Advancing reference-based super-resolution via mitigating clarity inconsistent issue with foundation model guidance
Yong Shi, Jiayu Xue · International Journal of General Systems · 2025
Reference-based super-resolution (RefSR) has yielded impressive advancements in recovering high-frequency details of a low-resolution (LR) image with an additional high-resolution reference image. Existing RefSR methods easily result in the clarity inconsistent issue, i.e. some areas are clear while others are blurred. To investigate the underlying causes, we visualize the intermediate features and find two key reasons: the feature extractor lacks robustness in semantic feature extraction, and the use of pixel-wise loss leads to over-smoothing. To address these challenges, we propose a Geometric-Semantic Feature Adaption module to achieve more accurate feature matching results by leveraging the guidance from vision foundation models. Furthermore, we calculate the patch distribution loss between the output, ground-truth, and reference images to alleviate the feature misalignment issue and facilitate the transfer of clear textures from the reference image. Extensive experiments have demonstrated significant improvements over state-of-the-art methods in both quantitative and qualitative evaluations.