BayesianFlow

Pixel-wise uncertainty estimation in Flow Matching generative models via Last Layer Laplace Approximation

BayesianFlow extends BayesDiff from diffusion models to flow matching, producing pixel-wise uncertainty estimates alongside generated images.

Method

A Last Layer Laplace Approximation is integrated into the U-Net. Its predictive variance is propagated through the generative trajectory, yielding an interpretable confidence map for each generated sample.

Results

Experiments on MNIST and Fashion-MNIST compare uncertainty quality and computational efficiency across diffusion and flow-matching models. Flow matching achieves 5× faster generation than DDIM at comparable quality.