BayesianFlow: Uncertainty in Generative Models

This project introduces BayesianFlow, an approach for estimating pixel-wise uncertainty in generated images. It extends the principles of BayesDiff to the more efficient Flow Matching generative models. We leverage the Last Layer Laplace Approximation to quantify uncertainty during the generative process, providing interpretable uncertainty maps. Experiments on MNIST and Fashion-MNIST demonstrate the method’s effectiveness.

🔗 GitHub Repository: BayesianFlow on GitHub