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

The paper introduces Sol-Attn, a training-free sparse-attention method for diffusion transformers used in image and video generation. It performs on-the-fly block thresholding during a single online-softmax pass, reuses proxy scores, and approximates the contribution of blocks that are not fully computed. This avoids materializing a proxy-score map and enables dynamic, controllable computation budgets. The authors report end-to-end speedups of 2.1x for video generation and 2.3x for video editing while preserving visual quality. The abstract does not provide detailed benchmark settings, model names, or quality metrics.