Abstract: The purpose of image super-resolution technology is to convert low-resolution (LR) images into high-resolution (HR) images without losing information. The realization of this technology on portraits has a wide range of application scenarios such as face recognition, face alignment, etc., but the traditional super-resolution method has a low degree of recovery on face images and is unstable. In this regard, we propose a SN-FSRGAN model. Face prior features were used to guide super-resolution; and spectral normalization was introduced to stabilize GAN-based super-resolution network training results. Experiments on the Helen and CelebA datasets show that the proposed method has achieved better results in terms of PSNR, SSIM and visual senses compared with models such as ESRGAN and FSRGAN.