AB-Attention accelerates video diffusion transformers
Bin Xu and eight other authors submitted “AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers” to arXiv on October 29, 2026. The training-free method combines block selection, inclusion-probability weighting and finer key/value gathering to approximate dense attention at high sparsity. On Wan2.1 and HunyuanVideo, the authors report skipping 89–92% of attention computation and achieving up to 2.23× end-to-end speedup.
AB-Attention targets the quadratic attention cost that dominates video diffusion transformer inference, with author-reported results indicating substantial computation savings on Wan2.1 and HunyuanVideo.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.39663 (cs)
[Submitted on 29 Oct 2026]
Title:AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers
Authors:Bin Xu, Mo Zhang, Zhengfeiyang Zhang, Xinfeng Yang, Yixing Ma, Wei Wang, HeXin Wang, Pengfei Zhou, Wangbo Zhao
View a PDF of the paper titled AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers, by Bin Xu and 8 other authors