Shuffling-SinGAN: Improvement on Generative Model from a Single Image

Abstract

Recently, SinGAN takes the use of GANs into a new realm – unconditional generation learned from a single natural image. Following the SinGAN architecture, we propose Shuffling- SinGAN, an efficient unconditional generative model that trained on a single natural image for general image manipulation. Our new network includes a pyramid of fully convolutional GANs, in which each layer is responsible for learning the patch distribution at a different scale of the image. We can generate new samples with variability through our network, which have the function of maintaining the texture and global structure of the original image. New random image generated by the model after multiple training is different from the original image in detail. Inspired by sinIR, we decided to add random pixel shuffling to the network. After experimentation, we found that the changed model generated more random new samples. Shuffling-SinGAN allows generating new samples of arbitrary size and aspect ratio, that have significant variability, yet maintain both the global structure and the fine textures of the training image. User tests confirm that the generated samples are commonly confused to be real images. With quantitative evaluation, we show that Shuffling-SinGAN has competitive performance on random image generation.