SCALINGCACHE: Extreme Acceleration of DiTs Through Difference Scaling and Dynamic Interval Caching
Published in ICLR 2026, 2026
Accepted to ICLR 2026 (CCF-A class conference).
We propose SCALINGCACHE, a method that leverages the similarity of adjacent denoising time-step features in DiT models to achieve extreme end-to-end acceleration through difference scaling and dynamic interval caching. Our method achieves 2.5x end-to-end acceleration on Wan2.1/Hunyuan video generation models and 3.0x on the Flux image generation model, surpassing SOTA solutions like TaylorSeer/EasyCache under the same acceleration ratio.
Recommended citation: Lihui Gu, et al. (2026). "SCALINGCACHE: Extreme Acceleration of DiTs Through Difference Scaling and Dynamic Interval Caching." ICLR 2026.