flatcad

flatcad is a framework designed for comparing various methods used for neural signed distance function training.

This repository contains:

• A pytorch implementation of a comparison framework, comparing state-of-the-art Neural-SDF reconstruction methods, like: DiGS, NeuralSingularHessian and NeurCADRecon.
• An official implementation of FlatCAD, as described in FlatCAD: Fast Curvature Regularization of Neural SDFs for CAD Models.
• A official implementation of finite difference losses, as described in A Finite Difference Approximation of Second Order Regularization of Neural-SDFs,

🔧 Setup

The repository uses uv dependency manager for environment setup and package installation, and relies on just for shortcut command execution. Please follow the tool specific installation guides.

To install all dependencies please run:

uv venv
source .venv/bin/activate
uv sync

To examine the setup please look at pyproject.toml

🚀 Getting Started

Start experimenting by running training for a selected method using the just command:

just run [method]

Available Methods: digs, nsh, nsh-fd, ncr, ncr-fd, odw or odw-fd.

Alternatively, you can run the training script manually:

python src/train.py --config ./configs/train_odw.yaml

using a .yaml config file from the ./configs directory:

📚 Citation

If you find this work useful in your research, please cite:

@article{Yin_2025,
  title={FlatCAD: Fast Curvature Regularization of Neural SDFs for CAD Models},
  volume={44},
  ISSN={1467-8659},
  url={http://dx.doi.org/10.1111/cgf.70249},
  DOI={10.1111/cgf.70249},
  number={7},
  journal={Computer Graphics Forum},
  publisher={Wiley},
  author={Yin, Haotian and Plocharski, Aleksander and Wlodarczyk, Michal Jan and Kida, Mikolaj and Musialski, Przemyslaw},
  year={2025},
  month=oct 
}

@article{yin2025finite,
  title={A Finite Difference Approximation of Second Order Regularization of Neural-SDFs},
  author={Yin, Haotian and Plocharski, Aleksander and Wlodarczyk, Michal Jan and Musialski, Przemyslaw},
  journal={arXiv preprint arXiv:2511.08980},
  year={2025}
}

@misc{yin2025schedulingoffdiagonalweingartenloss,
  title={Scheduling the Off-Diagonal Weingarten Loss of Neural SDFs for CAD Models}, 
  author={Haotian Yin and Przemyslaw Musialski},
  year={2025},
  eprint={2511.03147},
  archivePrefix={arXiv},
  primaryClass={cs.GR},
  url={https://arxiv.org/abs/2511.03147}, 
}

@misc{fan2025jointneuralsdfreconstruction,
  title={Joint Neural SDF Reconstruction and Semantic Segmentation for CAD Models}, 
  author={Shen Fan and Przemyslaw Musialski},
  year={2025},
  eprint={2510.03837},
  archivePrefix={arXiv},
  primaryClass={cs.GR},
  url={https://arxiv.org/abs/2510.03837}, 
}

🙏 Acknowledgement

• Credits to DiGS authors for sharing their method, which has been adapted for this framework:

  @inproceedings{ben2022digs,
  title={DiGS: Divergence guided shape implicit neural representation for unoriented point clouds},
  author={Ben-Shabat, Yizhak and Hewa Koneputugodage, Chamin and Gould, Stephen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19323--19332},
  year={2022}
  }

• Credits to NeuralSingularHessian authors for sharing their method, which has been adapted for this framework:

  @article{zixiong23neuralsingular,
  author = {Zixiong Wang, Yunxiao Zhang, Rui Xu, Fan Zhang, Pengshuai Wang, Shuangmin Chen, Shiqing Xin, Wenping Wang, Changhe Tu},
  title = {Neural-Singular-Hessian: Implicit Neural Representation of Unoriented Point Clouds by Enforcing Singular Hessian},
  year = {2023},
  journal = {ACM Transactions on Graphics (TOG)},
  volume = {42},
  number = {6},
  doi = {10.1145/3618311},
  publisher = {ACM}
  }

• Credits to NeurCADRecon authors for sharing their method, which has been adapted for this framework:

  @article{Dong2024NeurCADRecon,
  author={Dong, Qiujie and Xu, Rui and Wang, Pengfei and Chen, Shuangmin and Xin, Shiqing and Jia, Xiaohong and Wang, Wenping and Tu, Changhe},
  title={NeurCADRecon: Neural Representation for Reconstructing CAD Surfaces by Enforcing Zero Gaussian Curvature},
  journal={ACM Transactions on Graphics},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  year={2024},
  month={July},
  volume = {43},
  number={4},
  doi={10.1145/3658171},
  keywords = {CAD model, unoriented point cloud, surface reconstruction, signed distance function, Gaussian curvature}
  }

• Thanks to the SIREN codebase and the IGR/SAL codebases off whom original repos built upon.

⚖️ License

See LICENSE file.