DynamicRad: Content-Adaptive Sparse Attention for Long Video Diffusion

2026-04-22

AI Summary

This work addresses the challenge in long-video diffusion models where static sparse attention mechanisms struggle to balance computational efficiency with effective long-range dynamic modeling. To overcome this limitation, the authors propose DynamicRad, a content-adaptive sparse attention framework grounded in a radial locality prior. DynamicRad integrates dual modes—static sparsity ratios and dynamic thresholds—to achieve a Pareto-optimal trade-off between generation quality and inference efficiency without requiring online search. The method incorporates semantic motion routing, mask-aware LoRA fine-tuning, and MSE optimization via a physics-inspired proxy task, complemented by offline Bayesian hyperparameter tuning. Evaluated on HunyuanVideo and Wan2.1-14B, DynamicRad achieves 1.7–2.5× inference speedup with over 80% sparsity, while its dynamic mode matches or even surpasses dense baselines on certain long-sequence tasks.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

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 \textbf{DynamicRad}, a unified sparse-attention paradigm that grounds adaptive selection within a radial locality prior. DynamicRad introduces a \textbf{dual-mode} strategy: \textit{static-ratio} for speed-optimized execution and \textit{dynamic-threshold} for quality-first filtering. To ensure robustness without online search overhead, we integrate an offline Bayesian Optimization (BO) pipeline coupled with a \textbf{semantic motion router}. This lightweight projection module maps prompt embeddings to optimal sparsity regimes with \textbf{minimal runtime 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 \textbf{1.7$\times$--2.5$\times$ inference speedups} with \textbf{over 80\% effective 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.

Problem

• long video diffusion
• sparse attention
• spatiotemporal energy decay
• long-range information
• attention sparsity

Innovation

• sparse attention
• video diffusion
• content-adaptive
• Bayesian optimization
• radial locality