A Comparative study on variational autoencoders and generative adversarial networks

Abstract

Generative Adversarial Networks (GAN) have been remarkable at generating artificial data, especially realistic looking images. This learning framework has proven itself to be effective in synthetic image generation, semantic image hole filling, semantic image editing, style transfer and many more. On the other hand, variational auto-encoders (VAE) have also been quite effective, so much so that mathematically it is often more accurate at generating images resembling to its original dataset. Nevertheless, images generated by VAE suffer from blurriness and are generally less realistic looking from human perception. In this paper we take a broad view on both systems and propose a theoretical approach to combine them and bring out the best of both.