Calibrating Generative Models to Feature Distributions with MMD Finetuning

Nathaniel L. Diamant, Brian L. Trippe

Generative models can produce individually plausible samples while deviating substantially from a target set in the distribution of key features. For example, a model pretrained on…