Activation-Norm Maximization to Accelerate Training in Flow-Matching Transformers
CVPR 2026 Findings Workshop
Yash Belhe1, Wesley Chang1, Tzu-Mao Li1, Ravi Ramamoorthi1, and Michaël Gharbi2
1 University of California San Diego
2 Reve
Description
We propose a simple (a few lines of code) and lightweight (< 0.1% of training steps) initialization technique that reduces FID in flow-matching diffusion transformers, resulting in a 5.8x speedup.
BibTeX
@inproceedings{Belhe2026ActivationMaximization,
author = {Belhe, Yash and Chang, Wesley and Li, Tzu-Mao and Ramamoorthi, Ravi and Gharbi, Micha\"el},
title = {Activation-Norm Maximization to Accelerate Training in Flow-Matching Transformers},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},
month = {June},
year = {2026},
pages = {4089-4096}
}