DIGU: NAS-Inspired Gaussian Distribution Search for 3D Gaussian Splatting Compression

Abstract

3D Gaussian Splatting (3DGS) enables high-quality real-time novel view synthesis, but high-fidelity models usually contain a large number of Gaussian primitives, leading to substantial storage and rendering overhead. Existing compression methods mainly rely on pruning, where each Gaussian is discretely retained or removed, which may discard useful representation capacity and cause reconstruction degradation under aggressive compression. In this paper, we propose DIGU, a NAS-inspired Gaussian Distribution Search framework for compact 3DGS representation. Instead of directly pruning Gaussians, DIGU learns a compact Gaussian distribution through differentiable assignment optimization. Specifically, we introduce feature-aware grouping to decompose the global search into local subproblems, and importance-aware budget allocation to adaptively assign output Gaussian numbers to different groups. A learnable assignment matrix is then optimized under rendering supervision to generate compact Gaussian primitives. Experiments on representative 3DGS scenes demonstrate that DIGU achieves competitive rendering quality under high compression ratios and provides an effective alternative to pruning-based 3DGS compression.