Flow matching models can generate inconsistent or degraded samples, but they lack a built-in way to say how reliable a given output is. The authors propose UA-Flow, a lightweight extension that predicts the velocity field together with heteroscedastic uncertainty. By propagating that uncertainty through the flow dynamics, the model produces spatially localized reliability signals for each generated sample or region.
These uncertainty estimates are not just diagnostics. The authors use them for uncertainty-aware guidance (UG) and uncertainty-aware classifier-free guidance (UCFG), steering generation away from low-support regions. Experiments on ImageNet and CIFAR-10 show that UA-Flow's uncertainty tracks sample fidelity more consistently than element-wise baselines, and that the signals carry information about data-manifold support.
On ImageNet, UG increases precision at the cost of recall while improving FID. UCFG maintains lower FID than standard classifier-free guidance at large guidance scales. The results suggest that explicit uncertainty quantification can serve both as a quality signal and as a practical steering mechanism for flow-based generative models.