The core of flow matching models is designing a probability path from noise to data. Current options are limited: optimal transport (OT) paths move samples along straight lines, and rectified flow also travels straight. Mathematically, these paths are the same thing—the motion of a free particle minimizing the kinetic-energy Lagrangian.
Du, Zhang, and Li observed that in classical mechanics this is just a special case of the principle of least action: the simplest scenario where the particle feels no force, so it moves uniformly in a straight line. But real particles experience forces, and their trajectories can be curved, oscillatory, or spiral.
Lagrangian Flow Matching generalizes the problem: minimize the action of a general Lagrangian subject to the continuity equation and endpoint conditions. Within this framework:
- Optimal transport paths are the special case of the kinetic-energy Lagrangian (free particle motion).
- Trigonometric variance-preserving diffusion paths are the special case of the harmonic oscillator Lagrangian (particle under a simple harmonic restoring force).
- Do the new paths actually outperform straight-line paths on practical high-dimensional generation tasks (e.g., images, protein structures)?
- What is the selection criterion among different Lagrangians—is there a universal principle telling you which physical model to use? The theoretical framework offers "possibilities" but no guidance on "which one to pick."
The elegance of this framework lies in the fact that the dynamic problem has an equivalent static optimal transport formulation—meaning the training objective remains simulation-free (no need to simulate the full path during training). More general Lagrangians generate new probability paths and velocity fields. Numerical experiments show that different Lagrangians indeed change the learned dynamics and are competitive in generation quality.
What remains unclear
References
1. Du, S., Zhang, J., & Li, Y. (2026). *Lagrangian Flow Matching: A Least-Action Framework for Principled Path Design*. arXiv:2605.15419 [cs.LG]. 2. Lipman, Y., et al. (2023). *Flow Matching for Generative Modeling*. ICLR. 3. Liu, X., et al. (2023). *Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow*. ICLR.