.png){kind=icon}
.png){kind=icon}
Return
A structure-aware diffusion framework for generalizable mural restoration
.png){kind=icon}
OA
AI
DOI：10.1038/s40494-026-02923-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Mural paintings are important cultural heritage resources, but their digital restoration is challenged by peeling, cracks, fading, blur, and the need to preserve structural coherence in high-resolution images. This paper proposes SAG-MR, a structure-aware diffusion framework for mural restoration. SAG-MR integrates a Structure-Aware Cluster-Centric Scanning Module (SCCSM), a Structure-Guided Feature Modulation (SGFM) module, and an overlap-aware reconstruction strategy to enhance global structural reasoning and local detail recovery. We construct HRM-1550, containing 1550 paired mural samples with reference images, degraded images, masks, sketches, and degradation labels. Averaged over peeling, cracks, fading, and blur, SAG-MR achieves 46.091 PSNR, 0.886 SSIM, 0.289 LPIPS, and 19.99 FID on HRM-1550, outperforming restoration-oriented and general-purpose generative baselines under controlled degradation settings. Qualitative results further indicate improved contour continuity and motif consistency.