Abstract

We propose Energy-based generator matching (EGM), a modality-agnosticapproach to train generative models from energy functions in the absence ofdata. Extending the recently proposed generator matching, EGM enables trainingof arbitrary continuous-time Markov processes, e.g., diffusion, flow, and jump,and can generate data from continuous, discrete, and a mixture of twomodalities. To this end, we propose estimating the generator matching lossusing self-normalized importance sampling with an additional bootstrappingtrick to reduce variance in the importance weight. We validate EGM on bothdiscrete and multimodal tasks up to 100 and 20 dimensions, respectively.

 

Quick Read (beta)