Enhanced Conditions Based Deep Convolutional Generative Adversarial Networks

Generative adversarial network (GAN) is a prevalent generative model. While it is effective, it has been shown to be very hard to train in practice. This work demonstrates how an improvement to the GAN framework can be used in a stable training, and in a conditional manner able to restrict their generation according to some alternate information such as a class label. Additionally, we explore different GAN structures, showing stable training method between images and attributes.

I. Introduction

Generative adversarial network (GAN) [1] is a prevalent generative model. Deep convolutional generative adversarial network (DCGAN) [3], based on traditional generative adversarial networks, introduces convolutional neural networks (CNN) into the training for unsupervised learning to improve the effect of generative networks. Conditional generative adversarial network (CGAN) [2] is a conditional model which adds condition extension into GAN.