McGan: Mean and Covariance Feature Matching GAN

TL;DR

McGan introduces novel IPMs based on mean and covariance matching in feature space, enabling more stable GAN training by minimizing a meaningful distribution loss.

Contribution

The paper proposes new IPMs for GAN training that match mean and covariance statistics in feature space, improving stability and performance.

Findings

Enables stable GAN training with new IPMs.

Effectively matches distribution statistics in feature space.

Results show improved training stability and quality.

Abstract

We introduce new families of Integral Probability Metrics (IPM) for training Generative Adversarial Networks (GAN). Our IPMs are based on matching statistics of distributions embedded in a finite dimensional feature space. Mean and covariance feature matching IPMs allow for stable training of GANs, which we will call McGan. McGan minimizes a meaningful loss between distributions.