Research map / 3D vision / Gaussian splatting
3D Gaussian Rendering Can Be Sparser: Efficient Rendering via Learned Fragment Pruning
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 learned per-Gaussian fragment (pixel-footprint) truncation thresholds, fine-tuned via a sigmoid relaxation on pre-trained 3DGS/Mini-Splatting/4DGS models, to speed rendering on edge GPUs, orthogonal to primitive pruning.
Problem
3D Gaussian Splatting requires millions of primitives, and existing primitive pruning methods do not reduce rasterization cost proportionally because the number of fragments per Gaussian grows, limiting on-device rendering speed.
Approach
Learn a per-Gaussian truncation threshold on the exponent term and replace the non-differentiable indicator with a sigmoid, fine-tuning only opacity and threshold of a pre-trained Gaussian model so each Gaussian covers fewer pixels; this is applied orthogonally to existing primitive pruning.
Key results
- Fragment pruning on pre-trained 3DGS raises FPS by +17, +20, +26 on Mip-NeRF 360, Tanks&Temples, and Deep Blending while PSNR improves by +0.04, +0.04, +0.05.
- Combined with Mini-Splatting, fragment pruning reaches 60 FPS on Tanks&Temples and Deep Blending, a 71% FPS gain on Tanks&Temples on Jetson Orin NX.
- On dynamic scenes (4DGS), fragment pruning raises FPS from 14 to 22 with PSNR 32.01 to 32.04.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add fragment pruning to 3D Gaussian | 3D Gaussian · Mip-NeRF 360 | FPS: +17 | Helped |
| Add fragment pruning to 3D Gaussian | 3D Gaussian · Tanks&Temples | FPS: +20 | Helped |
| Add fragment pruning to 3D Gaussian | 3D Gaussian · Deep Blending | FPS: +26 | Helped |
| Add fragment pruning to Mini-Splatting | Mini-Splatting · Mip-NeRF 360 | FPS: +19; +54% | Helped |
| Add fragment pruning to Mini-Splatting | Mini-Splatting · Tanks&Temples | FPS: +25; +71% | Helped |
| Add fragment pruning to Mini-Splatting | Mini-Splatting · Deep Blending | FPS: +21; +54% | Helped |
| Add fragment pruning to 4DGS | 4DGS · Plenoptic Video Dataset | FPS: +8 FPS | Helped |
Surprising findings
- Pruning Gaussian primitives with Mini-Splatting reduces Gaussian count 7.79x but rasterization only speeds up 1.81x because fragments per Gaussian increase.
Limitations
- Fragment pruning is a post-training fine-tuning step, adding 16%-21% overhead relative to pre-training time.
- Fine-tuning only opacity and truncation threshold, since full fine-tuning overfits.
- Evaluated only on a few scenes per dataset with no error bars reported.
Experimental setup
Datasets: Mip-NeRF 360 Tanks&Temples Deep Blending Plenoptic Video Dataset
Benchmarks: Mip-NeRF 360 Tanks&Temples Deep Blending Plenoptic Video Dataset
Metrics: PSNR SSIM LPIPS FPS
Baselines: 3D Gaussian Splatting Mini-Splatting Plenoxels INGP-Big Mip-NeRF 360 4DGS K-Planes MixVoxels-L NeRFPlayer HexPlane HyperReel
Builds on: 3D Gaussian Splatting Mini-Splatting 4D Gaussian Splatting LightGaussian
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.93)
- MaskGaussian: Adaptive 3D Gaussian Representation from Probabilistic Masks (CVPR 2025 · similarity 0.91)
- LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · 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.