Research map / 3D vision / Gaussian splatting
LP-3DGS: Learning to Prune 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 LP-3DGS, which learns a per-Gaussian trainable mask with a Gumbel-Sigmoid relaxation on top of existing pruning importance scores (RadSplat, Mini-Splatting) so the optimal pruning ratio is found in a single training run.
Problem
3DGS produces millions of Gaussians, causing high memory use; existing pruning methods need a manually preset threshold or pruning ratio, requiring multiple training rounds per scene to find the best trade-off.
Approach
Apply a trainable mask to importance scores and binarize it with a differentiable Gumbel-Sigmoid instead of a straight-through estimator, so the number of pruned Gaussians is learned automatically during one training run with a sparsity (L1) regularizer.
Key results
- On MipNeRF360 with the RadSplat score, LP-3DGS reaches an average 0.63 pruning ratio at 27.47 PSNR versus a 27.48 baseline PSNR, with per-scene ratios from 0.58 to 0.74.
- With the Mini-Splatting score, LP-3DGS reaches an average 0.60 pruning ratio at 27.12 PSNR.
- Replacing the STE mask on the RadSplat score with LP-3DGS raises average PSNR from 26.955 to 27.470 at a lower pruning ratio (0.63 vs 0.86).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LP-3DGS with the RadSplat score instead of the 3DGS baseline | 3DGS · MipNeRF360 | PSNR: 27.48 -> 27.47 | No clear effect |
| Use LP-3DGS with the RadSplat score instead of the 3DGS baseline | 3DGS · MipNeRF360 | LPIPS: 0.2215 -> 0.2269 | Hurt |
| Use LP-3DGS with the RadSplat score instead of the 3DGS baseline | 3DGS · NeRF-Synthetic | PSNR: 33.798 -> 33.709 | No clear effect |
| Use LP-3DGS with the RadSplat score instead of the 3DGS baseline | 3DGS · Tanks & Temples | LPIPS: 0.1783 -> 0.1879 | Hurt |
| Replace the STE mask with a Gumbel-Sigmoid mask on opacity and scale (ablation) | Compact3D · MipNeRF360 | PSNR: 27.284 -> 27.442 | Helped |
| Replace the Gumbel-Sigmoid mask with an STE mask on the RadSplat importance score (ablation) | LP-3DGS · MipNeRF360, RadSplat score | PSNR: 27.470 -> 26.955 | Hurt |
| Replace the Gumbel-Sigmoid mask with an STE mask on the Mini-Splatting importance score (ablation) | LP-3DGS · MipNeRF360, Mini-Splatting score | PSNR: 27.12 -> 26.98 | Hurt |
Surprising findings
- The STE mask on the importance score over-prunes (about 0.86 pruning ratio) and loses substantial quality, while the learned Gumbel-Sigmoid mask keeps a moderate ratio.
- Pruning-score choice strongly affects final quality: LP-3DGS with the Mini-Splatting score loses more PSNR and LPIPS than with the RadSplat score.
Limitations
- Rendering quality after pruning depends on the chosen importance score.
- Multi-GPU training is not supported.
- Mask-learning memory overhead slightly increases peak training memory.
Experimental setup
Datasets: MipNeRF360 Tanks & Temples NeRF-Synthetic
Benchmarks: MipNeRF360 NeRF-Synthetic Tanks & Temples
Metrics: PSNR SSIM LPIPS FPS training time peak memory number of Gaussians
Baselines: RadSplat Mini-Splatting Compact3D STE mask on importance score 3DGS
Builds on: 3D Gaussian Splatting RadSplat Mini-Splatting Compact3D Gumbel-Softmax
Closest papers by meaning
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.93)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.93)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
- MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks (CVPR 2025 · similarity 0.92)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (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.