Original title: RealCADBench: Benchmarking Parametric CAD Modeling from Industrial Design Intents
JoyIndustrial VisCAD Team, Linxin Cai, Qiuhe Hong, Zhichao Huang +9 · September 2026
Designing 3D objects with computer-aided design software is hard to measure with just one test. Older tests rely on fake data, simple inputs, or only check if the code runs. To solve this, researchers built RealCADBench. It tests artificial intelligence models on real industrial design goals. The system takes text, engineering drawings, real product photos, or rendered pictures and asks models to write Python code for FreeCAD. A runtime program then runs this code to build the 3D shape. The researchers tested nine top models and found that no single model wins across all four tests: code execution, shape overlap, surface overlap, and visual identity. Across six large models on single parts, code execution ranged from 0.565 to 0.812. The models frequently failed by missing fine details, losing a part's identity, or placing assembly pieces incorrectly. Just checking if code runs does not tell you if the 3D model is actually right.
Imagine building a Lego set from a picture of the finished product. You might get the big bricks right but forget small connectors or put pieces in the wrong place. In this paper, AI writes code to build digital models, but it often misses tiny details or misplaces parts, showing that just making the code run isn't enough; the model must also be precise.
This benchmark helps engineers test if AI can handle real factory design tasks. By checking more than just code execution, it pushes for AI that produces complete and accurate models.
This work only evaluates existing AI models and does not propose new ones to fix the failure modes. It is limited to the specific tasks and metrics defined in the benchmark, which may not cover all types of CAD modeling.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
RealCADBench: Benchmarking Parametric CAD Modeling from Industrial Design Intents — JoyIndustrial VisCAD Team, Linxin Cai, Qiuhe Hong et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.