Research map / 3D vision / Gaussian splatting
Compact 3D Gaussian Representation for Radiance Field
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 a compact 3D Gaussian Splatting representation that learns a volume-based mask to prune Gaussians, encodes view-dependent color with a hash-grid neural field, and compresses scale/rotation with residual vector quantization, achieving ~25x storage reduction over 3DGS with comparable quality.
Problem
3D Gaussian Splatting requires a very large number of Gaussians and per-Gaussian attributes, leading to high memory and storage costs that hinder practical use of fast radiance field rendering.
Approach
Combine a learnable volume-and-opacity mask that prunes Gaussians throughout training, a hash-grid plus tiny MLP replacing spherical harmonics for view-dependent color, and residual vector quantization codebooks for scale and rotation, followed by optional quantization and Huffman coding.
Key results
- On Deep Blending, the method reaches 29.79 PSNR vs 29.46 for re-evaluated 3DGS, with 43.2 MB storage vs 663 MB and 181 vs 132 FPS.
- With post-processing (Ours+PP), storage is reduced to 29.1 MB on Mip-NeRF 360 (vs 746 MB for 3DGS*) and 20.9 MB on Tanks&Temples, with PSNR close to 3DGS.
- On NeRF-Synthetic, storage drops to 5.55 MB (x0.08 of 3DGS) and with post-processing to 2.67 MB (x0.04), with FPS rising from 359 to 545.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace per-Gaussian SH color with a hash-grid neural field | Compact 3D Gaussian Representation · Mip-NeRF 360 | storage: 746 MB -> 48.8 MB | Helped |
| replace per-Gaussian SH color with a hash-grid neural field | Compact 3D Gaussian Representation · Deep Blending | PSNR: 29.46 -> 29.79 | Helped |
| apply the full compact representation | Compact 3D Gaussian Representation · NeRF-Synthetic | rendering FPS: 359 -> 545 | Helped |
| add learnable volume masking (ablation) | 3DGS · Playroom | PSNR: 29.87 -> 29.91 | Helped |
| add learnable volume masking (ablation) | 3DGS · Playroom | storage: 553 MB -> 228 MB | Helped |
| add learnable volume masking (ablation) | 3DGS · Bonsai | PSNR: 32.16 -> 32.22 | Helped |
| add learnable volume masking and neural-field color representation (ablation) | 3DGS · Playroom | PSNR: 29.87 -> 30.33 | Helped |
| add learnable volume masking and neural-field color representation (ablation) | 3DGS · Bonsai | PSNR: 32.16 -> 32.08 | No clear effect |
| add neural-field color representation to the masking method (ablation) | Compact 3D Gaussian Representation · Playroom | storage: 228 MB -> 59 MB | Helped |
| add geometry codebook (ablation) | Masking + color field · Playroom | storage: 59 MB -> 44 MB | Helped |
| add geometry codebook (ablation) | Masking + color field · Playroom | PSNR: 30.33 -> 30.33 | No clear effect |
| add half-precision positions and opacities (ablation) | Masking + color field + geometry codebook · Playroom | storage: 44 MB -> 38 MB | Helped |
Surprising findings
- The masking strategy alone slightly improves PSNR on Playroom (29.87 to 29.91) while cutting Gaussians from 2.34M to 967K.
Limitations
- Training time is slightly longer than 3DGS on some datasets (e.g., NeRF-Synthetic 8m04s vs 6m14s).
- Slight quality drop on Mip-NeRF 360 (PSNR 27.08 vs 27.46 for 3DGS*) and Tanks&Temples SSIM/LPIPS.
- Color representation still requires relatively large storage in the end-to-end model, requiring post-processing to reduce it.
Experimental setup
Datasets: Mip-NeRF 360 Tanks&Temples Deep Blending NeRF-Synthetic
Benchmarks: Mip-NeRF 360 Tanks&Temples Deep Blending NeRF-Synthetic
Metrics: PSNR SSIM LPIPS training time FPS storage
Baselines: 3DGS Plenoxels INGP-base INGP-big Mip-NeRF 360
Builds on: 3D Gaussian Splatting Instant NGP residual vector quantization mip-NeRF 360
Closest papers by meaning
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.93)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.93)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.93)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.93)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- 3D-HGS: 3D Half-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.