Reproduce “AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bin Xu, Mo Zhang, Zhengfeiyang Zhang, Xinfeng Yang, Yixing Ma, Wei Wang, HeXin Wang, Pengfei Zhou and 1 more
AB-Attention reduces video-attention computation in text-to-video diffusion transformers by sampling small key/value blocks, weighting them to account for omitted blocks, and gathering them into GPU execution tiles. On HunyuanVideo-T2V-13B, it achieves 34.347 dB PSNR against dense attention and 2.23× end-to-end speedup. PSNR measures similarity to the dense reference, not video quality; results are H20-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.