CADReasoner

CADReasoner is a vision–language model for iterative CAD reverse engineering. It generates a runnable CadQuery program and refines it over multiple iterations using geometric feedback from the discrepancy between the target and the current prediction.

Accepted to CVPR 2026 Findings Track
Paper: https://arxiv.org/abs/2603.29847
Code: https://github.com/zhemdi/CADReasoner
HF paper page: https://huggingface.co/papers/2603.29847

Overview

Unlike single-pass Image2CAD methods, CADReasoner performs iterative program editing. At each step, it uses multi-view geometric inputs together with the previous prediction to improve reconstruction quality.

Usage

Please use the inference and evaluation scripts from the repository.

Example:

python3 test.py --dataset <test_dataset> --checkpoint kulibinai/CADReasoner --n_iters <n_iters> --outdir <outdir>

<test_dataset> can be one of:

• maksimko123/deepcad_test_mesh
• maksimko123/fusion360_test_mesh
• kulibinai/mcb_test
• kulibinai/deepcad_test_scan
• kulibinai/fusion360_test_scan
• kulibinai/mcb_test_scan

Citation

@article{kabisov2026cadreasoner,
  title={CADReasoner: Iterative Program Editing for CAD Reverse Engineering},
  author={Kabisov, Soslan and Kirichuk, Vsevolod and Volkov, Andrey and Savrasov, Gennadii and Barannikov, Marina and Konushin, Anton and Kuznetsov, Andrey and Zhemchuzhnikov, Dmitrii},
  journal={arXiv preprint arXiv:2603.29847},
  year={2026}
}