Flow Matching for One-Step Sampling

Svetlana Pavlova, Ivan Oseledets, Gleb Ryzhakov

Keywords: Flow Matching, Generative Models, Ordinary Differential Equations, One-step generation

TL;DR: The paper proposes a Flow Matching-based approach, that eliminates ODE solvers during sampling, speeding up without sacrificing performance.

Abstract:

Flow-based generative models have rapidly advanced as a method for mapping simple distributions to complex ones for which the distribution function is unknown. By leveraging continuous-time stochastic processes, these models offer a powerful framework for density estimation, i.e. an algorithm that samples new points based only on existing samples. However, their requirement of solving ordinary differential equations (ODEs) during sampling process incurs substantial computational costs, particularly for large amount of data and numerous time points. This paper proposes a novel solution, which is based on a theoretical analysis of Flow Matching (FM), to overcome this bottleneck, namely, we developed an algorithm to find the point prototype for a given point from the target distribution. By eliminating the need for ODE solvers, our method significantly accelerates sampling while preserving model performance. Numerical experiments validate the proposed approach, demonstrating its efficiency.

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: 14213