VSA: Faster Video Diffusion with Trainable Sparse Attention

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Peiyuan, Chen, Yongqi, Huang, Haofeng, Lin, Will, Liu, Zhengzhong, Stoica, Ion, Xing, Eric, Zhang, Hao
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908615918485504
author Zhang, Peiyuan
Chen, Yongqi
Huang, Haofeng
Lin, Will
Liu, Zhengzhong
Stoica, Ion
Xing, Eric
Zhang, Hao
author_facet Zhang, Peiyuan
Chen, Yongqi
Huang, Haofeng
Lin, Will
Liu, Zhengzhong
Stoica, Ion
Xing, Eric
Zhang, Hao
contents Scaling video diffusion transformers (DiTs) is limited by their quadratic 3D attention, even though most of the attention mass concentrates on a small subset of positions. We turn this observation into VSA, a trainable, hardware-efficient sparse attention that replaces full attention at \emph{both} training and inference. In VSA, a lightweight coarse stage pools tokens into tiles and identifies high-weight \emph{critical tokens}; a fine stage computes token-level attention only inside those tiles subjecting to block computing layout to ensure hard efficiency. This leads to a single differentiable kernel that trains end-to-end, requires no post-hoc profiling, and sustains 85\% of FlashAttention3 MFU. We perform a large sweep of ablation studies and scaling-law experiments by pretraining DiTs from 60M to 1.4B parameters. VSA reaches a Pareto point that cuts training FLOPS by 2.53$\times$ with no drop in diffusion loss. Retrofitting the open-source Wan-2.1 model speeds up attention time by 6$\times$ and lowers end-to-end generation time from 31s to 18s with comparable quality. These results establish trainable sparse attention as a practical alternative to full attention and a key enabler for further scaling of video diffusion models. Code will be available at https://github.com/hao-ai-lab/FastVideo.
format Preprint
id arxiv_https___arxiv_org_abs_2505_13389
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle VSA: Faster Video Diffusion with Trainable Sparse Attention
Zhang, Peiyuan
Chen, Yongqi
Huang, Haofeng
Lin, Will
Liu, Zhengzhong
Stoica, Ion
Xing, Eric
Zhang, Hao
Computer Vision and Pattern Recognition
Scaling video diffusion transformers (DiTs) is limited by their quadratic 3D attention, even though most of the attention mass concentrates on a small subset of positions. We turn this observation into VSA, a trainable, hardware-efficient sparse attention that replaces full attention at \emph{both} training and inference. In VSA, a lightweight coarse stage pools tokens into tiles and identifies high-weight \emph{critical tokens}; a fine stage computes token-level attention only inside those tiles subjecting to block computing layout to ensure hard efficiency. This leads to a single differentiable kernel that trains end-to-end, requires no post-hoc profiling, and sustains 85\% of FlashAttention3 MFU. We perform a large sweep of ablation studies and scaling-law experiments by pretraining DiTs from 60M to 1.4B parameters. VSA reaches a Pareto point that cuts training FLOPS by 2.53$\times$ with no drop in diffusion loss. Retrofitting the open-source Wan-2.1 model speeds up attention time by 6$\times$ and lowers end-to-end generation time from 31s to 18s with comparable quality. These results establish trainable sparse attention as a practical alternative to full attention and a key enabler for further scaling of video diffusion models. Code will be available at https://github.com/hao-ai-lab/FastVideo.
title VSA: Faster Video Diffusion with Trainable Sparse Attention
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2505.13389