Contrastive Generative Adversarial Networks

Conditional image synthesis is the task to generate high-fidelity diverse images using class label information. Although many studies have shown realistic results, there is room for improvement if the number of classes increases. In this paper, we propose a novel conditional contrastive loss to maximize a lower bound on mutual information between samples from the same class. Our framework, called Contrastive Generative Adversarial Networks (ContraGAN), learns to synthesize images using class information and data-to-data relations of training examples. The discriminator in ContraGAN discriminates the authenticity of given samples and maximizes the mutual information between embeddings of real images from the same class. Simultaneously, the generator attempts to synthesize images to fool the discriminator and to maximize the mutual information of fake images from the same class prior. The experimental results show that ContraGAN is robust to network architecture selection and outperforms state-of-the-art-models by 3.7% and 11.2% on CIFAR10 and Tiny ImageNet datasets, respectively, without any data augmentation. For the fair comparison, we re-implement the nine state-of-the-art-approaches to test various methods under the same condition. The software package that can re-produce all experiments is available at https://github.com/POSTECH-CVLab/PyTorch-StudioGAN.