Skip to content

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

Conference: NeurIPS 2025 arXiv: 2505.23158 Code: Available Area: 3D Vision / Large-Scale Scene Rendering Keywords: Gaussian Splatting, Level-of-Detail, Large-Scale Scene, Efficient Rendering, LOD

TL;DR

This paper proposes LODGE, which manages 3D Gaussian Splatting at multiple scales through a hierarchical Level-of-Detail (LOD) strategy. By dynamically selecting Gaussian representations of appropriate granularity based on camera distance, LODGE enables high-quality real-time rendering of large-scale scenes.

Background & Motivation

State of the Field

Background: 3DGS performs well on small scenes but faces an explosion in the number of Gaussians when scaled to large scenes at the city level.

Limitations of Prior Work: Millions of Gaussians lead to (1) GPU memory overflow; (2) enormous rasterization overhead; (3) wasteful computation on distant fine details.

Key Challenge: High quality requires dense Gaussians, whereas efficiency demands sparse representations.

Key Insight: LOD techniques from traditional computer graphics — distant objects are represented at low resolution, and nearby objects at high resolution.

Core Idea: Construct a multi-level LOD pyramid of Gaussians and select the appropriate level based on distance at render time.

Mechanism

Goal: ### Overall Architecture

Scene partitioning → LOD pyramid construction per block (fine to coarse) → distance-based level selection at render time → hybrid rasterization.

Method

Overall Architecture

Scene partitioning → LOD pyramid construction per block (fine to coarse) → distance-based level selection at render time → hybrid rasterization.

Key Designs

  1. LOD Pyramid Construction

    • Function: Progressively merges fine-grained Gaussians into coarser representations level by level.
    • Mechanism: Spatial clustering with attribute fusion — neighboring Gaussians are merged into larger Gaussians with weighted averaging of color and opacity.
    • Design Motivation: At large distances, small Gaussians are covered by individual pixels; merging them yields equivalent visual results with significantly fewer primitives.

  2. Distance-Adaptive Level Selection

    • Function: Selects the appropriate LOD level for each scene block based on camera distance.
    • Mechanism: \(\text{LOD}(d) = \lfloor \log_2(d / d_{min}) \rfloor\), i.e., one coarser level per doubling of distance.
    • Design Motivation: Human vision is insensitive to distant detail; Screen Space Error is used to govern visual quality.

  3. Cross-Level Smooth Transition

    • Function: Prevents abrupt visual changes (popping artifacts) at LOD level boundaries.
    • Mechanism: Gaussians from two adjacent levels are blended in boundary regions using linear interpolation weights.
    • Design Motivation: Sudden level switching produces visible flickering artifacts.

Loss & Training

Each level is trained independently: \(\mathcal{L} = \lambda_1 \mathcal{L}_{photometric} + \lambda_2 \mathcal{L}_{SSIM}\). Training begins at the finest level, with coarser levels constructed by progressive merging.

Key Experimental Results

Main Results

Method Mill19 PSNR↑ Render FPS↑ GPU Memory↓ # Gaussians↓
3DGS 26.8% 15 24GB 45M
Mega-NeRF 25.1% 0.5 8GB N/A
LODGE 26.5% 45 8GB 12M (at render)

Ablation Study

Configuration PSNR FPS Notes
Single level (finest) 26.8% 15 Memory overflow
LOD w/o smooth transition 26.1% 42 Popping artifacts
LOD + smooth transition 26.5% 45 Best

Key Findings

  • LODGE achieves a 3× FPS improvement (15→45) with only a 0.3 dB PSNR drop.
  • GPU memory is reduced from 24 GB to 8 GB, enabling large-scale scene rendering on consumer-grade hardware.
  • Smooth transition contributes +0.4 PSNR and effectively eliminates popping artifacts.

Highlights & Insights

  • Revival of Classic LOD in 3DGS: A well-established graphics technique demonstrates renewed vitality under a novel representation paradigm.
  • Practicality: LODGE makes large-scale 3DGS feasible on consumer hardware, substantially lowering the barrier to deployment.

Limitations & Future Work

  • LOD merging incurs information loss — certain fine details may degrade under close-up observation.
  • LOD update strategies for dynamic scenes are not addressed.
  • Seam artifacts at block boundaries require further handling.
  • vs. Mega-NeRF: Mega-NeRF performs block-based decomposition on NeRF; LODGE achieves greater efficiency on 3DGS.
  • vs. CityGaussian: CityGaussian also targets large-scale 3DGS but lacks an LOD strategy.

Rating

  • Novelty: ⭐⭐⭐⭐ The LOD concept is mature, but its adaptation to 3DGS introduces meaningful innovation.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Validated on large-scale scenes.
  • Writing Quality: ⭐⭐⭐⭐ Clear and well-organized.
  • Value: ⭐⭐⭐⭐⭐ A practical solution for large-scale 3DGS.