MGPC: Multimodal Network for Generalizable Point Cloud Completion With Modality Dropout and Progressive Decoding

This is the official repository of the paper "MGPC: Multimodal Network for Generalizable Point Cloud Completion With Modality Dropout and Progressive Decoding".

We have released the test code and data for now. The training code and dataset will be released after the paper is accepted.

We have tested on Ubuntu 20.04/22.04 with NVIDIA GeForce RTX 4090 with Python 3.10 and cuda12.1. The code may work on other systems.

• Setup a virtual environment

  python3.10 -m venv mgpc
  source mgpc/bin/activate

• Install pip dependencies

  Note that other versions of Pytorch may also work.

  cd MGPC
  pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121
  pip install -r requirements.txt

• Build extensions for Chamfer Distance and PointNet++

  cd extensions/Chamfer3D && python setup.py install
  cd ../pointnet2_ops_lib && python setup.py install
  cd ../..

• Download the datasets

  The test set of our MGPC_1M dataset can be downloaded from here.

• Download the model weight

  Our trained weights for 8192 and 2048 points are in here (containing pretrained weights of dinov2 and clip).

• Modify the configuration file

  Modify the arguments in the corresponding script. Specify all paths, batch size, and so on.

To evaluate the model on the test set, run:

python test.py

We provide some in-the-wild data in /demo, and you can also collect the data by yourselves. To test the zero-shot generalization, run:

python demo.py

Our code is partially built upon PoinTr, SeedFormer, TripoSR, and PUCRN. We thank them for their nicely open sourced code and their great contributions to the community.

@article{liu2026mgpc,
  title={MGPC: Multimodal Network for Generalizable Point Cloud Completion With Modality Dropout and Progressive Decoding},
  author={Liu, Jiangyuan and Ma, Hongxuan and Zhao, Yuhao and Liu, Zhe and Wang, Jian and Zou, Wei},
  journal={arXiv preprint arXiv:2601.03660},
  year={2026}
}