Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification
Sol-Attn reports 2.1x faster video generation inference and 2.3x faster video editing inference while preserving visual quality.
The paper targets diffusion transformers, where long video token sequences make attention the main inference bottleneck. Its method sparsifies attention during the online softmax pass, selecting key-value blocks by thresholding proxy scores instead of building a full proxy map. For blocks it skips, Sol-Attn reuses proxy scores to approximate their contribution rather than dropping them outright. HF Daily Papers' note
The paper targets diffusion transformers, where long video token sequences make attention the main inference bottleneck. Its method sparsifies attention during the online softmax pass, selecting key-value blocks by thresholding proxy scores instead of building a full proxy map. For blocks it skips, Sol-Attn reuses proxy scores to approximate their contribution rather than dropping them outright. HF Daily Papers' note
score 5