HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion

HyperVAttention (HVA) is a training-free sparse-attention framework for Video Diffusion Transformers. It combines 3D local-window clustering, a hybrid update schedule that performs full clustering only at anchor denoising steps, and hardware-aware cluster merging designed around GPU CTA-aligned execution costs. The paper argues that these components reduce clustering overhead, avoid redundant cluster updates, and improve sparse block density. On text-to-video generation experiments, HVA reportedly reduces end-to-end latency by up to 2.13x while achieving higher fidelity than existing training-free sparse-attention baselines. The supplied abstract does not specify models, datasets, hardware, or detailed quality metrics.