Diffusion models for image super-resolution: State-of-the-art and future directions

The single image super-resolution (SISR) task has received much attention due to the wide range of applications in many tasks. The progress in this SISR is mainly based on deep learning methods. In recent years, many methods have been developed based on using diffusion models to solve the SISR task. The performance of the diffusion-based model depends on many factors, such as the type of diffusion, noise schedule, and loss function. Based on this limitation, we developed a review paper about the state-of-the-art diffusion models, especially ones that were developed for SR. In this survey, the SR-based diffusion models are classified based on four metrics, namely the diffusion model type, the noise schedule, the loss function, and the used datasets. Then, we discuss the main categories in each metric and the developed diffusion models related to each category. Three main diffusion model types are considered, including denoising diffusion probabilistic models (DDPM), noise-conditioned score networks (NCSN), and stochastic differential equations models (SDEM), in addition to other diffusion models. In the end, we discuss the limitations of these diffusion models and show our expectations for future research in this domain.