LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering

LODGE enables efficient rendering of large-scale 3DGS even on mobile devices

Abstract

LODGE overview — **Left (LOD):** The scene is represented with multiple LODs; `active Gaussians' are selected during training based on camera distance. **Right (cluster-based rendering):** We cluster cameras into chunks, pre-compute `active Gaussians' per chunk, and render the two nearest chunks with `opacity blending'.

Ours

FPS: 257

ZipNeRF

FPS: 0.09

Ours

FPS: 219

Octree-GS

FPS: 119

Ours

FPS: 280

H3DGS

FPS: 33

Ours

FPS: 253

3DGS

FPS: 99

LOD representation for faster rendering

We show the number of visible Gaussians per pixel (lighter means more Gaussians). By adding the LOD representation with chunk-based rendering, we reduce the number of visible Gaussians per pixel, significantly improving the rendering speed with little loss in quality.

Ours

FPS: 258

Full representation

FPS: 66

Opacity interpolation to remove cross-chunk transitions

To ensure temporal consistency when transitioning between chunks, we propose opacity blending. Please focus on the black car. Notice, how opacity blending removes the sharp changes visible in LOD + chunks.

Ours

FPS: 258

LOD + chunks

FPS: 317

Citation

Please use the following citation:

@article{kulhanek2025lodge
  title={{LODGE}: Level-of-Detail Large-Scale {G}aussian Splatting with Efficient Rendering}, 
  author={Jonas Kulhanek and Marie-Julie Rakotosaona and Fabian Manhardt and Christina Tsalicoglou and Michael Niemeyer and Torsten Sattler and Songyou Peng and Federico Tombari},
  year={2025},
  journal={arXiv},
}