Reproduce “FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models” on OpenResearch

About this paper

Zhanwei Zhang, Shizhao Sun, Wenxiao Wang, Deng Cai, Jiang Bian

FLEXCAD introduces a unified framework leveraging fine-tuned Large Language Models for controllable CAD model generation, enabling precise modifications across all construction hierarchies from sketch-extrusion to individual curves. Developed by researchers from Microsoft Research Asia and Zhejiang University, the approach significantly improves the validity and realism of generated CAD models compared to existing methods.

Key insights

• Large Language Models can effectively learn and generate complex hierarchical CAD structures when CAD models are represented as structured text with explicit hierarchy tokens.
• A unified, hierarchy-aware masking strategy enables a single LLM to perform diverse controllable CAD generation tasks across all construction levels, from coarse sketch-extrusions to fine-grained curves.
• Leveraging the extensive pre-training of LLMs via efficient fine-tuning methods like LoRA is critical for achieving high-quality and valid CAD outputs, outperforming training from scratch.

Summary by alphaXiv.