Prior-Infused 3D Gaussian Splatting for Robust Sparse-View Reconstruction

Abstract

Sparse-view 3D Gaussian Splatting (3DGS) is prone to severe geometric failures and appearance inconsistency due to optimization ambiguity. We introduce Prior-Infused 3DGS (PI-3DGS), a COLMAP-free framework that systematically integrates robust multi-view geometric priors into the 3DGS pipeline. Our framework consists of two core stages: (1) High-fidelity geometric initialization using dense geometry from a state-of-the-art multi-view model, VGGT, to provide precise Gaussian positions, normals, and disk-shaped scales; and (2) Prior-guided joint optimization that enforces geometric consistency through confidence-weighted depth alignment and disk-regularized normal supervision, alongside photometric constraints. By unifying precise initialization and persistent prior-guided optimization, our approach substantially mitigates floaters, enforces cross-view consistency, and delivers sharp, high-fidelity reconstructions under extreme sparsity. Extensive experiments on Mip-NeRF 360, ETH3D, and Replica demonstrate that PI-3DGS significantly outperforms existing methods in novel view synthesis.