🛠️ Installation

# Clone the repository
git clone https://github.com/daidedou/diffumatch.git

Then install python>=3.9, pytorch>=2.5 with the method of your choice. I prefer conda (You can try with older version, it should work, as I did not use anything fancy. My guess is at least 1.8)

conda create -n diffumatch python=3.9

Run the command to install pytorch according to your version. For me it was:

conda activate diffumatch
conda install pytorch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 pytorch-cuda=12.4 -c pytorch -c nvidia

Then install the necessary packages

pip install -r requirements.txt

Note: In inference phase (training not affected), to be fast enough, our Zoomout implementation relies on Pykeops. Installation might fail from time to time (although it is much better since they switched to nvrtc). If you have problems with Pykeops, their github issues often have a solution for you! One of the common errors is ld cannot find -lnvrtc. The solution is to add cuda to library path:

export LIBRARY_PATH=$LIBRARY_PATH:/usr/local/cuda/lib

Befor launching any script/notebook.

---

Training the functional maps diffusion model

Download the Dynamic FAUST dataset here. Then, run

python preprocess_fmap_dataset.py --path dfaust_path

After that, run

python train_fmap.py

If you want to train a diffusion model on a different dataset, it should be possible if you save your shapes in a single folder and run the preprocess script with a different path and dataset argument. After that create a config file in config/diffuosion/custom_conf.yaml according to the dfaust_fmap template. The training is then:

python train_fmap.py --config custom_conf

The whole training relies of the EDM repository. Sorry if it is not compatible with your workflow.

Zero-shot shape matching

Matching two shapes

If you want to try the model, I suggest you to run the notebook, or alternatively the gradio demo script (may not be bug free). There is also a colab demo if you are brave enough to install the packages and debug (it is quite common that demos break after a single colab update).

python gradio_demo.py --config config/matching/sds.yaml

If you want to test your own trained diffusion model, modify the config lines:

diff_model:
  train_dir: pretrained

to

diff_model:
  train_dir: fmap_exps
  diff_num_exp: num_exp

where num_exp is the number of your experiences in the fmap_exps folder (can be custom too).

Shape Non-rigid Kinematics (SNK)

We also provide a (non-official) reproduction of Shape Non-rigid Kinematics (SNK), and you can try it in snk_demo.ipynb or use snk.yaml config

Paper experiments

Read experiments.md

📜 License

This project is licensed under the Apache Licence.

---

Acknowledgements

Thanks to amazing deep functional functional maps repositories that were very useful for us (non-exhaustive):

• AttentiveFMaps
• Robust shape matching
• SimplifiedFmaps

And also

• Pyfmaps
• Potpourri3d
• EDM

Citation

If you found our work useful, please cite

@article{pierson2025diffumatch,
  title={DiffuMatch: Category-Agnostic Spectral Diffusion Priors for Robust Non-rigid Shape Matching},
  author={Pierson, Emery and Li, Lei and Dai, Angela and Ovsjanikov, Maks},
  journal={arXiv preprint arXiv:2507.23715},
  year={2025}
}