Taming Outlier Tokens in Diffusion Transformers

Published in arXiv preprint, 2026

We show that high-norm outlier tokens appear in both pretrained vision encoders and diffusion transformers, where they degrade local patch semantics. Dual-Stage Registers address both sources with lightweight interventions that improve generation quality.

Recommended citation: Wu, X.*, Wang, Y.*, Fu, T.-J., et al. (2026). "Taming Outlier Tokens in Diffusion Transformers." arXiv:2605.05206.
Download Paper