Research map / 3D vision / Gaussian splatting
Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives
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.
Speedy-Splat accelerates 3D Gaussian Splatting rendering by precisely localizing Gaussians to tiles (SnugBox, AccuTile) and pruning Gaussians during and after training (Soft and Hard Pruning), yielding 6.71x faster rendering and 10.6x fewer Gaussians.
Problem
3D Gaussian Splatting rendering speed and model size are bottlenecks, especially in resource-constrained settings, due to conservative tile intersection and too many Gaussians.
Approach
Compute exact pixel extent of each Gaussian from its opacity-thresholded ellipse to produce tight bounding boxes and exact tile intersections, and prune Gaussians during densification and afterward using an efficient, memory-light gradient-squared sensitivity score derived from PUP 3D-GS.
Key results
- Average rendering speed-up of 6.71x (134 to 898 FPS) with 10.6x fewer Gaussians (2.93M to 0.28M) across evaluated scenes.
- SnugBox and AccuTile are lossless, giving 1.82x and 1.99x inference speed-ups respectively without changing renderings.
- Training time reduced by 1.47x; Speedy-Splat's pruning score outperforms PUP 3D-GS on PSNR in post-hoc pruning (26.87 vs 26.21).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SnugBox to precisely bound Gaussian-tile intersections (ablation) | 3D-GS rendering · All evaluated scenes | FPS: 134 -> 244 | Helped |
| Add AccuTile after SnugBox to identify exact tile intersections (ablation) | 3D-GS rendering · All evaluated scenes | FPS: 244 -> 267 | Helped |
| Add Soft Pruning during densification (ablation) | 3D-GS with SnugBox and AccuTile · All evaluated scenes | FPS: 267 -> 420 | Helped |
| Add Hard Pruning after densification (ablation) | 3D-GS with SnugBox, AccuTile, and Soft Pruning · All evaluated scenes | FPS: 420 -> 898 | Helped |
| Add SnugBox (ablation) | 3D-GS · Mip-NeRF 360 | PSNR: 27.55 -> 27.55 | No clear effect |
| Add AccuTile after SnugBox (ablation) | 3D-GS · Mip-NeRF 360 | PSNR: 27.55 -> 27.57 | No clear effect |
| Add Soft Pruning after SnugBox and AccuTile (ablation) | 3D-GS · Mip-NeRF 360 | PSNR: 27.57 -> 27.32 | Hurt |
| Add Hard Pruning after Soft Pruning (ablation) | 3D-GS · Mip-NeRF 360 | PSNR: 27.32 -> 26.94 | Hurt |
| Use the Speedy-Splat pruning score instead of the PUP score (ablation) | PUP prune-refine pipeline · All evaluated scenes | PSNR: 26.2136 -> 26.8658 | Helped |
Surprising findings
- Soft pruning at 80% ratio during densification preserves visual fidelity.
- Speedy-Splat's efficient pruning score achieves higher PSNR than PUP 3D-GS but lower SSIM and LPIPS in post-hoc pruning.
Limitations
- Slightly lower image quality than 3D-GS at high compression ratios (Hard Pruning drops SSIM to 0.782 and LPIPS rises to 0.296 on Mip-NeRF 360).
- Efficient pruning score shows a slight performance gap relative to the PUP 3D-GS score in post-hoc pruning.
Experimental setup
Datasets: Mip-NeRF 360 Tanks & Temples Deep Blending
Benchmarks: Mip-NeRF 360 Tanks & Temples Deep Blending
Metrics: FPS PSNR SSIM LPIPS number of Gaussians training time
Baselines: 3D-GS PUP 3D-GS Mini-Splatting LightGaussian EAGLES ELMGS Trimming StopThePop
Builds on: 3D Gaussian Splatting PUP 3D-GS LightGaussian Mini-Splatting
Closest papers by meaning
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.95)
- 3D Gaussian Rendering Can Be Sparser: Efficient Rendering via Learned Fragment Pruning (NeurIPS 2024 · similarity 0.93)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.93)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.93)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
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.