DiffuMural: a diffusion model for dunhuang murals restoration via multi-scale convergence and cooperative guidance

Ancient mural restoration is a challenging conditional image-generation problem because of large missing regions, semantic ambiguity, and strict demands for cultural and stylistic fidelity. We present DiffuMural, a multi-stage, condition-guided diffusion framework for high-fidelity mural inpainting. DiffuMural integrates structure-aware conditional modeling for spatial supervision, multi-scale collaborative diffusion for progressive, context-adaptive synthesis, and frequency-domain refinement to recover fine texture and chromatic detail. Trained end-to-end with multimodal spatial and structural conditioning, it leverages a dataset from 25 Tang-dynasty caves in Dunhuang’s Mogao Grottoes and produces 650,000 multi-scale samples. Evaluation combines standard metrics (PSNR, SSIM, FID, LPIPS), bespoke consistency indicators, and an expert humanistic assessment. DiffuMural outperforms the strongest baseline, yielding a PSNR gain of + 2.3 dB, SSIM + 0.013, FID − 2.5, and LPIPS − 0.035, while achieving a humanistic score of 96.7. The framework provides a scalable, deployable solution for AI-driven mural restoration in resource-constrained cultural heritage settings.