LiteAttention: A Temporal Sparse Attention for Diffusion Transformers

TL;DR

LiteAttention introduces a novel approach to accelerate diffusion transformers by exploiting temporal sparsity patterns, enabling efficient attention computation with no quality loss in video generation tasks.

Contribution

It leverages temporal coherence in diffusion attention to dynamically skip redundant computations, combining the benefits of static and dynamic sparsity methods.

Findings

Achieves significant speedups on production video diffusion models

Maintains high quality with no degradation in generated videos

Efficiently propagates skip decisions across denoising steps

Abstract

Diffusion Transformers, particularly for video generation, achieve remarkable quality but suffer from quadratic attention complexity, leading to prohibitive latency. Existing acceleration methods face a fundamental trade-off: dynamically estimating sparse attention patterns at each denoising step incurs high computational overhead and estimation errors, while static sparsity patterns remain fixed and often suboptimal throughout denoising. We identify a key structural property of diffusion attention, namely, its sparsity patterns exhibit strong temporal coherence across denoising steps. Tiles deemed non-essential at step $t$ typically remain so at step $t+\delta$. Leveraging this observation, we introduce LiteAttention, a method that exploits temporal coherence to enable evolutionary computation skips across the denoising sequence. By marking non-essential tiles early and propagating skip decisions forward, LiteAttention eliminates redundant attention computations without repeated profiling overheads, combining the adaptivity of dynamic methods with the efficiency of static ones. We implement a highly optimized LiteAttention kernel on top of FlashAttention and demonstrate substantial speedups on production video diffusion models, with no degradation in quality. The code and implementation details will be publicly released.