Research map / 3D vision / Gaussian splatting
LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS
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.
Compresses 3D Gaussian Splatting scenes ~15x via significance-based pruning with co-adaptation, SH distillation to degree 2 with pseudo-views, and vector quantization of low-significance SH coefficients, reaching 237 FPS on Mip-NeRF 360.
Problem
3D Gaussian Splatting produces millions of Gaussians per unbounded scene, requiring gigabytes of storage and slowing rendering; the goal is a compact representation that keeps rendering quality and speed.
Approach
Borrows network pruning ideas to score each Gaussian's global significance (hit count weighted by opacity, volume, and transmittance over training rays), prunes low-significance Gaussians followed by fine-tuning, distills high-degree SHs into lower-degree student SHs using pseudo-views, and applies significance-weighted vector quantization only to the least significant SH coefficients.
Key results
- On Mip-NeRF 360, LightGaussian reduces storage from 782MB to 45MB (about 15x) and raises FPS from 144 to 237 with PSNR 27.13 vs 27.40 for re-trained 3D-GS.
- On Tanks and Temples, LightGaussian reaches 357 FPS at 25MB versus 106 FPS and 433MB for re-trained 3D-GS.
- On Scaffold-GS, pruning 80% of neural Gaussians raises FPS from 152 to 178 and reduces size from 173.60 to 112.56 (units as reported) with a small quality drop (PSNR 27.96 to 27.78).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compress 3D-GS with LightGaussian | 3D-GS · Mip-NeRF 360 | Size: 782MB -> 45MB | Helped |
| compress 3D-GS with LightGaussian | 3D-GS · Mip-NeRF 360 | FPS: 144 to 237 | Helped |
| apply LightGaussian compression | 3D-GS · Mip-NeRF 360 | SSIM: 0.813 -> 0.806 | Hurt |
| apply LightGaussian compression | 3D-GS · Tank and Temple | FPS: 106 to 357 | Helped |
| apply LightGaussian compression | 3D-GS · Tank and Temple | PSNR: 23.66 -> 23.44 | Hurt |
| add Gaussian pruning without co-adaptation (ablation) | 3D-GS on Room · Room | PSNR: 31.34 -> 30.67 | Hurt |
| add co-adaptation after Gaussian pruning (ablation) | Gaussian pruning on Room · Room | PSNR: 30.67 -> 31.64 | Helped |
| reduce SH degree directly without distillation (ablation) | 3D-GS on Room · Room | PSNR: 31.64 -> 30.32 | Hurt |
| add photometric loss to compact SH (ablation) | SH compactness on Room · Room | PSNR: 30.32 -> 31.42 | Helped |
| add distillation and pseudo-views to compact SH (ablation) | SH compactness on Room · Room | PSNR: 31.42 -> 31.48 | Helped |
| add VQ finetuning after codebook quantization (ablation) | LightGaussian on Room · Room | PSNR: 31.09 -> 31.40 | Helped |
| use hit count only for pruning significance (ablation) | Gaussian pruning on Room · Room | PSNR: 31.34 -> 28.16 | Hurt |
Surprising findings
- Zero-shot opacity-based pruning drops PSNR sharply (27.2 to 25.3), while co-adaptation fine-tuning recovers most quality after pruning.
- Applying VQ to all attributes (position, rotation, scale) strongly degrades quality, so VQ is restricted to SH coefficients.
Limitations
- Zero-shot pruning without fine-tuning degrades quality noticeably (PSNR drop), requiring co-adaptation.
- Applying VQ to position, rotation, or scale causes significant accuracy loss, limiting compression to SH coefficients.
- Some metrics are slightly lower than 3D-GS on certain scenes and for the Scaffold-GS generalization experiment (LPIPS worsens from 0.2075 to 0.2197).
Experimental setup
Datasets: Mip-NeRF 360 Tanks and Temples NeRF-Synthetic (Blender)
Benchmarks: Mip-NeRF 360 Tanks and Temples NeRF-Synthetic
Metrics: FPS model size PSNR SSIM LPIPS
Baselines: Plenoxels Instant-NGP Mip-NeRF 360 VQ-DVGO Compressed 3D-GS Compact 3D-GS 3D-GS
Builds on: 3D Gaussian Splatting Scaffold-GS VQ-DVGO Compact 3D-GS
Closest papers by meaning
- MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks (CVPR 2025 · similarity 0.92)
- VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction (CVPR 2024 · similarity 0.92)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis (CVPR 2024 · similarity 0.90)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.90)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · 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.