Transformer-based conditional generative adversarial networks for image generation

Abstract

The recent transformers shows competitive performance on computer vision tasks, such as classification, detection, and segmentation. Inspire by its success, in this paper, we explore its application at some more notoriously difficult vision tasks such conditional generative adversarial networks and propose a transformer based conditional generative adversarial networks for image generation. Our model employs the transformer architecture to develop a generator discriminator and residual network as discriminator. To improve the performance, a spectral normalization technique is used to normalize the weights of discriminator and the hinge loss are determined for model optimization. The experimental results on four public datasets shows that our approach is capable of producing the high quality images with good consistence and diversity and outperforms existing works.