arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2409.13074cs.LGcs.CVstat.ML
收藏

引导机制的作用是什么?简单设定下的细粒度分析

What does guidance do? A fine-grained analysis in a simple setting

Duke University(杜克大学);MIT(麻省理工学院);Harvard University(哈佛大学);Johns Hopkins University(约翰斯·霍普金斯大学)

Muthu Chidambaram, Khashayar Gatmiry, Sitan Chen, Holden Lee, Jianfeng Lu

AI总结:

本研究严格证明了扩散模型中的引导机制无法从预期的倾斜分布中采样,并在混合紧支撑分布与混合高斯分布两种情形下,揭示了增大引导参数会使采样偏向条件分布支撑集边界甚至偏离支撑集,从而解释了大引导导致生成失真的现象,并为实际部署提供了理论指导。

摘要:

在扩散模型中使用引导机制的最初动机,基于这样一个前提:经引导修改后的分数,等于数据分布被提升至某次幂的条件似然所倾斜后的分布的分数。在本研究中,我们通过严格证明引导机制无法从预期的倾斜分布中进行采样,澄清了这一误解。我们的主要成果是在两种情形下对引导机制的动力学给出了细粒度刻画:(1)紧支撑分布的混合分布,以及(2)高斯分布的混合分布,这两种情形反映了引导机制在真实数据上表现出的显著特性。在这两种情形下,我们证明随着引导参数增大,引导模型会更多地从条件分布支撑集的边界进行采样。我们还证明,对于任意非零水平的分数估计误差,足够大的引导将导致采样偏离支撑集,这从理论上解释了“大引导会导致生成失真”这一经验发现。除了在合成设定中通过实验验证这些结果外,我们还展示了我们的理论见解如何为实际部署提供有用的指导原则。

英文摘要:

The use of guidance in diffusion models was originally motivated by the premise that the guidance-modified score is that of the data distribution tilted by a conditional likelihood raised to some power. In this work we clarify this misconception by rigorously proving that guidance fails to sample from the intended tilted distribution. Our main result is to give a fine-grained characterization of the dynamics of guidance in two cases, (1) mixtures of compactly supported distributions and (2) mixtures of Gaussians, which reflect salient properties of guidance that manifest on real-world data. In both cases, we prove that as the guidance parameter increases, the guided model samples more heavily from the boundary of the support of the conditional distribution. We also prove that for any nonzero level of score estimation error, sufficiently large guidance will result in sampling away from the support, theoretically justifying the empirical finding that large guidance results in distorted generations. In addition to verifying these results empirically in synthetic settings, we also show how our theoretical insights can offer useful prescriptions for practical deployment.

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。

↑