Comparison of AI-driven 3D Reconstruction Methods for Cultural Heritage Modeling

Abstract. Artificial intelligence-driven 3D reconstruction offers new opportunities for documenting cultural heritage under limited and low-cost image-acquisition conditions. This study compares Neural Radiance Fields (NeRF), 3D Gaussian Splatting (3DGS), and SAM 3D for reconstructing Saint Michael Church in Trabzon, Türkiye, using imagery representative of a short amateur UAV survey. From a UAV dataset of 178 images, 50 images were used to train NeRF and 3DGS, while SAM 3D was evaluated using four cardinal-view images. The methods were assessed through rendering speed, PSNR, SSIM, LPIPS, Chamfer mean, Chamfer RMSE, model size, and visual inspection against a photogrammetric reference model. 3DGS achieved the strongest overall performance among the scene-based methods, reaching 31.6 FPS, a PSNR of 29.71 dB, an SSIM of 0.920, and an LPIPS of 0.070, while also providing better average geometric proximity than NeRF. SAM 3D produced the lowest Chamfer mean and RMSE values despite using only four images. However, visual inspection revealed hallucination-like alterations in architecturally significant elements, including pillar dimensions, arch counts, and roof overhangs. The findings indicate that, in this case, low geometric distance does not necessarily imply documentary reliability. Under the tested configuration and with sufficient image overlap, 3DGS provides the most balanced solution for visually faithful and interactive heritage representation. SAM 3D remains promising for extremely sparse imagery, but its inferred geometry requires explicit verification. Future work may investigate multiview and video-based SAM 3D workflows and hallucination-resistant art-historical constraints.