FLoD: Integrating Flexible Level of Detail into 3D Gaussian Splatting for Customizable Rendering

This is the official implementation of FLoD.

Yunji Seo*, Young Sun Choi*, Hyun Seung Son, Youngjung Uh

Overview

We introduce integrating a Flexible Level of Detail (FLoD) to 3DGS, to allow a scene to be rendered at varying levels of detail according to hardware capabilities.

Installation

Our code was tested Ubuntu 20.04.6 LTS, on conda environment installed with environment.yml and the submodules below.

git clone https://github.com/3DGS-FLoD/flod.git
cd flod

Setup conda environment

conda env create -f environment.yml
conda activate flod

Clone submodules

git clone https://github.com/graphdeco-inria/diff-gaussian-rasterization submodules/diff-gaussian-rasterization
git clone https://gitlab.inria.fr/bkerbl/simple-knn.git submodules/simple-knn

Install dependencies

sudo apt install libglm-dev
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn

Hardware Requirements

Training

• Trained on: a single RTX A5000 GPU with 24GB VRAM
• VRAM requirement: ~22GB VRAM is required for training of all scenes

Evaluation/Inference

• Tested on:
  • a single RTX A5000 GPU with 24GB VRAM
  • a single GTX 1080 GPU with 8GB VRAM
  • a single GeForce MX250 GPU with 2GB VRAM (for laptop evaluations)
• VRAM requirement: ~3GB VRAM is required for full evaluation on all scenes

Training Duration

Average training duration per scene:

• DL3DV-10K: ~40min
• Mip-NeRF 360: ~80min
• Tanks and Temples: ~20min

Training and Evaluation

To reproduce, run... (Links to datasets used in the paper: Mip-NeRF 360, Tanks&Temples, DL3DV-10K)

train.sh 

Render and evaluate by...

render.sh 

Pretrained Models

Pretrained FLoD-3DGS models are available in Hugging Face.

Using Pretrained Models

The pretrained models are organized in the pretrained_models/ directory with the following structure:

pretrained_models/
├── mipnerf360/
│   ├── bonsai/
│   ├── kitchen/
│   ├── counter/
│   └── ...
├── dl3dv/
│   ├── [hash1]/
│   ├── [hash2]/
│   └── ...
└── tnt/
    ├── truck/
    └── train/

Each scene contains:

• point_cloud/ directory (model data)
• cfg_args file (configuration)

Important: Before running the pretrained models, you need to modify the cfg_args files:

1. Update source_path: Replace {path-to-data} with the actual path to your dataset:

   source_path='{path-to-data}/mipnerf360/bonsai'
   

   Change to:

   source_path='/path/to/your/mipnerf360/bonsai'
   

2. Verify model_path: Should already be correct:

   model_path='./pretrained_models/{dataset}/{scene}'
   

Viewer(Demo on Laptop, GeForce MX250)

demo-garden.mp4

To run viewer as demonstrated on our project page

convert4viewer.sh
SIBR_viewers/install/bin/SIBR_flodViewer_app -m /path/to/your/model

Licencse

We build our code for FLoD on top of the open-source code of 3D Gaussian Splatting.
Hence our licencse follows graphdeco-inria/gaussian-splatting

Acknowledgement

We would like to express our gratitude to the authors of the 3D Gaussian Splatting.
Their work has laid the foundation for this research.
Our code is largely based on their open-source project: graphdeco-inria/gaussian-splatting

Citation

@article{seo2025flod,
      author  = {Yunji Seo and Young Sun Choi and Hyun Seung Son and Youngjung Uh},
      title   = {FLoD: Integrating Flexible Level of Detail into 3D Gaussian Splatting for Customizable Rendering},
      journal = {ACM Transactions on Graphics (Proceedings of SIGGRAPH)},
      number  = {4},
      volume  = {44},
      year    = {2025},
      doi     = {10.1145/3731430}
}