The MintFlow framework enforces constraints on flow matching models by applying minimal perturbations to intermediate flow states. This training-free approach prevents samples from drifting too far from the pretrained data distribution. It allows generative models to satisfy physical laws or measurements without sacrificing sample quality.