Research map / 3D vision / Gaussian splatting
GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 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 GaussianSpa, which casts 3DGS simplification as an ℓ0-constrained optimization solved during training by alternating gradient "optimizing" steps and closed-form top-κ "sparsifying" projections, yielding compact high-quality splats.
Problem
3D Gaussian Splatting requires large memory to store many Gaussians; existing simplification methods prune via hand-crafted criteria or masks in a one-shot manner, causing irreversible quality loss.
Approach
Reformulate 3DGS training with an ℓ0 constraint on Gaussian opacities and solve it with an ADMM-style split into a gradient-based "optimizing" step and an analytic top-κ projection "sparsifying" step, gradually sparsifying Gaussians during training rather than removing them at once.
Key results
- On Deep Blending, GaussianSpa reaches 30.33 dB PSNR with 0.256M Gaussians versus 29.42 dB for 3DGS with 2.78M Gaussians.
- On Mip-NeRF 360, GaussianSpa achieves 27.85 dB PSNR with 0.547M Gaussians, outperforming 3DGS (27.45 dB, 3.11M) and Mini-Splatting (27.40 dB, 0.559M).
- On Tanks&Temples, GaussianSpa achieves 23.98 dB PSNR with 0.269M Gaussians, versus 23.63 dB for 3DGS with 1.83M Gaussians.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply optimizing-sparsifying simplification | 3DGS · Mip-NeRF 360 | PSNR: 27.45 -> 27.85 | Helped |
| Apply optimizing-sparsifying simplification | 3DGS · Tanks&Temples | PSNR: 23.63 -> 23.98 | Helped |
| Apply optimizing-sparsifying simplification | 3DGS · Deep Blending | PSNR: 29.42 -> 30.33 | Helped |
| Use GaussianSpa instead of Mini-Splatting | 3DGS simplification · Deep Blending | PSNR: 30.05 -> 30.33 | Helped |
| Use GaussianSpa instead of Taming 3DGS | 3DGS simplification · Tanks&Temples | PSNR: 23.95 -> 23.98 | No clear effect |
| Use GaussianSpa instead of LP-3DGS-R | 3DGS simplification · Mip-NeRF 360 | PSNR: 27.47 -> 27.85 | Helped |
Surprising findings
- After removing about 90% of Gaussians, the simplified model slightly exceeds the original 3DGS quality after light tuning.
- One-shot pruning baselines (e.g., opacity- or importance-based) lose quality permanently even with long subsequent training, as shown in Figure 2.
Limitations
- Relies on the same starting checkpoints as Mini-Splatting, so gains are measured relative to that initialization.
- Hyperparameters such as the start iteration (15K) and removal iteration (25K) are fixed and their sensitivity is not extensively studied.
Experimental setup
Datasets: Mip-NeRF 360 Tanks&Temples Deep Blending
Benchmarks: Mip-NeRF 360 Tanks&Temples Deep Blending
Metrics: PSNR SSIM LPIPS number of Gaussians
Baselines: 3DGS CompactGaussian LP-3DGS EAGLES Mini-Splatting Taming 3DGS CompGS LightGaussian
Builds on: 3D Gaussian Splatting Mini-Splatting ADMM LightGaussian
Closest papers by meaning
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.93)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.92)
- ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.92)
- PUP 3D-GS: Principled Uncertainty Pruning for 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.