Abstract

We propose a high-fidelity digital twin framework for preventive conservation of fragile heritage artifacts, addressing the critical challenge of balancing photorealism with computational efficiency in material-aware rendering. The proposed method integrates 3D Gaussian Splatting (3DGS) with a novel lightweight BRDF compression mechanism, replacing conventional high-dimensional reflectance data with compact analytic coefficients per Gaussian. At the core of our system is a neural compression module that maps complex material properties to low-dimensional representations using a hybrid basis of spherical harmonics and lightweight neural processes, enabling efficient storage and real-time rendering without sacrificing visual fidelity. The compression pipeline begins with an initial NeRF-based reconstruction to extract sparse 3D Gaussians and their BRDF estimates, then employs a transformer-encoder to predict compact coefficients dynamically adapted to heterogeneous materials such as patina or weathered stone. During rendering, the compressed BRDF is reconstructed on-the-fly, reducing memory usage by 10-20x while maintaining a PSNR above 35 dB for cultural heritage materials. Moreover, the framework supports conservation workflows by enabling degradation detection through temporal coefficient analysis and virtual restoration via coefficient manipulation. Key innovations include a hybrid basis selection for handling non-linear effects like subsurface scattering and a dynamic adaptation loop that ensures the digital twin evolves with the physical artifact. The method seamlessly integrates with existing NeRF-3DGS pipelines, requiring no modifications to core splatting routines, hence offering a practical solution for scalable heritage preservation.

Keywords