AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers
The authors report up to 2.23× faster video generation while skipping 89–92% of attention computation.
AB-Attention is a training-free method tested on Wan2.1 and HunyuanVideo. It estimates block importance through local query clustering, then samples and reweights blocks to approximate dense attention. Its kernel packs smaller key/value blocks into full tensor-core tiles, allowing finer selection without extra tensor-core cost. In the reported tests, baselines compute 1.8–3.6× as much attention, while AB-Attention improves PSNR by up to 4.4 dB over the strongest baseline.
HF Daily Papers' note
AB-Attention is a training-free method tested on Wan2.1 and HunyuanVideo. It estimates block importance through local query clustering, then samples and reweights blocks to approximate dense attention. Its kernel packs smaller key/value blocks into full tensor-core tiles, allowing finer selection without extra tensor-core cost. In the reported tests, baselines compute 1.8–3.6× as much attention, while AB-Attention improves PSNR by up to 4.4 dB over the strongest baseline.
HF Daily Papers' note
score 5