RAPID：面向高效视频扩散的跨步自适应注意力稀疏复用
RAPID: Reusing Attention Sparsity with Inter-step Adaptation for Efficient Video Diffusion
把视频扩散中 3D 注意力的稀疏模式当作可缓存的时序稳定信号：在生成早期做一次注意力块重要性估计并缓存稀疏掩码，后续步骤直接复用，从而免去逐步重算的开销；这套「一次性估计 + 跨步复用」思路可直接迁移到任何 DiT 类视频/图像生成模型的推理加速上。
原文依据 performs a one-shot attention block importance estimation early in the generation process
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论三维注意力的高昂开销阻碍了扩散模型生成高质量视频。现有稀疏注意力方法要么缺乏内容自适应性（静态），要么因逐步重算而带来过高开销（动态）。本工作挑战这一权衡的必要性，其依据是两点经验发现：(1) 视频扩散中的注意力模式表现出很强的时序稳定性；(2) 所需的计算密度逐步衰减。这一洞见促成了 RAPID，一个在生成过程早期执行一次性注意力块重要性估计的框架。所得分数与高保真稀疏掩码随后被缓存以便高效复用，从而消除重算开销。缓存的分数还支持一种可选的多阶段自适应剪枝（Turbo 模式），以实现最大加速。在 Wan2.1-14B 和 HunyuanVideo 等领先模型上，我们的高保真配置在受控计算预算下于各项关键质量指标（PSNR、SSIM、LPIPS）上超越所有基线。同时，其 Turbo 模式在 Wan2.1-14B 上取得最高 1.79 倍、在 HunyuanVideo 上最高 2.01 倍的加速，同时保持出色的视觉质量。
英文原文摘要 1180 字符
The prohibitive cost of 3D attention hinders high-quality video generation with diffusion models. Existing sparse attention methods either lack content adaptivity (static) or incur excessive overhead from per-step recalculation (dynamic). Our work challenges the necessity of this trade-off, based on a twofold empirical discovery: (1) attention patterns in video diffusion exhibit strong temporal stability, and (2) the requisite computational density progressively decays. This insight motivates RAPID, a framework that performs a one-shot attention block importance estimation early in the generation process. The resulting scores and high-fidelity sparse mask are then cached for efficient reuse, eliminating recalculation overhead. The cached scores also enable an optional, multi-stage adaptive pruning (Turbo mode) for maximum acceleration. On leading models like Wan2.1-14B and HunyuanVideo, our high-fidelity configuration surpasses all baselines across key quality metrics (PSNR, SSIM, LPIPS) under a controlled compute budget. Concurrently, its Turbo mode achieves speedups of up to 1.79x on Wan2.1-14B and 2.01x on HunyuanVideo while maintaining strong visual quality.