Seungjun Oh, Younggeun Lee, Hyejin Jeon, Eunbyung Park

3D-4D Gaussian Splatting introduces a hybrid representation that combines 3D and 4D Gaussians to model dynamic scenes efficiently — reducing memory, improving speed, and preserving quality.

• Efficient hybrid representation (3D for static, 4D for dynamic)
• Faster training than 4DGS with similar or better quality
• Drop-in replacement for existing 4DGS pipelines

git clone https://github.com/ohsngjun/3D-4DGS.git
cd 3D-4DGS
conda env create --file environment.yml
conda activate 3d4dgs

Download the dataset here. After downloading the data, preprocess it using:

python scripts/n3v2blender.py $path_to_dataset

Single sequence training:

python main.py --config configs/n3v/default.yaml --model_path <model save path> --source_path <dataset path>

Train all sequences:

bash train.sh

Don't forget to adjust dataset paths in train.sh.

python main.py --config configs/n3v/default.yaml --model_path <model path> --source_path <dataset path> --start_checkpoint <model_path>/chkpnt6000.pth --val

This project builds upon:

• Real-time 4D Gaussian Splatting
• Ex4DGS
• 4D-Rotor Gaussians (data preprocessing)
• @sorceressyidi (visualization code)

@article{oh2025hybrid,
  title={Hybrid 3D-4D Gaussian Splatting for Fast Dynamic Scene Representation},
  author={Oh, Seungjun and Lee, Younggeun and Jeon, Hyejin and Park, Eunbyung},
  journal={arXiv preprint arXiv:2505.13215},
  year={2025}
}