DynamicRad: Content-Adaptive Sparse Attention for Long Video Diffusion

Leveraging the natural spatiotemporal energy decay in video diffusion offers a path to efficiency, yet relying solely on rigid static masks risks losing critical long-range information in complex dynamics. To address this issue, we propose $\mathbf{\text{DynamicRad}}$, a unified sparse-attention paradigm that grounds adaptive selection within a radial locality prior. DynamicRad introduces a $\mathbf{\text{dual-mode}}$ strategy: $\mathit{\text{static-ratio}}$ for speed-optimized execution and $\mathit{\text{dynamic-threshold}}$ for quality-first filtering. To ensure robustness without online search overhead, we integrate an offline Bayesian Optimization (BO) pipeline coupled with a $\mathbf{\text{semantic}}\mathbf{\text{motion}}\mathbf{\text{router}}$. This lightweight projection module maps prompt embeddings to optimal sparsity regimes with $\mathbf{\text{minimal}}\mathbf{\text{runtime}}\mathbf{\text{overhead}}$. Unlike online profiling methods, our offline BO optimizes attention reconstruction error (MSE) on a physics-based proxy task, ensuring rapid convergence. Experiments on HunyuanVideo and Wan2.1-14B demonstrate that DynamicRad pushes the efficiency–quality Pareto frontier, achieving $\mathbf{\text{1.7}}\times \mathbf{\text{–2.5}}\times \mathbf{\text{inference}}\mathbf{\text{speedups}}$ with $\mathbf{\text{over}}\mathbf{\text{80\%}}\mathbf{\text{effective}}\mathbf{\text{sparsity}}$. In some long-sequence settings, the dynamic mode even matches or exceeds the dense baseline, while mask-aware LoRA further improves long-horizon coherence. Code is available at https://github.com/Adamlong3/DynamicRad.