VCL-GAN: A Variational Contrastive Learning Generative Adversarial Network for Image Synthesis

Generative Adversarial Networks (GANs) have worked well for image generation, but recent works have shown that their generated images lack diversity. In response to this problem, we propose an image generation network based on contrastive learning (CL) and Autoencoder (AE). Firstly, we try to let the network learn to extract different types of features by Siamese network (SN), which is a classic of contrastive learning. Secondly, we perform variational processing on the resulting codes, so that the GANs can generate high-quality images of multiple types by the codes. Finally, extensive experiments and comparisons with the state-of-the-art are conducted on the Tsinghua dog dataset that the method has made progress in generating realistic images of different types.