Drawing-Recode turns raster CAD scans into parametric code with LLMs

Researchers from industry and academia, including Mingi Kim, Yongjun Kim, and Hyungki Kim, introduced Drawing-Recode, a framework that automatically generates parametric CAD sequences from raster-format 2D drawings. The problem is critical for digitizing pre-digital engineering drawings, enabling part reproduction and manufacturing automation. Prior methods either required vector inputs or ignored the connection between dimensional annotations (like measurements and tolerances) and the visual geometry itself. Drawing-Recode explicitly bridges this gap by using an image encoder to extract geometric features and a separate text recognition module to capture annotations. A cross-attention mechanism then grounds these annotations to the correct geometric elements, reinforced by a novel Annotation Grounding Loss (AGL). The fused features are fed into a Large Language Model (LLM) that outputs code in Structured Parametric CAD Code (SPCC), a format designed to represent parametric CAD sequences.

Experiments demonstrate that Drawing-Recode significantly outperforms existing baselines on standard datasets. More importantly, it remains robust when applied to scanned drawings that resemble real industrial conditions, where noise, distortion, and imperfect text recognition are common. This robustness is a key step toward practical deployment in manufacturing workflows, where legacy raster drawings are still widely used. The framework's ability to recover editable, parametric CAD code from static images could reduce the manual effort required for part reproduction, enable automated design-to-manufacturing pipelines, and help companies unlock value from decades of archived drawings. The paper is available on arXiv under ID 2607.27558.