Title: Optimizing Source Distribution Dynamics: A Condition-Aware Approach to Flow Matching

Abstract:

Flow matching has recently gained traction as a compelling alternative to diffusion-based generative models, especially within the domain of text-to-image synthesis. Although the framework theoretically supports arbitrary source distributions, the majority of current methodologies continue to default to a standard Gaussian distribution—a convention inherited from diffusion models. Consequently, the source distribution is seldom treated as an optimization variable. In this study, we demonstrate that designing the source distribution in a principled manner is not only viable but also advantageous for modern text-to-image systems.

We introduce a method for learning a condition-dependent source distribution within the flow matching objective, thereby leveraging rich conditioning signals more effectively. Our analysis identifies critical failure modes associated with the direct integration of conditioning into the source, such as distributional collapse and training instability. We establish that maintaining appropriate variance regularization and ensuring directional alignment between the source and target distributions are essential for achieving stable and efficient learning. Furthermore, we investigate how the selection of the target representation space influences flow matching when employing structured sources, delineating the specific regimes where these designs yield the best results. Extensive evaluations across various text-to-image benchmarks reveal consistent and robust performance gains, including FID convergence speeds up to three times faster. These findings underscore the practical value of a carefully designed source distribution in conditional flow matching.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC