Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, Hanwang Zhang
Nanyang Technological University, Hong Kong University of Science and Technology, Singapore University of Technology and Design, Skywork AI
# clone this repository
git clone https://github.com/Yuxuan-W/VcEdit.git
cd VcEdit
# build environment
conda create -n vcedit python=3.8 -y
conda activate vcedit
# Our version is based on CUDA 11.8, see https://pytorch.org/get-started/locally/ to install other PyTorch versions.
conda install pytorch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 pytorch-cuda=11.8 -c pytorch -c nvidia
# install requirements
pip install -r requirements.txt
To get started, please download the input examples and place the unpacked folder under VcEdit/gs_data.
The example data are from InstructN2N and DreamEditor, where the trained 3DGS model are generated by first running colmap and then trained using gaussian splatting. Please refer to the original gaussian splatting repository for more details about preparing your own data.
We provide several sample scripts for running the editing process. For example:
bash script/man_clown.sh
The editing process typically takes 10 to 25 minutes on a single A100 GPU, depending on the scene size and the number of iterations. The GPU memory usage ranges between 20GB and 40GB. For panorama samples the memory cost for processing all views (which is unnecessary w.r.t. the performance) is a bit higher than 40GB, you can remove some views in the dataset.
Our editing framework builds upon the image editing capabilities provided by InfEdit. Due to the diverse nature of editing scenarios, not all prompts yield satisfying results. Some prompts may result in editing failures or too drastic multi-view inconsistencies when using InfEdit. Before using this framework, we recommend testing your prompts in InfEdit to ensure:
- Prompt compatibility with the editing process.
- Optimal hyper-parameter selection for achieving desired results.
If you find our work helps, please cite our paper:
@inproceedings{wang2025view,
title={View-consistent 3d editing with gaussian splatting},
author={Wang, Yuxuan and Yi, Xuanyu and Wu, Zike and Zhao, Na and Chen, Long and Zhang, Hanwang},
booktitle={European Conference on Computer Vision},
pages={404--420},
year={2025},
organization={Springer}
}
Our code is based on these wonderful repos: