README.md

May 6, 2026 · View on GitHub

MVGS: Multi-View Regulated Gaussian Splatting for Novel View Synthesis

If you like our project, please give us a star ⭐ on GitHub for latest update.

Xiaobiao Du123 · Yida Wang3 · Xin Yu2

1The University of Technology Sydney · 2The University of Queensland · 3Li Auto Inc.

🌐Project Page | 🖨️ArXiv | 📰Paper

😮 Highlights

MVGS can be easily embeded into existing Gaussian-based explict representation methods for better novel view synthesis results. We transform a traditional training paradigm only constrained by a single-view supervision per training iteration with our proposed multi-view training.

🔥 Four key Contributions

  • Multi-view constraint loss: We first propose to supplement multi-view constraints during optimization towards the 3D Gaussian attributes. As a general solution, we improve the overall accuracy in a variety of scenarios and different Gaussian-based methods.
  • Cross-intrinsic Guidance method: training in a coarse-to-fine manner with regard to different resolutions.
  • Cross-ray Densification: densifying more 3D Gaussians in ray-intersect regions.
  • Multi-view Augmented Densification Strategy: to enhance the effect of densification while each view is distinct dramatically.

🚩 Updates

Welcome to watch 👀 this repository for the latest updates.

✅ [2024.10.4] : Release project page.

✅ [2024.10.4] : Code Release.

Installation

  1. Clone this repo:
git clone https://github.com/xiaobiaodu/MVGS.git --recursive
cd MVGS
  1. Install dependencies (Skip, if you've installed environment for 3DGS.)
conda env create --file environment.yml
conda activate mvgs

Quickstart

Public Data

The MipNeRF360 scenes are provided by the paper's author and can be accessed here.

Custom Data

For custom data, process the image sequences using Colmap to obtain SfM points and camera poses.

Run

Below code requires at least 40G GPU memory.

python run_360.py

If your GPU memory is limited, try decrease mv value like below code:

python train.py -s  {data_path}   --eval --white_background  -m {save_path}  --mv  6

Interactive Viewer

Please see the original 3DGS repository. Our method can be seemlessly integrated in the original 3DGS

✏️ Citation

If you find our paper and code useful in your research, please consider giving a star :star: and citation :pencil:.

@misc{du2026mvgs,
      title={MVGS: Multi-view Regulated Gaussian Splatting for Novel View Synthesis}, 
      author={Xiaobiao Du and Yida Wang and Xin Yu},
      year={2026},
      eprint={2410.02103},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.02103}, 
}