Adversarial Self Flow Matching: Few-steps Image Generation with Straight Flows

Hengyuan Cao, Yunhong Lu, Qichao Wang, Tiexiang Li, Xiaoyin Xu, Min Zhang

Keywords: generative models, flow matching, adversarial training

TL;DR: We propose Adversarial Self Flow Matching, which can straighten flows and align the generated data distribution with the real data distribution.

Abstract:

Flow Matching provides a method to train Ordinary Differential Equation (ODE)-based generative models and facilitates various probability path designs between initial and target distributions. Among these designs, straight flows are particularly interesting for reducing sampling steps. While some works have successfully straightened flows and achieved image generation in a few steps, they often suffer from cumulative errors or provide only piecewise or minibatch-level straightness. We propose Adversarial Self Flow Matching (ASFM), which can straighten flows and align the generated data distribution with the real data distribution. ASFM consists of two complementary components. Online Self Training straightens flows by constructing a conditional vector field using paired data, enabling one-step image generation during training. Adversarial Training aligns the one-step generated data with real data, thereby reducing cumulative errors when straightening flows. Experiments demonstrate that ASFM can build straight flows across the entire time span between two complete distributions and achieve highly competitive results across multiple datasets among Flow Matching-based methods. For instance, ASFM achieves 8.15 and 14.9 FID scores with NFE=6 on CelebA-HQ (256) and AFHQ-Cat (256), respectively.

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 7055