Optimized Minimal 3D Gaussian Splatting

TL;DR

This work addresses the storage and compute bottlenecks of 3D Gaussian Splatting by proposing Optimized Minimal Gaussian representation (OMG), which reduces the number of Gaussians and compresses their attributes without sacrificing fidelity. It achieves this via (i) a local distinctiveness-based importance metric, (ii) Sub-Vector Quantization (SVQ) for compact per-Gaussian attributes, and (iii) a lightweight neural field to capture coarse spatial features. Extensive experiments on standard 3DGS benchmarks show OMG reduces storage by nearly 50% and enables 600+ FPS rendering with competitive visual quality, across multiple datasets and variants. The approach also demonstrates strong ablations, showing SVQ efficacy and the importance of space features and LD scoring, with good generalization to related methods.

Abstract

3D Gaussian Splatting (3DGS) has emerged as a powerful representation for real-time, high-performance rendering, enabling a wide range of applications. However, representing 3D scenes with numerous explicit Gaussian primitives imposes significant storage and memory overhead. Recent studies have shown that high-quality rendering can be achieved with a substantially reduced number of Gaussians when represented with high-precision attributes. Nevertheless, existing 3DGS compression methods still rely on a relatively large number of Gaussians, focusing primarily on attribute compression. This is because a smaller set of Gaussians becomes increasingly sensitive to lossy attribute compression, leading to severe quality degradation. Since the number of Gaussians is directly tied to computational costs, it is essential to reduce the number of Gaussians effectively rather than only optimizing storage. In this paper, we propose Optimized Minimal Gaussians representation (OMG), which significantly reduces storage while using a minimal number of primitives. First, we determine the distinct Gaussian from the near ones, minimizing redundancy without sacrificing quality. Second, we propose a compact and precise attribute representation that efficiently captures both continuity and irregularity among primitives. Additionally, we propose a sub-vector quantization technique for improved irregularity representation, maintaining fast training with a negligible codebook size. Extensive experiments demonstrate that OMG reduces storage requirements by nearly 50% compared to the previous state-of-the-art and enables 600+ FPS rendering while maintaining high rendering quality. Our source code is available at https://maincold2.github.io/omg/.

Figures (5)

• Figure 1: Our approach focuses on minimizing storage requirements while using only a minimal number of Gaussian primitives. To achieve this, we introduce a local distinctiveness metric to identify the important Gaussians. Additionally, we propose a more efficient attribute representation, particularly for sparse Gaussians, by exploiting their irregularity and continuity. As a result, our method enables scene representation under 5 MB while achieving 600+ FPS rendering.
• Figure 2: The overall architecture of our proposed OMG. OMG learns per-Gaussian geometric and appearance features, applying Sub-Vector Quantization (SVQ) to all of them. The SVQ-applied geometric attributes are used for rendering, while the space feature based on the Gaussian center position is integrated into the appearance features to define the final appearance.
• Figure 3: Conceptual diagram of (a) vector quantization, (b) residual vector quantization, and (c) sub-vector quantization. + and $\oplus$ denote the element-wise summation and the vector concatenation.
• Figure 4: Qualitative results of OMG compared to 3DGS and LocoGS with COLMAP initialization. We provide per-image rendering PSNR with storage requirements for each scene.
• Figure 5: Evaluation without attribute compression.