Face Super-Resolution Model Based on Diffusion Model

The problem of restoring high-resolution images from blurry images has long been a concern, and traditional methods of directly interpolating low-resolution images to obtain high-resolution images are simple but ineffective. Inspired by SR3, we propose a super-resolution model of human faces based on the diffusion model, which achieves super-resolution through a random iterative denoising process. In this paper, we have used a residual block that integrates multi-scale spatial attention and coordinate attention. Additionally, we have enhanced the restoration of image details through a global attention model. These advancements effectively address the discrepancy between automated evaluation metrics and human perception in high-frequency details for super-resolution models. Through evaluation of the standard eight-fold super-resolution task on CelebA-HQ, our model performs well and achieves competitive scores on SSIM and PSNR metrics.