Researchers prove closed-form formula for computing uncertainty in flow matching generative models without retraining

A mathematical identity shows that the uncertainty (posterior covariance) of flow matching models can be computed directly from the velocity field's divergence, with no model retraining or architectural changes needed.

The approach works post-hoc on any pre-trained flow matching model. For one-step generators like MeanFlow, it yields end-to-end generation uncertainty in a single forward pass, eliminating multi-step variance propagation required by prior methods.

Experiments on MNIST show the resulting per-pixel uncertainty maps concentrate on digit boundaries where variation is highest, and the uncertainty score tracks actual prediction error, all at roughly 10,000× less total compute than ensembling or Monte Carlo dropout.