arXiv:2610.02260 releases MintFlow for minimal-perturbation constrained flow matching

The method formulates constraint satisfaction as a single minimal state intervention at an adaptively chosen timestep. Subsequent integration under the unchanged pretrained vector field then meets the target constraint while limiting distribution shift. Experiments on generative vision and physical system tasks show constraint satisfaction rates matching or exceeding prior samplers with lower distributional displacement measured by FID and Wasserstein distances.

Flow matching models, introduced in Lipman et al. 2022 and extended in subsequent works such as Liu et al. 2023, already support fast sampling. Existing constrained variants rely on optimization loops or classifier guidance that move samples far from the learned manifold. MintFlow replaces those loops with an explicit perturbation derived from the adjoint sensitivity of the flow ODE, eliminating per-sample solver iterations.

Operationally this enables plug-in use on deployed flow models in simulation and measurement-conditioned generation pipelines. Because the intervention time is selected by balancing perturbation size against remaining integration length, the approach scales to stiff physical constraints without manual schedule tuning. No training data or model weights are modified.

Next steps include integration into open-source flow matching libraries and benchmarking against emerging ODE-constrained baselines on larger-scale scientific datasets.