3D Gaussian Pruning

3D Gaussian pruning studies how to remove or combine primitives in 3D Gaussian Splatting scenes while retaining accurate novel views and geometry. Methods score primitives using opacity, visibility, gradients, reconstruction sensitivity, geometric structure, or multi-view consistency; some predict importance without rendering, while others allocate pruning budgets by region or scene detail. Evaluations measure reconstruction quality, model size, memory, and speed on benchmarks such as Mip-NeRF 360, Tanks and Temples, and DTU, as well as in outdoor and SLAM settings. The central tension is that cheap, camera-independent scores and aggressive budgets can discard view-dependent detail, whereas more accurate rendering-aware selection costs computation and may not transfer across scenes or resolutions.