English

Time dependent loss reweighting for flow matching and diffusion models is theoretically justified

Machine Learning 2025-11-21 v1 Machine Learning

Abstract

This brief note clarifies that, in Generator Matching (which subsumes a large family of flow matching and diffusion models over continuous, manifold, and discrete spaces), both the Bregman divergence loss and the linear parameterization of the generator can depend on both the current state XtX_t and the time tt, and we show that the expectation over time in the loss can be taken with respect to a broad class of time distributions. We also show this for Edit Flows, which falls outside of Generator Matching. That the loss can depend on tt clarifies that time-dependent loss weighting schemes, often used in practice to stabilize training, are theoretically justified when the specific flow or diffusion scheme is a special case of Generator Matching (or Edit Flows). It also often simplifies the construction of X1X_1-predictor schemes, which are sometimes preferred for model-related reasons. We show examples that rely upon the dependence of linear parameterizations, and of the Bregman divergence loss, on tt and XtX_t.

Keywords

Cite

@article{arxiv.2511.16599,
  title  = {Time dependent loss reweighting for flow matching and diffusion models is theoretically justified},
  author = {Lukas Billera and Hedwig Nora Nordlinder and Ben Murrell},
  journal= {arXiv preprint arXiv:2511.16599},
  year   = {2025}
}

Comments

19 pages, 0 figures

R2 v1 2026-07-01T07:47:43.759Z