arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2510.00658cs.CVcs.AIcs.LG
收藏

对齐你的切线:通过流形对齐切线训练更优的一致性模型

Align Your Tangent: Training Better Consistency Models via Manifold-Aligned Tangents

Graduate School of AI, KAIST(人工智能研究生院,韩国科学技术院)

Beomsu Kim, Byunghee Cha, Jong Chul Ye

AI总结:

针对一致性模型训练耗时长、需大批量的问题,提出流形特征距离损失以生成流形对齐切线,大幅加速训练并支持小批量训练,样本质量更优。

摘要:

随着扩散模型和流匹配模型取得当前最优的生成性能,学界目前的研究兴趣转向在不牺牲样本质量的前提下缩短推理时间。一致性模型(Consistency Models, CMs)经训练后可在扩散或概率流常微分方程(probability flow ordinary differential equation, PF-ODE)轨迹上保持一致性,能够实现单步或两步流采样或扩散采样。然而,CMs通常需要超大批量的长时间训练才能获得具备竞争力的样本质量。本文研究了CMs在收敛附近的训练动态,发现CM切线——即CM输出更新方向——具有很强的振荡性,表现为它们沿数据流形平行移动,而非朝向流形推进。为缓解振荡切线问题,我们提出了一种名为流形特征距离(manifold feature distance, MFD)的新型损失函数,它能够生成指向数据流形的流形对齐切线。由此,我们提出的方法——命名为Align Your Tangent (AYT)——可以将CM训练速度提升数个数量级,甚至在学习感知图像块相似性指标(learned perceptual image patch similarity metric, LPIPS)上表现更优。此外,我们发现该损失函数支持在极小批量下训练且不会降低样本质量。代码链接:https://github.com/1202kbs/AYT

英文摘要:

With diffusion and flow matching models achieving state-of-the-art generating performance, the interest of the community now turned to reducing the inference time without sacrificing sample quality. Consistency Models (CMs), which are trained to be consistent on diffusion or probability flow ordinary differential equation (PF-ODE) trajectories, enable one or two-step flow or diffusion sampling. However, CMs typically require prolonged training with large batch sizes to obtain competitive sample quality. In this paper, we examine the training dynamics of CMs near convergence and discover that CM tangents -- CM output update directions -- are quite oscillatory, in the sense that they move parallel to the data manifold, not towards the manifold. To mitigate oscillatory tangents, we propose a new loss function, called the manifold feature distance (MFD), which provides manifold-aligned tangents that point toward the data manifold. Consequently, our method -- dubbed Align Your Tangent (AYT) -- can accelerate CM training by orders of magnitude and even out-perform the learned perceptual image patch similarity metric (LPIPS). Furthermore, we find that our loss enables training with extremely small batch sizes without compromising sample quality. Code: https://github.com/1202kbs/AYT

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑