HRGAN: A Generative Adversarial Network Producing Higher-Resolution Images than Training Sets.

The generative adversarial network (GAN) has demonstrated superb performance in generating synthetic images in recent studies. However, in the conventional framework of GAN, the maximum resolution of generated images is limited to the resolution of real images that are used as the training set. In this paper, in order to address this limitation, we propose a novel GAN framework using a pre-trained network called evaluator. The proposed model, higher resolution GAN (HRGAN), employs additional up-sampling convolutional layers to generate higher resolution. Then, using the evaluator, an additional target for the training of the generator is introduced to calibrate the generated images to have realistic features. In experiments with the CIFAR-10 and CIFAR-100 datasets, HRGAN successfully generates images of 64 × 64 and 128 × 128 resolutions, while the training sets consist of images of 32 × 32 resolution. In addition, HRGAN outperforms other existing models in terms of the Inception score, one of the conventional methods to evaluate GANs. For instance, in the experiment with CIFAR-10, a HRGAN generating 128 × 128 resolution demonstrates an Inception score of 12.32, outperforming an existing model by 28.6%. Thus, the proposed HRGAN demonstrates the possibility of generating higher resolution than training images.