AR) video generation can produce long-horizon sequences, but it comes at a steep computational price: every step demands a forward pass through a large diffusion transformer. Training-free caching has emerged as a way to cut this cost by reusing intermediate computation across steps, avoiding the need for additional training.

The new paper introduces Frequency-Step Cache, a method that extends this idea by explicitly separating synchronization and repair. Rather than treating all cached information uniformly, the approach distinguishes between these two functions, aiming to reduce redundant computation while preserving the quality of the generated video.

Because the method is training-free, it could in principle be dropped into existing autoregressive video models without retraining. The abstract does not yet report specific efficiency gains, but the conceptual split between synchronization and repair points toward a more principled way to make long-video synthesis practical.