Generative models with kernel distance in data space

TL;DR

This paper introduces the LCW generator, a new generative model that uses kernel distance instead of adversarial training, aiming to combine the strengths of autoencoders and GANs while avoiding their main weaknesses.

Contribution

The paper proposes the LCW generator, a novel non-adversarial generative model that employs kernel distance and a two-phase training process to improve data generation quality.

Findings

Achieves competitive FID scores.

Avoids mode collapse and training instability.

Utilizes a kernel-based approach instead of adversarial training.

Abstract

Generative models dealing with modeling a~joint data distribution are generally either autoencoder or GAN based. Both have their pros and cons, generating blurry images or being unstable in training or prone to mode collapse phenomenon, respectively. The objective of this paper is to construct a~model situated between above architectures, one that does not inherit their main weaknesses. The proposed LCW generator (Latent Cramer-Wold generator) resembles a classical GAN in transforming Gaussian noise into data space. What is of utmost importance, instead of a~discriminator, LCW generator uses kernel distance. No adversarial training is utilized, hence the name generator. It is trained in two phases. First, an autoencoder based architecture, using kernel measures, is built to model a manifold of data. We propose a Latent Trick mapping a Gaussian to latent in order to get the final model. This results in very competitive FID values.