DOG: Discriminator-only Generation Beats GANs on Graphs

TL;DR — We investigate discriminator-only generation (DOG) using gradient descent on the input of a discriminator to generate samples and beat GANs on graphs.

Abstract

We propose discriminator-only generation (DOG) as a generative modeling approach that bridges the gap between energy-based models (EBMs) and generative adversarial networks (GANs). DOG generates samples through iterative gradient descent on a discriminator's input, eliminating the need for a separate generator model. This simplification obviates the extensive tuning of generator architectures required by GANs. In the graph domain, where GANs have lagged behind diffusion approaches in generation quality, DOG demonstrates significant improvements over GANs using the same discriminator architectures. Surprisingly, despite its computationally intensive iterative generation, DOG produces higher-quality samples than GANs on the QM9 molecule dataset in less training time.