RASR: Retinex-Guided Adaptive One-Step Diffusion for Low-Light Image Super-Resolution

Low-light image super-resolution aims to recover normal-light high-resolution images from dark low-resolution observations captured by image sensors, in which illumination attenuation, sensor noise, blur, and low resolution are entangled, making it more challenging than conventional super-resolution. Diffusion-based methods perform well on real-world super-resolution but usually require costly multi-step inference; recent one-step methods either rely on a globally fixed timestep that cannot adapt to per-sample degradation, or they directly encode the dark image into the latent space, coupling illumination bias with content degradation. To address these issues, we propose RASR, a Retinex-guided adaptive one-step diffusion framework for low-light super-resolution. We first decompose the observation into reflectance and illumination, and we use the reflectance as the content carrier to align it with the normal-light prior of the pretrained model. A latent-space teacher then constructs per-sample supervision from the low/high-quality latent discrepancy, while a lightweight student predicts the noise level solely from the Retinex representation, removing the dependence on high-quality references at inference. Finally, a single velocity-field integration on Stable Diffusion 3 yields the result, updating only low-rank adapters and lightweight modules during training. Extensive experiments on the RELLISUR benchmark show that RASR overall outperforms existing low-light and mainstream super-resolution methods in PSNR, SSIM, and LPIPS, with especially prominent gains in perceptual quality, and ablation studies validate the effectiveness of each key design.