One-step diffusion for real-world image super-resolution via degradation removal and text prompts

Abstract

Pre-trained Text-to-Image (T2I) diffusion models have shown remarkable progress in Real-world Image Super-Resolution (Real-ISR) by leveraging powerful latent space priors. However, these models typically require tens or even hundreds of diffusion steps for high-quality reconstruction, posing two critical challenges: (1) excessive computational overhead, hindering practical deployment; and (2) inherent stochasticity, leading to output uncertainty. To overcome these limitations, we propose a One-Step Diffusion framework for Real-ISR via Degradation Removal and Text Prompts (OSD-DRTP). Specifically, the proposed OSD-DRTP comprises two principal components: (1) a Degradation Removal Module (DRM), which eliminates complex real-world image degradations to restore fidelity; and (2) a Detail Enhancement Module (DEM), which integrates a fine-tuned diffusion model with text prompts from a large language model to enhance perceptual quality. In addition, we introduce Variational Score Distillation (VSD) in the latent space to ensure high-fidelity reconstruction across diverse degradation patterns. To further exploit the latent capacity of the VAE decoder, we employ a hybrid loss combining mean squared error (MSE) and perceptual loss (LPIPS), enabling accurate texture restoration without auxiliary modules. Extensive experiments demonstrate that the proposed OSD-DRTP outperforms state-of-the-art methods in both perceptual quality and computational efficiency.