提出一种无需训练的视频生成注意力机制，显著降低计算量。
Token Radius Attention for Efficient Video Generation
- 根据查询熵动态分配注意力范围，实现自适应稀疏化
- 仅保留9-19%的注意力交互，速度提升1.56至2.05倍
- 适用于高保真视频生成，适合资源受限场景
视频扩散变换器（VDiTs）虽能生成高质量视频，但其密集3D自注意力带来二次方复杂度。现有头级与块级稀疏方法将计算预算均分给所有查询，忽略查询间的注意力需求差异。我们发现，不同查询的保留密度与其注意力熵呈对数线性关系，且主要交互呈现以查询为中心、依赖于标记的邻域结构，具有可变半径。基于此，提出无需训练的令牌半径注意力（TRA），将查询熵映射为解析的令牌预算，并转换为随时间衰减的半径，无需显式键值排序。融合熵提取、预热复用和块稀疏掩码构建进一步降低开销。在七个Wan2.1、Wan2.2及HunyuanVideo文本到视频/图像到视频配置中，TRA仅保留9%-19%的注意力交互，实现1.56x-2.05x加速，同时保持竞争力生成质量。代码已开源。
原文摘要 · Abstract (English)
Video Diffusion Transformers (VDiTs) enable high-fidelity generation but incur quadratic cost from dense 3D self-attention. Existing head- and block-level sparse methods share computation budgets across queries, overlooking token-specific attention demand. We observe that retained density varies across queries yet correlates log-linearly with attention entropy, while dominant interactions form query-centered neighborhoods with token-dependent radii. Based on these findings, we propose Token Radius Attention (TRA), a training-free framework that maps query entropy to an analytic token budget and converts it into a temporally decayed radius without explicit key ranking. Fused entropy extraction, warm-up reuse, and block-sparse mask construction further reduce overhead. Across seven Wan2.1, Wan2.2, and HunyuanVideo T2V/I2V configurations, TRA retains only 9-19% of attention interactions and achieves 1.56x-2.05x speedup with competitive generation quality. Code is available at https://github.com/IF-LAB-PKU/Token-Radius-Attention.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。