Adaptive Control for 3D Gaussian Splatting: A Systematic Regularization Framework

Regularization in 3D Gaussian Splatting (3D-GS) is often piecemeal, applying uniform penalties that fail to resolve the interdependent trade-offs between detail, smoothness, and stability. This paper moves beyond such ad-hoc solutions by introducing a systematic, context-aware regularization framework for 3D Half-Gaussian Splatting (3D-HGS). Our method acts as an adaptive control system, featuring three synergistic techniques that respond to local scene properties and temporal dynamics. First, we introduce an adaptive opacity consistency loss that uses a dynamic, view-dependent geometric proxy to suppress appearance artifacts on smooth surfaces while preserving sharp boundaries. Second, a selective normal smoothness loss leverages a high-performance CUDA KNN search to enforce geometric coherence exclusively within object interiors, critically protecting edge and corner details from over-smoothing. Finally, a novel EMA-based normal anchoring mechanism provides temporal stability, safeguarding learned geometry against parameter drift during the volatile densification and pruning stages. Our integrated framework establishes a new state-of-the-art. Applied to the strong 3D-HGS baseline, it yields remarkable average PSNR gains, including an exceptional +7.63dB on the challenging Deep Blending dataset. These modular yet synergistic techniques offer a new, principled paradigm for robust and high-fidelity primitive-based rendering. Our source code is available at https://github.com/Archaic-Atom/Adaptive-GS.