BayesianFlow
Pixel-wise uncertainty estimation in Flow Matching generative models via Last Layer Laplace Approximation
BayesianFlow carries the uncertainty estimation of BayesDiff from diffusion models over to flow matching, so each generated image comes with a pixel-wise uncertainty map.
Method
A last-layer Laplace approximation is fitted in the U-Net. Predictive variance is propagated along the generative trajectory, so each sample ends with a per-pixel variance estimate.
Results
MNIST and Fashion-MNIST experiments compare uncertainty quality and sampling cost between the diffusion and flow-matching versions. Flow matching generates about 5× faster than DDIM at comparable quality.