CAD-Recode
CAD-Recode is an open research model that predicts executable Python CadQuery scripts from 3D point clouds so engineers can inspect, edit, and reuse CAD logic recovered from scanned geometry.
CAD-Recode targets reverse engineering: it maps a point cloud into CadQuery source code instead of only a mesh, which keeps the output readable and editable in a text-based CAD workflow (University of Luxembourg and collaborators, 2025). The authors report large gains on public CAD benchmarks, including roughly ten times lower mean Chamfer distance than several prior baselines on the DeepCAD test set in their tables (Rukhovich et al., 2025).
Training relies on a large synthetic corpus of one million procedurally generated CadQuery programs, mixing sketch primitives and higher-level solids so the model sees realistic command patterns. That design choice matters for AEC and manufacturing teams who already capture sites and objects as point clouds and want structured CAD rather than a frozen mesh.
Because outputs are Python scripts, you can pair them with general coding assistants for edits or questions, as the project page notes in its discussion of downstream use. The work ships with open weights, code links, and a Hugging Face demo referenced from the official project site.
Expect a research-grade workflow: you must validate geometry against field measurements, standards, and your own QA rules before using recovered CAD in production deliverables. The model is not a turnkey scan-to-BIM substitute for buildings, but it is a concrete step toward interpretable CAD from captured data.
Specifications
Pricing
Platforms
Used for
Used by
Tasks
Pros and cons
Pros
- Strong reported accuracy on standard CAD benchmarks
- Open code and weights for experimentation
- Outputs editable scripts instead of opaque meshes only
Cons
- Research code path with manual setup
- Not tuned as an end-user AEC product
- Building-scale scans may need different handling than benchmark objects
Key features
Point cloud to CadQuery ??? Encodes a cloud with a lightweight encoder and decodes Python CAD programs you can run and edit.
Interpretable scripts ??? Uses full Python plus CadQuery rather than a fixed token soup, which helps human review.
Open artifacts ??? Project page links to arXiv, GitHub code, Hugging Face weights, dataset, and an interactive demo.
Pricing
Research release
Free
No vendor fee; follow repository licenses for usage.
Frequently asked questions
What does CAD-Recode output?
It produces CadQuery Python programs that reconstruct CAD solids from an input point cloud. You can execute the scripts in a Python environment with CadQuery installed, then inspect or change the code like any other program.
Is CAD-Recode free?
The project shares research code and model weights under an open academic release on GitHub and Hugging Face. Check the license files in those repositories for redistribution and commercial-use terms before you ship anything.
Can CAD-Recode replace commercial scan-to-BIM software?
No. It is a research model focused on CAD program recovery from clouds, not on architectural BIM objects, IFC export, or construction-ready modeling. Use it as a building block in custom pipelines, not as a full replacement for commercial scan-to-BIM tools.
Which point clouds does CAD-Recode support?
The public materials emphasize general 3D point clouds used in the paper benchmarks. Real building scans vary in noise, density, and scope, so expect to test preprocessing, cropping, and validation for your own capture hardware.
How accurate is CAD-Recode?
Authors report strong numeric gains on DeepCAD and Fusion360-style benchmarks in their paper tables. Always compare outputs against ground truth or manual CAD for your domain because real-site data can differ from academic sets.
Where can I try CAD-Recode?
The official project page lists a Hugging Face Space demo plus links to code and model checkpoints. Start there for interactive tests before installing locally.