中文VDE: 基于速度分解与估计的无训练整流流模型加速方法

整流流模型在图像、视频和3D生成中表现出色，但推理速度慢。现有加速方法重用缓存特征，但静态缓存与动态输入不匹配导致保真度下降。本文提出无训练加速方法VDE（速度分解与估计），从缓存重用转向分解估计，有效提升推理速度并保持输出质量。

arXiv:2605.23381v1 Announce Type: new Abstract: Though rectified flow models have achieved remarkable performance in image, video, and 3D generation, their practical deployments are challenged by slow inference speeds. Prior acceleration methods reuse cached features from previous steps, which neglects the growing mismatch between static caches and the evolving input, leading to reduced output fidelity. This work proposes Velocity Decomposition and Estimation (VDE), a training-free acceleration method that shifts the paradigm from caching-and-reusing to decomposing-and-estimating. Specifically