Generative models that know when to doubt themselves

Generative models often produce low-quality or physically impossible outputs, yet existing confidence measures require running multiple ensembles or stochastic trajectories at k× cost. Flow Matching with Confidence (FMwC) injects input-dependent noise at selected layers and propagates variance through the network in closed form, yielding a confidence score during standard sampling. The score filters unrealistic images and unstable crystal structures, enables trajectory editing by rewinding to decision points, and concentrates ODE integration where the model is uncertain. The confidence correlates with velocity field divergence, offering interpretability and guiding where to focus model refinement.