Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers

ALLOT is a training-free method for allocating block-sparse attention densities across layer-step pairs in video Diffusion Transformers. It estimates each cell's contribution to final-video error via cell probing and trajectory calibration, then optimizes densities under a measured-latency budget while keeping existing selection strategies and kernels. The paper reports an 11.8 dB PSNR gain over XAttention's default on HunyuanVideo at equal or shorter wall-clock time, and says it outperforms or ties defaults in all nine settings across four additional pipelines on Wan2.1, Wan2.2, and HunyuanVideo.

Why it matters

Attention is a major cost in video generation, so a training-free way to spend sparsity where it hurts output fidelity least could make these models cheaper to run without retraining.

Read with care

Source claims not independently checked; the reported gains come from the authors' own experiments on a specific set of pipelines.