Abstract
Image super-resolution (SR) reconstruction is a critical task aimed at enhancing low-quality images to obtain high-quality counterparts. Existing denoising diffusion models have demonstrated commendable performance in handling image SR reconstruction tasks; however, they often require thousands-or even more-diffusion sampling steps, significantly prolonging the training duration for the denoising diffusion model. Conversely, reducing the number of diffusion steps may lead to the loss of intricate texture features in the generated images, resulting in overly smooth outputs despite improving the training efficiency. To address these challenges, we introduce a novel diffusion model named RapidDiff. RapidDiff uses a state-of-the-art conditional noise predictor (CNP) to predict the noise distribution at a level that closely resembles the real noise properties, thereby reducing the problem of high-variance noise produced by U-Net decoders during the noise prediction stage. Additionally, RapidDiff enhances the efficiency of image SR reconstruction by focusing on the residuals between high-resolution (HR) and low-resolution (LR) images. Experimental analyses confirm that our proposed RapidDiff model achieves performance that is either superior or comparable to that of the most advanced models that are currently available, as demonstrated on both the ImageNet dataset and the Alsat-2b dataset.