Real-World Super-Resolution with Residual Consistency

Finding or collecting paired datasets for realworld super-resolution is a challenging process. Some studies have approached this problem with a GAN-based degradation generator trained using an unpaired dataset. However, this approach does not need real-world low-resolution images after degradation generator training. To benefit more from these images that contain important domain information, a method called Residual Consistency has been proposed. It is aimed to increase performance by directly incorporating these images into training using Residual Consistency. Experiments were conducted on two datasets used in similar studies and comparable results were obtained. Additionally, the evaluation metric was examined with sample visuals.