Clone our repo first:
git clone https://github.com/huanngzh/MV-Adapter.git
Create a new conda env:
conda create -n Pro3DEditor python=3.10
conda activate Pro3DEditor
Install necessary packages:
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
cd ./MV-Adapter
pip install -r requirements.txt
cd ../GaussianObject
pip install -r requirements.txt
cd ./models
python download_hf_models.py
Install necessary models:
- huanngzh/mv-adapter
- stabilityai/stable-diffusion-xl-base-1.0
- madebyollin/sdxl-vae-fp16-fix
- ZhengPeng7/BiRefNet
- clip-vit-large-patch14
cd ../..
bash ./huggingface/hf_download_model.sh huanngzh/mv-adapter
bash ./huggingface/hf_download_model.sh stabilityai/stable-diffusion-xl-base-1.0
bash ./huggingface/hf_download_model.sh madebyollin/sdxl-vae-fp16-fix
bash ./huggingface/hf_download_model.sh ZhengPeng7/BiRefNet
bash ./huggingface/hf_download_model.sh openai/clip-vit-large-patch14
In this case, our initial input is an object represented in 3DGS, specifically the file dust3r_4.ply located in ./GaussianObject/data/render_keyviews/rccar. In the same directory, camera.json contains the camera extrinsics of six key views, cfg_args is the configuration file for 3DGS initialization, and sparse_6.txt contains the indices of the views.
cd ./GaussianObject
bash render.sh
The rendering results are saved in ./data/render_keyviews/rccar/train/ours_None/renders
In the previous step, we render six key views of the 3D object. Then, we use these key views for fine-tuning the multi-view diffusion model. After that, we will use the fine-tuned multi-view diffusion model to edit the key views.
cd ../MV-Adapter
cp ../GaussianObject/data/render_keyviews/rccar/train/ours_None/renders/* ./dataset/rccar
bash quick_start.sh
The edited key views are saved in ./results/images
In the previous step, we finish editing all the key views. Next, we will edit the entire 3D object.
cd ../GaussianObject
cp -r ../MV-Adapter/results/* ./realcap/cat_rccar
bash edit.sh cat_rccar
The edited 3D object is saved in ./output/gaussian_object/rccar/save
Our code is based on MV-Adapter and GaussianObject. Thanks for their works.
If you find this work helpful, please consider citing our paper:
@article{zheng2025pro3d,
title={Pro3D-Editor: A Progressive-Views Perspective for Consistent and Precise 3D Editing},
author={Zheng, Yang and Huang, Mengqi and Chen, Nan and Mao, Zhendong},
journal={arXiv preprint arXiv:2506.00512},
year={2025}
}