Research map / 3D vision / Gaussian splatting
LOD-GS: Achieving Levels of Detail using Scalable Gaussian Soup
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes scalable Gaussian Soup, a triangle-primitive 3DGS representation with a triangle-MLP that decodes interior neural Gaussians, enabling LOD by downsampling Gaussians within triangles and reducing memory.
Problem
3D Gaussian Splatting methods use unstructured, redundant Gaussians with high memory usage and no structured way to provide levels of detail (LOD) under tight memory budgets.
Approach
Organize Gaussians inside discrete triangle primitives (triangle soup) with barycentric interpolation of vertex features, decode Gaussian attributes via a triangle-MLP conditioned on view direction, and produce LODs by downsampling Gaussians within each triangle, avoiding storage of explicit SH coefficients.
Key results
- On NeRF-Synthetic at LOD0, the method reaches 33.24 PSNR at 5.31M storage versus 33.88 PSNR at 71.6M for 3DGS and 33.36 PSNR at 18.2M for Scaffold-GS.
- At LOD3 on NeRF-Synthetic, the method achieves 31.08 PSNR at 1.4M storage, versus 29.91 PSNR at 1.9M for Scaffold-GS, and the best efficiency (21.11).
- On NSVF at LOD3, the method reaches 31.61 PSNR at 1.5MB, compared with 30.86 PSNR at 1.6MB for Scaffold-GS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use triangle primitives with triangle-MLP and downsample Gaussians within each primitive instead of unstructured 3DGS | LOD-GS · NeRF-Synthetic LOD3 | PSNR: 32.00 -> 31.08 | Hurt |
| use triangle primitives with triangle-MLP and downsample Gaussians within each primitive instead of unstructured 3DGS | LOD-GS · NSVF LOD3 | PSNR: 31.94 -> 31.61 | Hurt |
| use triangle primitives with triangle-MLP rather than voxel anchors | LOD-GS · NeRF-Synthetic LOD3 | PSNR: 29.91 -> 31.08 | Helped |
| use triangle primitives with triangle-MLP rather than voxel anchors | LOD-GS · NSVF LOD3 | PSNR: 30.86 -> 31.61 | Helped |
| use N=5 instead of N=10 interpolation samples per barycentric coordinate (ablation) | LOD-GS · NeRF-Synthetic | memory efficiency: 6.18 -> 4.72 | Hurt |
| use N=15 instead of N=10 interpolation samples per barycentric coordinate (ablation) | LOD-GS · NeRF-Synthetic | memory efficiency: 6.18 -> 5.70 | Hurt |
Surprising findings
- Full-size 3DGS achieves higher PSNR at each LOD but needs far more memory, so the PSNR-per-MB efficiency favors the proposed method.
Limitations
- Full-size PSNR is slightly lower than Scaffold-GS and 3DGS on both datasets.
- Evaluated only on two synthetic/object-centric datasets (NeRF-Synthetic with 8 scenes and NSVF with 4 scenes).
Experimental setup
Datasets: NeRF-Synthetic NSVF
Benchmarks: NeRF-Synthetic NSVF
Metrics: PSNR SSIM LPIPS PSNR per unit storage (efficiency)
Baselines: 3DGS Scaffold-GS
Builds on: 3D Gaussian Splatting Scaffold-GS triangle soup
Closest papers by meaning
- LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering (NeurIPS 2025 · similarity 0.93)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.91)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.90)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · similarity 0.90)
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.