ALLOT: training-free block-sparse attention allocation for video DiTs
A new arXiv preprint by Tianze Xia and five co-authors proposes ALLOT (Allocating Layer-step Latency by Output Tests), a training-free method for allocating block-sparse attention in video Diffusion Transformers. The authors observe that sparsifying attention in different layers and denoising steps affects final-video fidelity to very different degrees, while the most sensitive layer-step pairs stay nearly the same across inputs. ALLOT probes each layer-step cell and calibrates along the denoising trajectory, then jointly optimizes cell densities within a measured latency budget; the paper reports an 11.8 dB PSNR gain over XAttention at equal or shorter wall-clock time on HunyuanVideo, and wins or ties the default in all nine settings across Wan2.1, Wan2.2 and HunyuanVideo.
The authors say ALLOT keeps existing selection strategies and kernels and replaces only the allocation component, so it could be layered onto existing sparse-attention video generation pipelines without retraining.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.29441 (cs)
[Submitted on 21 Oct 2026]
Title:Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers
Authors:Tianze Xia, Xiaoling Zhou, Ruya Lan, Yiqiao Liao, Chengru Song, Shiming Xiang
View a PDF of the paper titled Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers, by Tianze Xia and 5 other authors