Research map / 3D vision / Gaussian splatting
Locality-aware Gaussian Compression for Fast and High-quality Rendering
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 LocoGS, a locality-aware 3DGS compression framework that stores locally coherent Gaussian attributes in a hash-grid neural field while keeping the standard 3DGS rendering pipeline, achieving 54.6-96.6x storage reduction and 2.1-2.4x faster rendering than 3DGS.
Problem
3D Gaussian Splatting requires large storage (often over 1GB per scene), and existing compression methods either degrade quality under aggressive quantization (global statistics) or slow rendering (anchor-based neural decoding).
Approach
Analyze local coherence of 3D Gaussian attributes and encode the implicit attributes (opacity, normalized scale, rotation, residual SH) in a multi-resolution hash grid queried by Gaussian position, while keeping base scale and base color explicit; decompress back to a standard explicit 3DGS representation so rendering is unchanged, combined with dense Nerfacto-based initialization, soft-movement p
Key results
- On Mip-NeRF 360, Ours reaches 27.33 PSNR at 13.89 MB and 270 FPS versus 3DGS at 27.44 PSNR, 822.6 MB and 127 FPS.
- Ours-Small reaches 7.90 MB on Mip-NeRF 360 at 310 FPS, the smallest size among compared methods; HAC is 16.95 MB at 110 FPS.
- Ablation shows the full pipeline gives a 3.18x compression ratio over the 3DGS + pruning + compression variant with minimal quality loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LocoGS instead of compact baselines | Mip-NeRF 360 · Mip-NeRF 360 | FPS: 110 -> 270 | Helped |
| use LocoGS instead of compact baselines | Tanks and Temples · Tanks and Temples | PSNR: 24.08 -> 23.84 | Hurt |
| use LocoGS instead of compact baselines | Deep Blending · Deep Blending | PSNR: 29.99 -> 30.11 | Helped |
| add dense initialization (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | PSNR: 27.16 -> 27.33 | Helped |
| add dense initialization (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | FPS: 246 -> 270 | Helped |
| add adaptive SH bandwidth (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | FPS: 238 -> 270 | Helped |
| add adaptive SH bandwidth (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | PSNR: 27.29 -> 27.33 | Helped |
| add G-PCC position compression (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | Size (MB): 44.14 -> 38.99 | Helped |
| add explicit attribute compression (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | Size (MB): 38.99 -> 32.25 | Helped |
| add implicit attribute compression (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | Size (MB): 32.25 -> 13.89 | Helped |
| remove dense initialization (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | PSNR: 27.33 -> 27.16 | Hurt |
| treat base scale as implicit (ablation) | LocoGS on Mip-NeRF 360 · Mip-NeRF 360 | PSNR: 27.33 -> 20.48 | Hurt |
Surprising findings
- Applying dense initialization to existing compression methods often increases their storage and reduces FPS, and HAC does not benefit from it under compression tuning.
- Treating base scale as an implicit (neural-field) attribute makes training highly unstable and yields very few Gaussians with poor quality.
Limitations
- Training takes about one hour per scene and requires more memory than 3DGS, which challenges large-scale scenes.
- Hash grid hyperparameters assume cube-shaped, object-centric scenes and may be less effective for large-scale scenes.
- Compression results are reported primarily with a single hash-grid configuration per variant; dense initialization depends on a Nerfacto pre-stage.
Experimental setup
Datasets: Mip-NeRF 360 Tanks and Temples Deep Blending
Benchmarks: Mip-NeRF 360 Tanks and Temples Deep Blending
Metrics: PSNR SSIM LPIPS storage size (MB) FPS
Baselines: 3DGS Scaffold-GS CompGS Compact-3DGS C3DGS LightGaussian EAGLES SSGS HAC
Builds on: 3D Gaussian Splatting multi-resolution hash encoding Nerfacto Compact-3DGS soft movement pruning (Mask-based pruning)
Closest papers by meaning
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.92)
- Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis (CVPR 2024 · similarity 0.92)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.92)
- CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression (ICLR 2025 · similarity 0.92)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.91)
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.