AB-Attention: training-free sparse attention speeds video DiTs up to 2.23x
bdsqlsz · x · 2026-11-02
A multi-institution paper introduces AB-Attention (Accurate Block Attention), a training-free block-sparse attention method for accelerating 3D full attention in video diffusion transformers.
- Results: Skips 89–92% of video attention on Wan2.1 and HunyuanVideo, delivering up to 2.23x end-to-end speedup (measured on NVIDIA H20) and up to +4.4 dB PSNR over the strongest sparse baseline.
- Approach: Targets fidelity loss under aggressive sparsity caused by coarse block-importance scoring, pruned blocks vanishing from softmax, and whole-tile keep/drop decisions. Its Intra-Block Clustering Selector (IBCS) estimates block importance cheaply by clustering queries only within each query block, then computes exact attention on selected blocks with full-tile execution.
- Availability: No training required; code coming soon.
More from Multimodal
- The 'Rian Johnson app' generates strikingly cinematic AI photos — threepointone · 2026-11-02
- Seedance 2.5 short film: a bunker escape and a heartbreaking goodbye — SimplyAnnisa · 2026-11-02
- Speridlabs open-sources Idol: single image to full 3D Gaussian splat in under 2s — pbaylies · 2026-11-02
- How Claude Code made a 54-second 4K60 launch film from raw screen recordings — No_Coast6186 · 2026-11-02
- Semi-transparent Canny trick reduces reference bleeding in Ref2V MM-H3 video generation — Vortexneonlight · 2026-11-02
- Open-source AnyShot: a Claude agent acts as cinematographer, cuts 6-hour renders to 3 minutes — Scobleizer · 2026-11-02