CAD-Recode: Reverse Engineering CAD Code from Point Clouds

Danila Rukhovich1, Elona Dupont1, Dimitrios Mallis1, Kseniya Cherenkova1,2, Anis Kacem1, Djamila Aouada1
1SnT, University of Luxembourg, 2Artec3D
ICCV 2025
arXiv Code Huggingface Model Huggingface Dataset Huggingface Demo

TL;DR

CAD-Recode is the first model to translate 3D point clouds into executable Python CAD scripts, combining a novel output format (CadQuery code), a lightweight Qwen2-1.5B-based architecture with a single-layer point cloud encoder, and a procedurally generated dataset of 1 million CAD programs - setting a new standard in interpretable 3D reverse engineering.
Sheme

CAD Representation as Code

Unlike DeepCAD's closed and limited token vocabulary, CAD-Recode leverages the full expressiveness of Python with the CadQuery library, enabling rich, interpretable, and modular representations of CAD programs that mirror real-world design logic and support human editing, inspection, and reuse.
CadQuery

Procedurally Generated Dataset

To train CAD-Recode, we introduce a dataset of 1 million procedurally generated Python CadQuery programs, featuring both low-level sketch primitives—like lines, circles, and arcs—and higher-level geometric abstractions such as rectangles, boxes, and cylinders.
CAD-Recode Dataset

Results

While previous methods often fail to restore the general shape, CAD-Recode outputs only slightly deviate from ground truth in most cases.
qualitative results
Trained solely on a procedurally generated dataset, CAD-Recode significantly outperforms prior methods on DeepCAD and Fusion360 benchmarks, achieving 10× lower Chamfer Distance and over 20% higher IoU.
Method Train Dataset DeepCAD Test Set Fusion360 Test Set
Name Size Mean CD↓ Med. CD↓ IoU↑ IR↓ Mean CD↓ Med. CD↓ IoU↑ IR↓
DeepCADDeepCAD160k42.59.6446.77.133089.239.925.2
PrismCADDeepCAD127k–4.2872.116.2–4.7565.318.0
Point2CylDeepCAD35k–4.2773.83.9–4.1867.53.2
HNC-CADDeepCAD125k–8.6465.35.6–36.863.57.3
MultiCADDeepCAD160k–8.09–11.5–42.2–16.5
TransCADDeepCAD140k32.34.5165.51.178.633.460.22.4
CAD-DiffuserDeepCAD160k–3.0274.31.5–3.6263.31.7
CAD-SIGNetDeepCAD160k3.430.2877.60.97.374.0870.41.1
CAD-RecodeDeepCAD160k0.890.2086.20.01.770.3075.60.0
CAD-RecodeOurs1M0.300.1692.00.40.350.1587.80.5

Given the code output from CAD-Recode and a generic prompt, GPT-4o allows automated and interactive editing of the CAD model or CAD question answering on SGP-Bench.
editability

BibTeX

@article{rukhovich2025cad-recode,
  author    = {Danila Rukhovich, Elona Dupont, Dimitrios Mallis, Kseniya Cherenkova, Anis Kacem, Djamila Aouada},
  title     = {CAD-Recode: Reverse Engineering CAD Code from Point Clouds},
  journal   = {ICCV},
  year      = {2025},
}