Reproduce “Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for 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
Tianze Xia, Xiaoling Zhou, Ruya Lan, Yiqiao Liao, Chengru Song, Shiming Xiang
ALLOT allocates attention across transformer layers and denoising steps by probing sparse attention and calibrating one-step errors against final-video damage. It builds a density table under a latency budget while retaining the existing block-selection strategy and kernel. On XAttention/Wan2.1-T2V-1.3B, it achieves 26.39 dB PSNR versus 24.86 dB at 151.3 seconds and 1.21× speedup. PSNR compares with dense output, not absolute video quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.