---

Introduction

CADFit is a hybrid optimization-based framework that recovers complex, editable CAD construction sequences from watertight meshes by incrementally fitting and validating parametric operations using geometric feedback. Rather than predicting programs in a single forward pass, CADFit formulates reconstruction as an explicit IoU-driven optimization over executable CAD programs, supporting a rich operator set (extrusions, revolutions, fillets, chamfers) composed using Boolean union and cut.

---

Key Features

• Rich operator set. Extrude, Revolve, Fillet, Chamfer, Union, Cut — far beyond sketch-and-extrude pipelines.
• IoU-driven optimization, kernel-validated. Every candidate is built through CadQuery; the optimizer maximizes the same volumetric IoU it reports. Invalid Ratio = 0 on every benchmark slice.
• Compact construction sequences. Backward marginal pruning produces shorter, cleaner programs than the originals, giving higher-quality supervision for downstream learning models.
• Iterative residual refinement. Positive and negative residuals are reconstructed with the same pipeline and composed with Union / Cut, until residual volume falls below tolerance.
• Modality-agnostic. The same algorithm handles meshes from STL exports, point clouds, or images-to-3D models (we use Hunyuan3D) — no retraining.

---

Qualitative Mesh-to-CAD Results

DeepCAD, Fusion360, and ABC (easy/medium/hard). CADFit produces a valid, editable CAD program on every input where baselines either output Invalid or visibly diverge from the target geometry.

---

Multimodal Image-to-CAD

CADFit is agnostic to where the mesh comes from. Pairing it with a pretrained image-to-3D model — we use Hunyuan3D — and standard post-processing (watertight enforcement, Taubin smoothing, mesh decimation) gives a fully end-to-end image-to-CAD pipeline.

---

Example CadQuery Output

Every CADFit run emits an executable Python script that reproduces the solid in CadQuery:

---

Quick Start

conda create -n cadfit-env python=3.10 && conda activate cadfit-env
pip install -r requirements.txt

Reconstruct every STL in a folder:

python run_pipeline.py <input_folder>

Outputs land in <input>_runs/<stl_id>/ (override with --output-folder OUT):

<input>_runs/<stl_id>/
├── best_greedy_parallel_iterative.stl   # final reconstruction
├── best_greedy_parallel_iterative.py    # equivalent CadQuery program
└── final_iou.json                       # { final_iou, duration, ... }

Common flags:

For best throughput on a multi-core box, raise the inner worker pool:

CADFIT_INNER_WORKERS=25 python run_pipeline.py abc_hard

Call from Python directly:

from run_pipeline import process_single_stl
process_single_stl("abc_hard/00013080.stl",
                    folder_name="abc_hard_runs",
                    max_iterations=1)

---

Citation

If you find our work helpful, please consider citing:

@article{nehme2026cadfit,
  title={CADFit: Precise Mesh-to-CAD Program Generation with Hybrid Optimization},
  author={Nehme, Ghadi and Whalen, Eamon and Ahmed, Faez},
  journal={arXiv preprint arXiv:2605.01171},
  year={2026}
}

---

License

Released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license — free for academic and research use; see LICENSE for full terms.

A provisional patent application has been filed covering the CADFit method. Commercial licensing inquiries: ghadi@mit.edu.

---

Contact

For questions, issues, or collaboration, please contact: ghadi@mit.edu

---