SplatQ: Interaction-Aware Pruning of 3D Gaussian Splatting via Quadratic Selection

arXiv cs.GR details SplatQ, a pruning approach for 3D Gaussian Splatting that aims to do more than just keep the “most important” splats. The core idea is to avoid over-concentrating budget in one part of a scene, which can leave other regions underrepresented and hurt visual consistency.

SplatQ partitions Gaussians into spatially and appearance-similar blocks, smooths retention quotas across those blocks, and then models joint deletion effects with a quadratic objective. In practice, that means it considers not only each Gaussian’s individual contribution, but also how pairs of Gaussians interact when one or both are removed.

The optimization is solved with simulated annealing or a coherent Ising machine, which is an unusual but interesting angle for graphics work. The authors say the method improves aggregate reconstruction quality, spatial balance, and rendering speed across five benchmarks, with ablations pointing to quota smoothing and interaction modeling as the main wins.

For teams experimenting with 3DGS in production, the practical takeaway is that pruning can be treated as a structured optimization problem rather than a greedy score cutoff. That could matter for real-time rendering, scene compression, and any pipeline where quality cliffs in specific regions are more painful than a small global PSNR drop.