Research map / 3D vision / Gaussian splatting
PUP 3D-GS: Principled Uncertainty Pruning for 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 PUP 3D-GS, a post-hoc multi-round prune-refine pipeline that prunes pretrained 3D-GS scenes using a Fisher-information-based sensitivity score over Gaussian spatial parameters, reaching 90% pruning with higher rendering speed than LightGaussian.
Problem
3D Gaussian Splatting scenes contain millions of Gaussians, causing high storage and memory costs; existing heuristic pruning methods degrade visual fidelity and foreground details at high compression ratios.
Approach
A principled per-Gaussian sensitivity score computed as the log determinant of the Fisher information (a second-order approximation of reconstruction error) block for each Gaussian's mean and scale parameters, computed on 4x4 patches, combined with multi-round prune-refine fine-tuning.
Key results
- At 90% pruning, PUP 3D-GS achieves 204.81 FPS vs 3D-GS 95.59 FPS on Mip-NeRF 360 (3.56x average speedup), with PSNR 26.67 vs LightGaussian 26.28.
- Two-round prune-refine (80% then 50%) outperforms one-round 90% pruning across all metrics on Mip-NeRF 360 (PSNR 26.67 vs 26.12).
- With Vectree quantization, final models are about 51-53x smaller than unpruned 3D-GS, and PUP 3D-GS outperforms LightGaussian on PSNR/SSIM/LPIPS/FPS on Mip-NeRF 360 and Deep Blending.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PUP 3D-GS pruning instead of LightGaussian pruning | 3D-GS · Mip-NeRF 360 | PSNR: 26.28 -> 26.67 | Helped |
| Use PUP 3D-GS pruning instead of LightGaussian pruning | 3D-GS · Tanks & Temples | PSNR: 23.08 -> 22.72 | Hurt |
| Use PUP 3D-GS pruning instead of LightGaussian pruning | 3D-GS · Deep Blending | PSNR: 28.51 -> 28.85 | Helped |
| Use PUP 3D-GS pruning instead of LightGaussian pruning | 3D-GS · Mip-NeRF 360 | FPS: 162.12 -> 204.81 | Helped |
| Use PUP 3D-GS pruning instead of LightGaussian pruning | 3D-GS · Tanks & Temples | FPS: 329.03 -> 391.10 | Helped |
| Use PUP 3D-GS pruning instead of LightGaussian pruning | 3D-GS · Deep Blending | FPS: 234.10 -> 301.43 | Helped |
| Split pruning into two rounds with refinement after each, rather than one round (ablation) | PUP 3D-GS · Mip-NeRF 360 | PSNR: 26.12 -> 26.67 | Helped |
| Include rotation parameters in the sensitivity score (ablation) | PUP 3D-GS · Mip-NeRF 360 | PSNR: 26.67 -> 26.79 | Mixed |
| Use RGB color parameters to compute the sensitivity score instead of spatial mean and scale (ablation) | PUP 3D-GS · Mip-NeRF 360 | PSNR: 26.67 -> 26.40 | Hurt |
Surprising findings
- Including rotation parameters in the sensitivity score lowers image quality, and using RGB color parameters instead performs worse than the spatial score.
Limitations
- Assumes the scene is converged so the Fisher approximation (dropping second-order residual term) is accurate.
- Memory for Fisher matrices scales with number of Gaussians (N x 36), requiring 16GB+ GPUs.
- Including rotation parameters increases computational cost and requires more expensive hardware.
Experimental setup
Datasets: Mip-NeRF 360 Tanks & Temples Deep Blending
Benchmarks: Mip-NeRF 360 Tanks & Temples Deep Blending
Metrics: PSNR SSIM LPIPS FPS model size (MB)
Baselines: LightGaussian 3D-GS EAGLES Compact-3DGS Mini-Splatting RadSplat
Builds on: 3D Gaussian Splatting LightGaussian FisherRF BayesRays
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.95)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- 3D Gaussian Rendering Can Be Sparser: Efficient Rendering via Learned Fragment Pruning (NeurIPS 2024 · similarity 0.93)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · similarity 0.92)
- MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks (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.