Digital Restoration: DiffuMural Sets New Standards in the Reconstruction of Dunhuang Murals

The restoration of ancient murals presents a complex challenge, especially when large parts of the artworks are damaged and only limited reference materials are available. The Dunhuang Grottoes in China, famous for their magnificent Buddhist murals, are a striking example of this restoration need. A new method called DiffuMural now promises to make a significant contribution to the digital reconstruction of these valuable cultural assets using Artificial Intelligence (AI).

DiffuMural is based on so-called diffusion models, a class of AI models that have achieved impressive results in the field of image generation in recent years. In contrast to conventional methods, which often have difficulty filling large gaps coherently while maintaining the style of the original, DiffuMural offers an innovative approach. By combining multi-scale convergence and collaborative diffusion with ControlNet and cyclic consistency loss, DiffuMural optimizes the match between the generated images and the given template. ControlNet allows precise control of the generation process, while the cyclic consistency loss ensures that the restored areas are seamlessly integrated into the overall image.

DiffuMural was trained with data from 23 large-scale Dunhuang murals that exhibit a consistent visual aesthetic. The model is characterized by its ability to reconstruct fine details, create a coherent overall image, and address the specific challenges of incomplete murals without a factual basis. The results are impressive: DiffuMural not only restores missing image areas, but also preserves the artistic style and cultural significance of the originals.

To objectively evaluate the quality of the restoration, a special evaluation framework was developed that considers four key metrics: factual accuracy, texture details, contextual semantics, and holistic visual coherence. In addition, humanistic value assessments were also integrated to ensure that the restored murals retain their cultural and artistic significance. Detailed experiments have shown that DiffuMural surpasses the current state of the art in both qualitative and quantitative metrics.

The development of DiffuMural represents a significant advance in the field of digital restoration. By combining state-of-the-art AI technology with a deep understanding of the artistic and cultural aspects of the Dunhuang murals, DiffuMural opens up new possibilities for the preservation and research of this unique cultural heritage. The technology could also be applied to other areas of restoration in the future, thus contributing to securing damaged works of art for future generations.

The application of AI in art restoration also raises ethical questions. It is important that the technology is used responsibly and that the authenticity of the originals is respected. DiffuMural is understood as a tool that supports restorers in their work, and not as a replacement for human expertise and artistic judgment.

Han, P., Kang, J., Pan, Y., Pan, E., Zhang, Z., Jin, Q., Jiang, J., Liu, Z., & Gong, L. (2025). DiffuMural: Restoring Dunhuang Murals with Multi-scale Diffusion. arXiv preprint arXiv:2504.09513. Wang, X., Sun, J., & Duan, Y. (2024). Dunhuang Murals Image Restoration Method Based on Generative Adversarial Network. Applied Sciences, 15(3), 1422. Pan, Y., Han, P., Zhang, Z., Kang, J., Pan, E., Jin, Q., ... & Gong, L. (2025). A comprehensive dataset for digital restoration of Dunhuang murals. arXiv preprint arXiv:2504.09514.