CADBench: 18K-Sample Multimodal Benchmark for CAD Program Generation
June 19, 2026
CADBench provides 18,000 evaluation samples across six dataset families, five input modalities (clean mesh to photorealistic renders), and six metrics covering geometric fidelity and executability. Eleven CAD-specialized and general VLMs were benchmarked, generating over 1.4 million CAD programs.
HOW THIS AFFECTS YOU
●
builderYou can use CADBench to evaluate AI-assisted CAD tools against a standardized suite before shipping.
●
researcherUnified evaluation across modalities and complexity levels fills a real gap for comparing CAD generation models systematically.