Representation by Design in Generation: Cross-View Class-Token Alignment in Diffusion Transformers
Published in arXiv preprint, 2026
We train diffusion transformers to learn stronger semantic representations alongside generation. Aligning class tokens across independently noised views improves both global and patch features: ImageNet class-token linear probing gains 9.4 percentage points and frozen-backbone VOC2012 segmentation gains 3.6 mIoU, with comparable ImageNet generation FID.
Recommended citation: Wu, X., Wang, Y., and Wei, C. (2026). "Representation by Design in Generation: Cross-View Class-Token Alignment in Diffusion Transformers." arXiv:2609.36348.
Download Paper