NanoGS: Training-Free Gaussian Splat Simplification

Abstract: 3D Gaussian Splat (3DGS) enables high-fidelity, real-time novel view synthesis by representing scenes with large sets of anisotropic primitives, but often requires millions of Splats, incurring significant storage and transmission costs. Most existing compression methods rely on GPU-intensive post-training optimization with calibrated images, limiting practical deployment. We introduce NanoGS, a training-free and lightweight framework for Gaussian Splat simplification. Instead of relying on image-based rendering supervision, NanoGS formulates simplification as local pairwise merging over a sparse spatial graph. The method approximates a pair of Gaussians with a single primitive using mass preserved moment matching and evaluates merge quality through a principled merge cost between the original mixture and its approximation. By restricting merge candidates to local neighborhoods and selecting compatible pairs efficiently, NanoGS produces compact Gaussian representations while preserving scene structure and appearance. NanoGS operates directly on existing Gaussian Splat models, runs efficiently on CPU, and preserves the standard 3DGS parameterization, enabling seamless integration with existing rendering pipelines. Experiments demonstrate that NanoGS substantially reduces primitive count while maintaining high rendering fidelity, providing an efficient and practical solution for Gaussian Splat simplification. Our project website is available at https://saliteta.github.io/NanoGS/.

• The paper introduces a training-free, CPU-efficient method for Gaussian Splat simplification that effectively reduces primitives while preserving photorealism.
• It employs a locality-sensitive k-nearest-neighbor graph and mass-preserved moment matching, achieving significant PSNR and SSIM improvements across benchmarks.
• The method enables scalable deployment on edge devices and diverse 3D content pipelines, eliminating the need for retraining, extra imagery, or calibration.

NanoGS: Training-Free Gaussian Splat Simplification

Motivation and Context

3D Gaussian Splatting (3DGS) has rapidly supplanted traditional volumetric radiance field techniques for real-time novel view synthesis. Its rasterization-friendly, explicit representation enables high-fidelity rendering with practical performance, supporting diverse applications from VR and cinematic production to 3D content generation and semantic distillation. However, the requirement for millions of Gaussian primitives to achieve photorealism introduces prohibitive storage, transmission, and rendering overhead, curtailing scalability in both deployment and distribution. Existing compression schemes, predominately focused on bit-level optimization or pruning guided by rendering supervision and still highly dependent on calibrated imagery, fail to deliver lightweight, agnostic, and fully representation-compatible compaction for arbitrary 3DGS assets generated or edited via alternative pipelines.

NanoGS directly addresses this critical bottleneck by introducing a training-free, CPU-efficient framework for splat simplification. Its pipeline operates exclusively on a given set of splats, requiring neither extra images nor GPU optimization, and preserves the standard parameterization for seamless renderer compatibility.

Figure 1: NanoGS reduces Gaussian Splat primitive count from the full model to increasingly compact representations, preserving visual fidelity without GPU-intensive optimization or calibration.

Methodology

NanoGS achieves simplification via progressive, locality-sensitive pairwise merging over a sparse $k$-nearest-neighbor graph. The process is structured as follows:

1. Sparse Merge Graph Construction: The initial splat set is organized into a KNN graph, restricting candidate merges to spatially proximate pairs. This reduces complexity from $\mathcal{O}(N^2)$ pairwise comparisons to $\mathcal{O}(kN)$.
2. Merge Cost Evaluation: Each edge in the merge graph is scored using a composite cost: a geometric term and an appearance term. The geometric discrepancy leverages the I-divergence between the unnormalized two-splat mixture and the moment-matched single Gaussian approximation, effectively quantifying spatial distortion. Appearance features (e.g., SH coefficients) are compared by the squared $\ell_2$ distance.
3. Mass-Preserved Moment Matching (MPMM): Merging is governed by mass-weighted moment matching—the merged primitive’s parameters are biased towards the more massive component, with covariance expanded to cover dispersion. Opacity is aggregated using the probabilistic union rule (Porter-Duff source-over) and appearance blended proportionally.
4. Progressive Batched Edge Collapse: Edges with minimal merge cost are greedily selected for parallel merging; neighborhoods are periodically refreshed to ensure the graph accurately reflects evolving spatial adjacency.

   Figure 2: NanoGS pipeline overview: local pairwise merge candidates, principled merge cost computation, and mass-preserved moment matching guide progressive simplification.

Experimental Results

NanoGS was evaluated across four major benchmarks: NeRF-Synthetic, Mip-NeRF 360, Tanks and Temples, and Deep Blending, covering both synthetic and real-world scenes with variable geometry and appearance complexity. Three compaction ratios ($\rho\in\{0.1, 0.01, 0.001\}$) were tested, consistently demonstrating NanoGS’s superiority in visual fidelity compared to LightGS, PUP3DGS, and GHAP baselines.

Quantitative Metrics:

• PSNR and SSIM trends: NanoGS achieves +2.40 dB, +4.84 dB, and +5.46 dB PSNR improvements at the three respective compression ratios, outperforming all baselines at every budget.
• Graceful degradation: NanoGS preserves contiguous surface coverage and avoids floater artifacts at extreme compaction, evidenced by higher SSIM under aggressive pruning.

  Figure 3: Qualitative comparison on NeRF Synthetic chair scene; NanoGS retains structure as compaction ratio decreases.

  

  Figure 4: Qualitative comparison on Mip-NeRF 360 data; NanoGS maintains scene coherence even at $\rho=0.001$.

Ablation Analysis:

The ablation study underscores the critical role of KNN-based candidate selection, opacity filtering, and I-divergence for robust merging—each component contributes measurably to final fidelity, especially under heavy simplification.

Implications and Future Directions

NanoGS unlocks practical deployment of 3DGS assets across edge devices, browsers, and large-scale distribution platforms by dramatically reducing primitive counts without the need for retraining or image calibration. Its CPU-friendly, training-free nature enables integration into content pipelines—including generative and edited 3DGS assets—where rendering supervision is unavailable.

Figure 5: Generated results and corresponding compression results: NanoGS is applicable to models produced via 2D-3D diffusion pipelines.

Figure 6: Compression results with multimodal inputs, demonstrating NanoGS's compatibility with diverse content sources.

Theoretical Contributions:

NanoGS leverages information-theoretic measures (I-divergence) to guide merge distortion, integrating spatial mass into moment matching for coherent simplification. This alignment with Gaussian mixture reduction literature enables principled, scalable hierarchy construction for splat sets.

Prospects:

Accelerating candidate graph construction and merging via GPU could further boost scalability. Jointly learned, view-dependent merge cost criteria may improve perceptual relevance. Extending NanoGS to dynamic scenes and temporally evolving splat fields represents a natural evolution.

Conclusion

NanoGS provides an accessible, principled method for training-free Gaussian Splat simplification, outperforming existing compaction schemes in both quantitative and qualitative fidelity across standard benchmarks and diverse compression ratios. It is agnostic to asset provenance, preserves renderer compatibility, and exhibits scalable performance, setting a new baseline for structural simplification of radiance field representations and paving the way for more flexible, efficient, and democratized deployment of photorealistic 3D content.