Augmented Flow Matching via Variance Reduction with Auxiliary Variables

Jongmin Yoon, Juho Lee

Keywords: generative modeling, flow matching

TL;DR: Dimension augmentation on training data reduces training variance and achieve fast and efficient sampling.

Abstract: Flow matching is a simulation-free approach that scalably generates an ODE, in which its path traverses between two different distributions. However, conventional flow matching relies on the training pairs drawn independently, inducing high variance that might slow down training process and degrade the performance upon training. To mitigate this, we propose augmented flow matching, a simple yet efficient framework that can be ubiquitously applied to flow matching with slight modification to the models. We first find that when some auxiliary variables that are correlated to the training data, then they contribute on variance reduction of the flow matching loss estimation, when used together with the training data pair. With this observation, we construct auxiliary variables that are correlated to the training pair, which is obtained by simple and effective linear operation from the input data. Finally, we show that with this simple modification on the training phase, we achieve the improved model flexibility and performance when the ODE is applied on the learned model.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4726