Controllable Reference-Guided Diffusion With Local–Global Fusion for Real-World Remote Sensing Super Resolution

Super-resolution (SR) techniques enhance the spatial resolution of remote sensing images, supporting large-scale earth observation. While single-image SR methods improve low-resolution (LR) inputs, they often ignore valuable information from auxiliary data. Reference-based super-resolution (RefSR) leverages historical high-resolution (HR) images to guide the reconstruction of current LR observations, but existing methods struggle with cross-sensor resolution gaps and land cover changes, leading to under-generation or over-reliance on the reference. We propose CRefDiff, a controllable reference-guided diffusion model for real-world remote sensing image SR. To address the under-generation problem, CRefDiff leverages a powerful generative prior to produce accurate structures and textures. To mitigate over-reliance on the reference, we introduce a dual-branch fusion mechanism that adaptively fuse both local and global information from the reference image. The model also supports adjustable reference strength during inference for enhanced flexibility. In addition, a Better Start strategy significantly accelerates inference by reducing denoising steps. We further construct Real-RefRSSRD, a real-world RefSR dataset with HR NAIP and LR Sentinel-2 image pairs exhibiting diverse land cover and temporal variations. Experiments demonstrate that CRefDiff achieves state-of-the-art performance and improves downstream tasks such as scene classification and semantic segmentation. The dataset and code is publicly available athttps://github.com/wwangcece/CRefDiff