Toward clean and efficient 3D Gaussian representations

3D Gaussian splatting (3DGS) enables high-quality novel-view synthesis, but its photometric-only optimization often favors per-pixel color fitting over geometric regularity, producing unsupported outliers and unnecessarily dense Gaussian distributions. We introduce a geometry-aware optimization pipeline that requires no external supervision and yields a cleaner, more efficient Gaussian representation while preserving reconstruction fidelity. The pipeline consists of three modules: (1) initial point purification (IPP) radially stratifies the seed point cloud and uses k-NN density estimates to remove upper-tail outliers, producing a well-conditioned initialization; (2) anchor-constrained densification (ACD) regulates the displacement and radius of Gaussians spawned during splitting to suppress color-driven drift and uncontrolled growth; and (3) opacity-sparsity joint pruning (OSJP) periodically removes Gaussians that are jointly sparse in opacity and density, progressively tightening the distribution. Experiments show that our method reduces training time by 32.6%, lowers rank-weighted dispersion (RW-Disp) by 29.8%, decreases the 95th-percentile inter-point distance (P95) by 42.2%, and reduces symmetric Chamfer distance by 80.5%, while maintaining comparable reconstruction quality. The resulting Gaussian set is more compact and cleaner, improving downstream usability. The framework requires no pretrained models, preserves the original 3DGS renderer and photometric loss, and is simple to integrate, offering a concise and effective pathway for optimizing 3DGS under joint geometric cleanliness and efficiency objectives.