A mathematical identity involving the divergence of the velocity field calculates generative model uncertainty instantly without needing ensembles or retraining.
Generative models usually struggle to tell you how sure they are without running expensive secondary tests. This method uses the divergence of the velocity field to provide an exact posterior covariance in one step. It turns what used to be a computationally heavy guessing game into a simple, closed-form calculation. Users can now get reliable confidence scores for AI-generated images or data points for free. This makes flow-matching models much more reliable for high-stakes applications where knowing the error margin is mandatory.
Divergence is Uncertainty: A Closed-Form Posterior Covariance for Flow Matching
arXiv · 2605.00941
Flow matching has become a leading framework for generative modeling, but quantifying the uncertainty of its samples remains an open problem. Existing approaches retrain the model with auxiliary variance heads, maintain costly ensembles, or propagate approximate covariance through many integration steps, trading off training cost, inference cost, or accuracy. We show that none of these trade-offs is necessary. We prove that, for any pre-trained flow matching velocity field, the trace of the post