Engineers Propose Hybrid GAN-VAE Objectives to Improve Generative Models

Minimizing something like MMD against the real reference distribution (either against single sample or generated distribution) feels like a safer route somewhat for the goal of real/fake distinction and blending in at a distribution level