On Distilling Generator Matching Models

Generator Matching (GM) is a new framework which encompasses the current workhorse generative modeling methods. However GM suffers from the computationally intensive sampling process common to these ODE/SDE based models. We introduce "Implicit Generator Matching" (IGM), a general framework for one-step distillation of generator matching models. Our method generalizes the recently proposed one-step diffusion distillation \citep{zhou2024score,luo2024one} methods to Generator Matching. We present promising initial results on image generation.