DART-3DGS: Distance-Aware Replication and Age-Gated Densification for 3D Gaussian Splatting

3D Gaussian Splatting (3DGS) renders novel views in real time, but quality is limited by two coupled factors: a Structure-from-Motion (SfM) seed that is thin in scene peripheries and weakly textured regions, and an adaptive density-control rule that lets new Gaussians split repeatedly, inflating primitive count, memory, and training time. We address both with two backbone-agnostic add-ons that leave the rasterizer and pruning untouched. Distance-Aware Replication on Tangent-planes (DART) enriches the seed at initialization, replicating each SfM point on its tangent plane within a distance-aware radius, and adding coverage in the scene periphery, where the SfM cloud is typically sparsest, without depth or external cues. Age-Gated Eligibility (AGE) bounds the budget by letting a Gaussian split or clone only after reaching a minimum age, turning the splitting cascade into controlled refinement. Applied unchanged to vanilla 3DGS and anchor-based Scaffold-GS on three standard benchmarks, both add-ons improve both backbones: the Scaffold-GS variant is strongest on six of nine dataset–metric combinations and second on the rest, and under a second protocol leads all three Tanks & Temples metrics and both Deep Blending SSIM/LPIPS, with PSNR within 0.5% of the top result. Gains concentrate on sparsely seeded scenes, while AGE bounds the added cost, lowering primitive count, memory, and training time relative to DART alone at unchanged quality.