Lyapunov Guidance: Stabilizing Generative Flows with One-Line Code

Jingdong Zhang, Xinze Li, Yize Jiang, Luan Yang, Minkai Xu, Junhong Liu

Keywords: flow matching guidance, Lyapunov control, projection operation

TL;DR: All the guidance methods of flow matching can be regarded as Lyapunov control, and can be improved substantially by our one-line code of Lyapunov projection.

Abstract: Flow matching has recently emerged as a powerful approach to learning complex data distributions with excellent performance across diverse generative tasks, yet adapting pre-trained flow models to new tasks typically requires costly retraining. To mitigate this issue, post-training guidance methods were proposed as they are lightweight and user-friendly for downstream applications. However, existing guidance methods are unreliable since they usually rely on function approximations and lack structural guarantees of sampling stability. In this paper, we address this challenge by proposing a unified framework, LyaGuide (Lyapunov Guidance for flow matching), which reformulates the guidance in flow matching as a Lyapunov control problem. LyaGuide supports two modes depending on whether the Lyapunov function is a known priori: a model-driven mode for developer-oriented scenarios where the guidance distribution is explicitly specified, and a data-driven mode for user-oriented scenarios where pre-trained models can be adapted with downstream task-specific data. Furthermore, to enforce the stability, we introduce a pseudo projection operator with a closed-form expression that strictly satisfies the Lyapunov condition. Notably, LyaGuide is compatible with any guidance method and can be implemented with a single line of code. Experiments on synthetic datasets and image inverse problems demonstrate that our framework consistently improves sample quality and guidance fidelity while preserving efficiency, and it significantly enhances the performance of existing guidance methods.

Primary Area: generative models

Submission Number: 13480