SpecKCast: Accelerating Diffusion via Spectral Kalman Forecasting
October 18, 2026
SpecKCast accelerates Diffusion Transformers (DiTs) by using a Spectral Kalman filter to forecast features across skipped denoising steps. It addresses the mismatch in current methods by accounting for the non-uniform way spatial-frequency components change during the denoising process.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference costs for DiTs without significant quality loss by implementing spectral forecasting.
●
researcherThis provides a more mathematically grounded approach to feature caching and extrapolation.