Research map / 3D vision / Gaussian splatting
LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering
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 LODGE, a level-of-detail 3D Gaussian Splatting representation with depth-aware 3D smoothing, importance pruning, and chunk-based pre-computed active Gaussians with opacity blending for real-time large-scale rendering on memory-constrained devices.
Problem
Real-time rendering of large-scale 3D Gaussian Splatting scenes is slow and memory-hungry, especially on mobile and low-end devices, since existing LOD methods recompute active Gaussians per frame and keep all LOD levels in GPU memory.
Approach
Build discrete LOD levels from a trained 3DGS by adding a depth-dependent 3D smoothing filter, pruning with importance scores and fine-tuning; partition the scene into chunks with precomputed fixed active Gaussian sets, and blend opacity between the two nearest chunks to avoid boundary artifacts.
Key results
- On SmallCity, LODGE reaches 26.57 PSNR / 0.815 SSIM / 0.325 LPIPS at 257.46 FPS with 877K Gaussians, outperforming Octree-GS (25.98 PSNR, 120.27 FPS).
- On the Zip-NeRF dataset, LODGE achieves the best FPS on NYC (280.22) and London (252.58) with the highest SSIM on NYC and London among real-time methods.
- On mobile, LODGE renders at 41 FPS on iPhone 13 Mini, 35 FPS on iPhone 15 Pro, 43 FPS on MacBook M3 and 22 FPS on Chromebook, while full H3DGS crashes on iPhones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add one LOD level at d=10 (ablation) | 3DGS representation · SmallCity | Time: 15.17 -> 5.61 | Helped |
| add one LOD level at d=10 (ablation) | 3DGS representation · SmallCity | PSNR: 26.62 -> 26.45 | Hurt |
| use four LOD levels at d=10,28,47,63 instead of two at d=10,28 (ablation) | LOD representation · SmallCity | Time: 4.75 -> 4.07 | Helped |
| manually set LOD thresholds to 15 and 70 instead of automatic thresholds (ablation) | LOD representation · SmallCity | Time: 4.75 -> 5.24 | Hurt |
| add chunk clustering and adjust depth thresholds by chunk radius (ablation) | LOD (d=10,28) · SmallCity | PSNR: 26.50 -> 26.54 | Helped |
| add chunk clustering and adjust depth thresholds by chunk radius (ablation) | LOD (d=10,28) · SmallCity | Time: 4.75 -> 3.62 | Helped |
| add visibility filtering to clustered LOD (ablation) | LOD (d=10,28) + clusters · SmallCity | Time: 3.62 -> 3.15 | Helped |
| add visibility filtering to clustered LOD (ablation) | LOD (d=10,28) + clusters · SmallCity | PSNR: 26.54 -> 26.55 | No clear effect |
| add opacity blending between adjacent chunks (ablation) | clustered LOD with visibility filtering · SmallCity | PSNR: 26.55 -> 26.57 | Helped |
| add opacity blending between adjacent chunks (ablation) | clustered LOD with visibility filtering · SmallCity | Time: 3.15 -> 3.88 | Hurt |
| use LOD levels at d=10,28 instead of no LOD (ablation) | 3DGS representation · SmallCity | Time: 15.17 -> 4.75 | Helped |
| use LOD levels at d=10,28 instead of no LOD (ablation) | 3DGS representation · SmallCity | PSNR: 26.62 -> 26.50 | Hurt |
Surprising findings
- Adding more than two LOD levels yields diminishing returns in speed and quality, so two LODs were used for all experiments.
- Octree-GS and CityGS achieve higher PSNR on Campus but lower SSIM and LPIPS than LODGE.
Limitations
- Assumes Gaussians can be loaded and reloaded efficiently at chunk boundaries, which requires streaming infrastructure not evaluated here.
- Opacity blending uses only the two nearest chunks, which may be insufficient for diagonal trajectories through three or more chunks (argued empirically rather than proven).
Experimental setup
Datasets: Hierarchical 3DGS dataset (SmallCity, Campus) Zip-NeRF dataset (Alameda, London, NYC)
Benchmarks: Hierarchical 3DGS dataset (SmallCity, Campus) Zip-NeRF dataset (Alameda, London, NYC)
Metrics: PSNR SSIM LPIPS (VGG) FPS number of Gaussians in GPU memory
Baselines: Zip-NeRF 3DGS Mip-Splatting Scaffold-GS H3DGS FLoD Octree-GS CityGS
Builds on: 3D Gaussian Splatting Mip-Splatting RadSplat importance pruning H3DGS Zip-NeRF
Closest papers by meaning
- LOD-GS: Achieving Levels of Detail using Scalable Gaussian Soup (CVPR 2025 · similarity 0.93)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.91)
- Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting (AAAI 2025 · similarity 0.91)
- Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering (CVPR 2024 · similarity 0.91)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · 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.