APLDiff: an adaptive perception-driven lightweight diffusion framework for digital mural inpainting

Abstract Mural degradation presents significant challenges to cultural heritage preservation. To address this, a hierarchical mural inpainting model, APLDiff, based on a lightweight diffusion model, is proposed. A physics-based degradation simulation is introduced, which simulates real damage patterns by modeling material aging and environmental factors, thereby enhancing the model’s generalization ability. An efficient diffusion network is constructed, with parameters reduced by 83% compare to the original Diffusion model, and an adaptive perception weight mechanism is incorporated to alleviate quality loss caused by model compression. The two-stage multi-scale sampling strategy allows for coarse structure restoration at low resolution, followed by high-fidelity detail enhancement in the latent space. These innovations provide a scientific foundation and practical solution for the digital inpainting of mural heritage, improving inference efficiency while maintaining visual authenticity.