CADEvolve: Creating Realistic CAD via Program Evolution
Maksim Elistratov, Marina Barannikov, Gregory Ivanov, Valentin Khrulkov, Anton Konushin, Andrey Kuznetsov, Dmitrii Zhemchuzhnikov
2026-02-19
Summary
This paper focuses on improving how computers can automatically create 3D models, a process called Computer-Aided Design or CAD. It tackles the challenge of teaching AI to design complex shapes, going beyond simple designs.
What's the problem?
Currently, training AI for CAD is difficult because the available data isn't good enough. Most existing datasets only show how to make very basic shapes, like stretching a 2D sketch into a 3D object. They lack the complexity of real-world designs, don't show how to combine multiple operations, and don't capture the designer's original *intent* behind the design. Existing AI models also struggle to understand 3D shapes well enough to create valid designs.
What's the solution?
The researchers developed a system called CADEvolve. It starts with simple 3D shapes and gradually makes them more complex. An AI model guides the process, suggesting edits and checking if the changes are valid. This creates a large dataset of 1.3 million CAD scripts, along with images of the resulting 3D models. They then used this dataset to train a new AI model specifically for converting images into CAD designs.
Why it matters?
This work is important because it allows AI to create much more sophisticated 3D models than before. The new AI model performs better than previous ones on standard tests, meaning it's a significant step towards fully automating the CAD process. This could speed up engineering and manufacturing, allowing for faster creation of new products and designs.
Abstract
Computer-Aided Design (CAD) delivers rapid, editable modeling for engineering and manufacturing. Recent AI progress now makes full automation feasible for various CAD tasks. However, progress is bottlenecked by data: public corpora mostly contain sketch-extrude sequences, lack complex operations, multi-operation composition and design intent, and thus hinder effective fine-tuning. Attempts to bypass this with frozen VLMs often yield simple or invalid programs due to limited 3D grounding in current foundation models. We present CADEvolve, an evolution-based pipeline and dataset that starts from simple primitives and, via VLM-guided edits and validations, incrementally grows CAD programs toward industrial-grade complexity. The result is 8k complex parts expressed as executable CadQuery parametric generators. After multi-stage post-processing and augmentation, we obtain a unified dataset of 1.3m scripts paired with rendered geometry and exercising the full CadQuery operation set. A VLM fine-tuned on CADEvolve achieves state-of-the-art results on the Image2CAD task across the DeepCAD, Fusion 360, and MCB benchmarks.