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

Sol-Attn is a training-free dynamic sparse attention method that speeds up video generation inference.

It thresholds block scores during a single online-softmax pass, reuses proxy scores to approximate skipped blocks, and avoids expensive proxy-map materialization.

This reduces the attention bottleneck in diffusion transformers while preserving visual quality and improving efficiency-accuracy trade-offs.

Reported end-to-end gains reached 2.1x for video generation and 2.3x for video editing.