All-in-one mural restoration with prompt-guided residual diffusion
Chao Jiang, Tiantian Ren, Zhengyun Cheng · npj Heritage Science · 2025
Ancient murals, as invaluable cultural heritage, are prone to degradation from natural erosion and human activities. Traditional manual restoration methods have inherent limitations, which in turn make virtual restoration a promising and innovative alternative. This paper thus proposes a diffusion-based virtual mural restoration method. To enable unified restoration of diverse degradation types, we first introduce a prompt-guided block. This block leverages the strong text feature extraction capability of pre-trained large language models to guide the extraction of mural image features. Secondly, we account for the semi-transparent nature of degradation patches. Damaged areas are not completely opaque, so we design a novel residual diffusion model. This model employs a prompt-guided UNet to predict semi-transparent residuals and time-dependent Gaussian noise. Our all-in-one model achieves the restoration of damaged murals across multiple dynasties, regions, and degradation types. Comprehensive experiments and ablation studies validate the method’s effectiveness, demonstrating that it achieves state-of-the-art performance and brings significant advancements to the field of ancient mural virtual restoration.