M3SFormer: multi-stage semantic and style-fused transformer for mural image inpainting

Digital restoration of ancient murals is crucial for preserving cultural heritage. However, existing methods often suffer from semantic distortion and stylistic inconsistencies when repairing large damaged areas. This paper proposes M3SFormer, an innovative restoration framework. Built on an enhanced P-VQVAE module, it employs continuous feature modeling without quantization to retain details. A new Semantic-Style Consistency Module (SSCM) integrates regional semantics with multi-scale style features, ensuring coherent outputs. Furthermore, the Flow-Guided Refinement Module (FGRM) reconstructs key textures through network guidance, improving visual quality. Experiments on multiple benchmarks show that M3SFormer surpasses mainstream methods across all metrics, with significant gains in PSNR, SSIM, and LPIPS. It also excels in reconstructing complex structures and preserving styles under high mask coverage, offering a reliable solution for high-quality digital mural preservation. The dataset and code are available at: https://github.com/LPDLG/M3SFormer