Image Restoration of Latent Diffusion Models Based on Compact A Priori Representation

Image restoration is the process of reconstructing the undesired part of an image by algorithm, which is one of the research hotspots in the field of computer vision. In this paper, for the problems of insufficient performance of denoising network and insufficient estimation of a priori information of image in Diffusion model for image restoration (DiffIR) model, the u-net network is improved to NAFNet and FreeU module is added. The performance of the denoising network in the diffusion model is effectively improved. Experimental results on Places and CelebA-HQ datasets show that compared with other models, the improved model in this paper leads about 0.5% and 3.1% in the evaluation metrics of LPIPS and FID, respectively.