Staged Progressive Single Image Generative Adversarial Network

This paper focuses on exploring unconditional generative model training methods for single image learning and proposes a new staged progressive single image generative adversarial network (SPSI-GAN). By conducting staged multiscale training on a single image, we gradually learn the feature block information of different areas within the image and pay more systematic attention to the overall structure of the image, the basic outline of the scene, and the details of the content. Experiments verified the application effect of this method in multiple image tasks and compared it with other mainstream methods. This method shows excellence in generated image quality and feature distribution and could improve training efficiency and performance metrics while maintaining high quality generated images.