Quantifying Epistemic Uncertainty in Diffusion Models

Researchers have introduced FLARE (Fisher-Laplace Randomized Estimator), a new scalable method that quantifies epistemic uncertainty in diffusion models like Stable Diffusion. Current methods often mix uncertainty types, leading to unreliable outputs. FLARE explicitly isolates epistemic variance using a random subset of parameters, producing more reliable plausibility scores. It improved uncertainty estimation in synthetic time-series tasks, achieving more accurate filtering than other methods, with theoretical bounds on its convergence rate.