Research map / 3D vision / Gaussian splatting
Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis
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 optimized 3D Gaussian splat scenes ~26-31x via sensitivity-aware vector clustering, quantization-aware fine-tuning, and entropy coding, and renders them with hardware rasterization for up to 4x faster rendering.
Problem
3D Gaussian splatting scenes require gigabytes of memory and a specialized compute pipeline, which hinders network streaming and rendering on low-power devices and integration with polygon rasterization.
Approach
Sensitivity-weighted k-means clustering of SH color coefficients and normalized Gaussian shape (covariance) parameters into codebooks, followed by quantization-aware fine-tuning, Morton-order sorting and DEFLATE encoding, plus a GPU hardware-rasterization renderer in WebGPU.
Key results
- Average compression ratio of 25.77x (up to 31x) with about 0.23 PSNR average drop on real-world scenes (e.g., 640 MB to 24 MB average size).
- Rendering speed up to 4x higher than the 3DGS compute pipeline (e.g., Bicycle at 1080p: 93 to 321 FPS on RTX A5000; 54 to 211 FPS on RTX 3070M).
- Compression takes about 5-6 minutes and increases reconstruction time by roughly 10%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Compress 3D Gaussian splats | 3D Gaussian Splatting · Mip-NeRF360 | PSNR: 27.21 -> 26.981 | Hurt |
| Compress 3D Gaussian splats | 3D Gaussian Splatting · Mip-NeRF360 | SIZE (MB): 795.26 -> 28.80 | Helped |
| Use compressed representation with hardware rasterization | 3D Gaussian Splatting · Bicycle, 1080p | FPS: 93 -> 321 | Helped |
| Add pruning (ablation) | 3D Gaussian Splatting compression pipeline · Mip-NeRF360 garden | PSNR: 27.179 -> 27.083 | Hurt |
| Add color clustering (ablation) | Pruned 3D Gaussian Splatting compression pipeline · Mip-NeRF360 garden | PSNR: 27.083 -> 25.941 | Hurt |
| Add Gaussian clustering (ablation) | Color-clustered 3D Gaussian Splatting compression pipeline · Mip-NeRF360 garden | PSNR: 25.941 -> 25.781 | Hurt |
| Add quantization-aware fine-tuning (ablation) | Clustered 3D Gaussian Splatting compression pipeline · Mip-NeRF360 garden | PSNR: 25.781 -> 26.746 | Helped |
| Add entropy encoding (ablation) | Quantization-aware fine-tuned 3D Gaussian Splatting · Mip-NeRF360 garden | SIZE (MB): 86.69 -> 58.40 | Helped |
| Add Morton-order sorting (ablation) | Entropy-encoded 3D Gaussian Splatting · Mip-NeRF360 garden | SIZE (MB): 58.40 -> 46.57 | Helped |
| Vary color codebook size from 1024 to 8192 (ablation) | Ours · Mip-NeRF360 | PSNR: 26.95(-0.67) -> 27.00(-0.58) | No clear effect |
| Vary Gaussian codebook size from 1024 to 8192 (ablation) | Ours · Mip-NeRF360 | PSNR: 26.95(-0.79) -> 26.97(-0.60) | No clear effect |
| Vary color sensitivity threshold βc (ablation) | Ours · Mip-NeRF360 | PSNR: 27.22(-0.25); 26.74(-0.94) | Mixed |
Surprising findings
- Pruning splats with zero sensitivity in color parameters removes up to 15% of Gaussians with no quality loss.
- Roughly half of the 4x speedup comes from reduced memory bandwidth alone, with the remainder from hardware rasterization.
Limitations
- Gaussian positions cannot be aggressively compressed without significant rendering error.
- Training memory footprint is not reduced.
- Compression tested only on 3DGS-reconstructed scenes.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples Deep Blending NeRF-Synthetic
Benchmarks: Mip-NeRF360 Tanks&Temples Deep Blending NeRF-Synthetic
Metrics: PSNR SSIM LPIPS model size (MB) compression ratio rendering FPS
Baselines: 3D Gaussian Splatting (Kerbl et al.)
Builds on: 3D Gaussian Splatting NeRF compression via vector quantization (Li et al.) k-bit quantization-aware training DEFLATE
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
- Locality-aware Gaussian Compression for Fast and High-quality Rendering (ICLR 2025 · similarity 0.92)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.91)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 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.