CS-IntroVAE: Cauchy-Schwarz Divergence-Based Introspective Variational Autoencoder

Abstract

Although generative models are still being developed, image reconstruction and generation tasks have evolved dramatically. Since the most popular generative models still have some limitations, it is still challenging. For example, while generative adversarial network (GAN) produces clear images, it is hard to train. The hybrid VAE-GAN incorporates the benefits of both, although it is computationally intensive and prone to drawbacks such as overfitting and gradient disappearance. A novel generative model called the Cauchy-Schwarz Divergence-based Introspective Variational Autoencoder (CS-IntroVAE) is based for this challenge. Extensive experiments show that our model has good performance on both tasks by employing mixed Gaussian distributions as prior distributions and Cauchy-Schwarz divergence as a measure of the distance between prior and posterior distributions. The source code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/CoderSnack/CS-IntroVAE</uri> .