Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration

Jiaqi Han Juntong Shi Puheng Li Haotian Ye Qiushan Guo Stefano Ermon · 2026

Read Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration by Jiaqi Han Juntong Shi Puheng Li Haotian Ye Qiushan Guo Stefano Ermon (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-01623.txt.

Description

Diffusion models make fine images and video, but each output takes many slow passes through a large network. Jiaqi Han, Juntong Shi, Puheng Li, and colleagues cut that cost in this 2026 work by forecasting hidden features. Past tricks reuse old features for a few steps, yet errors grow fast when skips stretch far.

Their method, Spectrum, views hidden features as curves over time and fits them with smooth basis curves. A small regression sets the curve weights, which then forecast features many steps ahead at once. Theory gives an error bound that does not balloon as the skip grows. On top image and video models, the method runs several times faster than normal. Sample quality stays well above rival speed-up tricks at the same pace.

Catalog record

CREATORJiaqi Han Juntong Shi Puheng Li Haotian Ye Qiushan Guo Stefano Ermon
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58084
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.01623

Browse the full library · About the Tower of Babel