Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification

2026-07-27

AI Summary

Diffusion Transformers for high-fidelity video generation are hindered by the computational cost of long-sequence attention, and existing training-free sparse attention methods struggle to balance efficiency and accuracy. This work proposes a training-free online sparse attention mechanism that unifies dynamic routing, sparse computation, and approximation correction within a single Softmax pass. It dynamically allocates block-level budgets and applies thresholding to identify critical blocks, while reusing block-wise proxy scores to efficiently approximate the contributions of unselected blocks—eliminating the need for explicit proxy graph construction. The method achieves substantial acceleration in both image and video generation tasks, delivering 2.1× and 2.3× end-to-end speedups for video generation and editing, respectively, without compromising visual quality.

Abstract

Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected key-value blocks, yet existing methods struggle to sparsify attention both efficiently and accurately for two reasons: (1) Rigid, unpredictable, and costly routing: selecting a fixed fraction of top-ranked blocks by proxy score imposes fixed budgets, whereas retaining blocks to reach a target cumulative proxy probability mass yields dynamic but potentially imbalanced budgets; both incur non-negligible overhead from computing and materializing proxy scores. (2) Lossy keep-or-drop sparsification: unselected blocks are discarded entirely, degrading accuracy under aggressive sparsity. These limitations motivate cheaper dynamic-budget routing while limiting accuracy degradation. In this paper, we introduce training-free Sol-Attn (Sparsifying online attention), which unifies dynamic routing, sparse computation, and approximation correction in a single online-softmax pass, achieving a better accuracy-efficiency trade-off in sparse attention. The core of Sol-Attn is on-the-fly block thresholding with proxy-score reuse, which selects critical blocks by comparing block proxy scores against a threshold during online softmax. This design enables dynamic yet controllable block budgets without materializing the proxy map, while directly reusing the proxy scores of unselected blocks to approximate their contribution. Experiments across image and video generation tasks show that Sol-Attn advances the quality-efficiency frontier of training-free sparse attention, delivering 2.1 times and 2.3 times end-to-end speedups for video generation and editing, respectively, while preserving visual quality.

Problem

• sparse attention
• video generation
• diffusion transformers
• inference acceleration
• attention sparsification

Innovation

• dynamic sparse attention
• training-free
• online softmax
• attention sparsification
• video generation