2026 Joglekar Ortho2CAD Orthographic Drawings

Engineering design intent is often communicated through rasterized orthographic drawings, which need to be transformed into editable and parametric 3D computer-aided design (CAD) models for downstream applications. This paper introduces vision language model (VLM) frameworks designed to convert such drawings into CadQuery code, effectively enabling 3D CAD model generation. Given the lack of comprehensive datasets for orthographic drawings, we developed a pythonOCC-based drawing generator that produces over 1 million drawings from 813 CAD model datasets, along with a specialized set of 100 manually crafted drawings. Our experiments reveal that fine-tuning small open-source VLMs, when paired with CadQuery code, significantly enhances reconstruction accuracy in supervised settings. For datasets lacking code labels, we apply geometry-grounded reinforcement learning using solid intersection-over-union (IoU) metrics for reward optimization, leading to improved code validity and better cross-dataset generalization. Additionally, we present a self-refinement framework for front-end VLMs that iteratively repairs incorrect code and aligns generated 3D models with the original orthographic projections, achieving 100% valid code generation and a superior mean IoU performance across test sets, surpassing competing methods by more than 11%. These results demonstrate the potential of leveraging VLMs in the transition from orthographic drawings to 3D CAD models.