SinGANxSR Super-Resolution Transfer with SinGAN

Data efficient machine learning models for perception tasks and generative modeling, have been a part of the human drive towards understanding the incredible ability of the human mind to make sense of real-life tasks with very little to no experience. SinGAN, is an extremely data efficient generative model that holds prior knowledge that has been learned only by viewing a single image. In this research, we explore the ability of SinGAN to generalize its super-resolution capabilities to unseen images, under the assumption that these images share semantic information with the single training image of the model.
Our contributions are 2-fold. Firstly, we perform an extensive research to learn, not only that this model has very impressive generalization capabilities, but also to test how far is the extent of this ability. Secondly, we introduce the Discriminator Super-Resolution Fitness (DSRF) metric - a built-in method that allows a pretrained model to indicate a-priorly if it can perform the SR task successfully on an unseen image.

two dog pictures

SinGANxSR Super-Resolution Transfer with SinGAN