Fast Video Generation with Sliding Tile Attention

TL;DR

This paper introduces sliding tile attention (STA), a hardware-efficient local attention mechanism that significantly accelerates 3D video diffusion models without quality loss, enabling faster video generation.

Contribution

The paper proposes STA, a novel local attention method with hardware-aware design, achieving substantial speedups in video diffusion models without retraining.

Findings

STA accelerates attention by up to 17x over FlashAttention-2.

End-to-end video generation latency reduced from 945s to 685s without quality loss.

Code released publicly for reproducibility.

Abstract

Diffusion Transformers (DiTs) with 3D full attention power state-of-the-art video generation, but suffer from prohibitive compute cost -- when generating just a 5-second 720P video, attention alone takes 800 out of 945 seconds of total inference time. This paper introduces sliding tile attention (STA) to address this challenge. STA leverages the observation that attention scores in pretrained video diffusion models predominantly concentrate within localized 3D windows. By sliding and attending over the local spatial-temporal region, STA eliminates redundancy from full attention. Unlike traditional token-wise sliding window attention (SWA), STA operates tile-by-tile with a novel hardware-aware sliding window design, preserving expressiveness while being hardware-efficient. With careful kernel-level optimizations, STA offers the first efficient 2D/3D sliding-window-like attention implementation, achieving 58.79% MFU. Precisely, STA accelerates attention by 2.8-17x over FlashAttention-2 (FA2) and 1.6-10x over FlashAttention-3 (FA3). On the leading video DiT, HunyuanVideo, STA reduces end-to-end latency from 945s (FA3) to 685s without quality degradation, requiring no training. Enabling finetuning further lowers latency to 268s with only a 0.09% drop on VBench. We make our codebase public at https://github.com/hao-ai-lab/FastVideo.