This is the offical repository of ACM MM 2025 "Wavelet-GS: 3D Gaussian Splatting with Wavelet Decomposition"
Paper PDF: https://arxiv.org/abs/2507.12498
We recommend deploying Wavelet-GS on a GPU with VRAM ≥ 24GB for stable training and rendering.
git clone https://github.com/ALEX5874/Wavelet-GS.git
cd Wavelet-GS
The environment configuration follows Scaffold-GS.
Please refer to the official Scaffold-GS repository for detailed installation instructions:
https://github.com/city-super/Scaffold-GS
After completing the Scaffold-GS environment setup, the same environment can be used for Wavelet-GS with slightly adjustment.
The training pipeline is consistent with standard 3D Gaussian Splatting and Scaffold-GS workflows.
Example command (adjust dataset paths and configs as needed):
python train.py -m output_dir -s your_dataset_dir
python render.py
Wavelet-GS shares common limitations with Gaussian Splatting–based methods.
Performance may degrade in scenarios involving:
- Extremely sparse input views
- Heavy occlusions
- Highly reflective or transparent surfaces
Further improvements may require integration with stronger geometric priors or learned visibility modeling.
Including but not limited to:
3D Gaussian Splatting, Scaffold-GS, Octree-GS,
Wavelet-based representations, Neural Radiance Fields (NeRF),
Multi-resolution 3D reconstruction.
If you find this work useful, please consider citing:
@inproceedings{zhao2025wavelet,
title={Wavelet-GS: 3D Gaussian Splatting with Wavelet Decomposition},
author={Zhao, Beizhen and Zhou, Yifan and Yu, Sicheng and Wang, Zijian and Wang, Hao},
booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
pages={8616--8625},
year={2025}
}