Research map / 3D vision / Gaussian splatting
Optimized Minimal 3D Gaussian Splatting
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 OMG, a 3DGS compression framework that reduces Gaussian count via a local-distinctiveness importance score and compresses attributes with sub-vector quantization plus a tiny neural space feature, reaching ~4 MB on Mip-NeRF 360 at 600+ FPS.
Problem
Compressing 3D Gaussian Splatting scenes when the number of Gaussians is drastically reduced; existing compression methods rely on over a million Gaussians and degrade when attributes of sparse Gaussians are quantized.
Approach
Combine a local distinctiveness importance term (appearance-feature difference to Morton-order neighbors) with a sparse-Gaussian attribute representation (per-Gaussian features plus a tiny neural space feature) and sub-vector quantization, which uses small codebooks and avoids the storage cost of residual VQ.
Key results
- On Mip-NeRF 360, OMG-XS reaches 4.06 MB with 27.06 PSNR and 350 FPS (612 FPS on RTX 4090), roughly halving storage versus LocoGS-S (7.90 MB).
- OMG-XL achieves 27.34 PSNR and 0.819 SSIM at 6.82 MB, surpassing LocoGS-L in all metrics on Mip-NeRF 360.
- Applied to 3DGS-MCMC on Zip-NeRF, MCMC+OMG reduces size from 675 MB to 29.0 MB while improving PSNR from 26.37 to 26.69.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OMG-XS instead of LocoGS-S | Mip-NeRF 360 compression · Mip-NeRF 360 | Size (MB): 7.90 -> 4.06 | Helped |
| Use OMG-XS instead of LocoGS-S | Mip-NeRF 360 compression · Mip-NeRF 360 | PSNR: 27.04 -> 27.06 | No clear effect |
| Use OMG-XS instead of LocoGS-S | Mip-NeRF 360 rendering · Mip-NeRF 360 | FPS: 310 -> 350 | Helped |
| Apply OMG representation to 3DGS-MCMC | 3DGS-MCMC · Zip-NeRF | Size: 675 MB -> 29.0 MB | Helped |
| Apply OMG representation to 3DGS-MCMC | 3DGS-MCMC · Zip-NeRF | PSNR: 26.37 -> 26.69 | Helped |
| Replace SVQ with VQ (ablation) | OMG · Mip-NeRF 360 | PSNR: 27.06 -> 27.06; 27.21 -> 27.18; 27.34 -> 27.27 | Hurt |
| Replace SVQ with VQ (ablation) | OMG · Mip-NeRF 360 | CB Init (seconds): 4.8 -> 86.4; 7.6 -> 100.9; 7.9 -> 117.4 | Hurt |
| Replace SVQ with RVQ (ablation) | OMG · Mip-NeRF 360 | Size: 4.06 -> 4.19; 5.31 -> 5.52; 6.82 -> 7.10 | Hurt |
| Replace SVQ with RVQ (ablation) | OMG · Mip-NeRF 360 | PSNR: 27.06 -> 27.01; 27.21 -> 27.13; 27.34 -> 27.25 | Hurt |
| Use Morton-order neighbor approximation instead of KNN (ablation) | OMG-XS local-distinctiveness scoring · Mip-NeRF 360 | PSNR: 27.06 -> 27.07 | No clear effect |
| Use Morton-order neighbor approximation instead of KNN (ablation) | OMG-XS local-distinctiveness scoring · Mip-NeRF 360 | Time (ms): 4100 -> 28 | Helped |
| Add G-PCC position compression and Huffman encoding (ablation) | OMG variants · Mip-NeRF 360 | Size (MB): 5.82 -> 4.06; 6.83 -> 4.75; 7.66 -> 5.31; 9.47 -> 6.52; 9.89 -> 6.82 | Helped |
Surprising findings
- Naively combining Mini-Splatting with existing compressors (Compact-3DGS, LocoGS) degrades quality substantially, and Compact-3DGS training with Mini-Splatting pruning failed to converge.
Limitations
- Performance is bounded by the baselines (3DGS and Mini-Splatting), which fail to densify enough Gaussians on large-scale scenes such as Zip-NeRF.
- No error bars or statistical significance reported.
Experimental setup
Datasets: Mip-NeRF 360 Tanks&Temples Deep Blending Zip-NeRF
Benchmarks: Mip-NeRF 360 Tanks&Temples Deep Blending Zip-NeRF
Metrics: PSNR SSIM LPIPS Size (MB) FPS training time number of Gaussians
Baselines: 3DGS Scaffold-GS Mini-Splatting CompGS Compact-3DGS C3DGS LightGaussian EAGLES SOG HAC LocoGS
Builds on: 3D Gaussian Splatting Mini-Splatting LocoGS Product Quantization 3DGS-MCMC
Closest papers by meaning
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.94)
- Locality-aware Gaussian Compression for Fast and High-quality Rendering (ICLR 2025 · similarity 0.93)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.93)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.92)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
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.