Research map / 3D vision / Gaussian splatting
Faster-GS: Analyzing and Improving Gaussian Splatting Optimization
Florian Hahlbohm, Linus Franke, Martin Eisemann, Marcus Magnor
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the per-Gaussian backward pass for alpha blending (ablation) | Faster-GS basis · Mip-NeRF360 indoor | training time: 14m29s -> 7m52s (1.84x) | Improved |
| add the per-Gaussian backward pass (ablation) | Faster-GS basis · Mip-NeRF360 outdoor | peak VRAM: 6.33GiB -> 7.69GiB | Hurt |
| replace non-fused Adam with custom fused Adam (ablation) | Faster-GS · Mip-NeRF360 outdoor | training time: 1.33x vs 1.22x PyTorch fused, 1.30x Apex | Improved |
| add load-balanced Gaussian/tile instancing (ablation) | Faster-GS · Mip-NeRF360 outdoor | training time: 1.00x | No clear effect |
| add repeated z-ordering during densification (ablation) | Faster-GS full · Mip-NeRF360 outdoor | training time: 5m52s -> 5m31s | Improved |
| add separate SH buffers (ablation) | Faster-GS basis · Mip-NeRF360 outdoor | peak VRAM: 6.39GiB -> 5.72GiB | Improved |
| add fused activations (ablation) | Faster-GS basis · Mip-NeRF360 outdoor | training time: 17m07s -> 16m19s | Improved |
| add separate sorting stage (ablation) | Faster-GS basis · Mip-NeRF360 outdoor | training time: 1.01x | No clear effect |
| fuse optimizer step into rasterization backward (ablation) | Faster-GS full · Mip-NeRF360 outdoor | training time: 5m31s -> 5m04s | Improved |
| skip parameter updates for invisible Gaussians (ablation) | Faster-GS fused updates · Mip-NeRF360 outdoor | PSNR: 24.73 -> 24.59 | Hurt |
| omit view-dependent SH coefficients (degree 0) (ablation) | Faster-GS fused updates · Mip-NeRF360 outdoor | PSNR: 24.73 -> 24.38 | Hurt |
| run on newer GPU (RTX 5090 vs RTX 3090) | Faster-GS · all 13 scenes | speedup over 3DGS: 3.9x -> 4.8x | Improved |
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- DiffGS: Functional Gaussian Splatting Diffusion (NeurIPS 2024 · similarity 0.91)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- POp-GS: Next Best View in 3D-Gaussian Splatting with P-Optimality (CVPR 2025 · similarity 0.90)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
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.
