We conducted a study on combining Generative Adversarial Networks and learnt about the compression to create an advanced generative lossy compression system by utilizing KNN and GAN’s approach. Our study focused on examining factors such as normalization layers, generator and discriminator architectures, training strategies, and perceptual losses. Our system is capable of producing visually pleasing reconstructions that are similar to the original input, can operate at a wide range of bitrates, and can handle high-resolution images. We tested our system using various perceptual metrics and a user study, which showed that our approach was better than existing approach, even when using more than 2 x bitrate. In summary, our study bridged the gap between rate-distortion-perception theory and practice.

#