AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers

Researchers present AB-Attention, an accurate block attention method for accelerating video diffusion transformers, addressing the quadratic cost of 3D full attention that dominates inference. The work targets a fidelity problem in dynamic block-sparse attention, which reduces cost by computing exact attention only on key/value blocks selected from current activations but whose accuracy degrades under that approximation. The source provides no measured speedup, benchmark figure, or named author for the method.

Video diffusion transformers generate high-quality videos, but the quadratic cost of 3D full attention dominates their inference. Dynamic block-sparse attention reduces this cost by computing exact attention only on key/value blocks selected from current activations. However, its fidelity degrades u