AutoBrep: Autoregressive B-Rep Generation with Unified Topology and Geometry (SIGGRAPH Asia 2025)

Xiang Xu*, Pradeep Jayaraman*, Joseph Lambourne*, Yilin Liu, Durvesh Malpure, Pete Meltzer

AutoBrep is a decoder-only Transformer model that autoregressively generates B-Rep geometry and topology tokens following a BFS order of the B-Rep topology graph. Geometric information is tokenized as bounding boxes paired with encoded UV-grid shape codes. Topological structure is represented via a special face identifier that maps face–edge adjacencies into reference tokens.

Installing environments

cd core
conda env create -f dev-env.yaml
conda activate autobrep

Inference

sh scripts/sample.sh

Download the pretrained checkpoints

Modify configs/sample.json for different sampling parameters.

complexity can be random, easy, medium, or hard.

Training

sh scripts/train.sh

Download the deduplicated ABC-1M dataset

Modify configs/autobrep.yaml with the path to ABC parquets and pretrained FSQ checkpoints.

Citation

@inproceedings{xu2025autobrep,
  title={AutoBrep: Autoregressive B-Rep Generation with Unified Topology and Geometry},
  author={Xu, Xiang and Jayaraman, Pradeep and Lambourne, Joseph and Liu, Yilin and Malpure, Durvesh and Meltzer, Pete},
  booktitle={Proceedings of the SIGGRAPH Asia 2025 Conference Papers},
  pages={1--12},
  year={2025}
}