Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

This paper introduces a new approach to single image super-resolution (SR) using a generative adversarial network (GAN). The SRGAN is designed to recover finer texture details from heavily downsampled images. By using a GAN, the generated high-resolution images are more realistic compared to those produced by traditional SR methods, as they are closer to the natural image manifold. The network consists of a generator, which attempts to create high-resolution images, and a discriminator, which distinguishes between the generated images and real high-resolution images, leading to more visually appealing results.