GeoCAD-LLM: CAD Sequence Generation via Multimodal LLMs with Equivariant Geometric Features🛠️

• Github
• Paper

GeoCAD-LLM_4B_pc🛠️

• Base_model: Qwen3-4B-Instruct
• Max sequence length: 8,192
• Epoch: 2
• Learning rate: 1e-4
• Batch size: 128
• This model specialized for pc-text-to-CAD. However, it also supports multi-modality.

GeoCAD-LLM Contributions🔥

• State-of-the-art Performace🏆 in Text2CAD datasets. (as shown in below Tables)
• Multimodal CAD Generation🌐: Both text-to-CAD and pc-text-to-CAD.
• GeoCAD-LLM directly generate CAD vector sequence as natural language.
• Novel Two Stage Training Pipeline🧭: In stage1, training semantic geometry alignment. In stage2, training fine-grained geometry. Especially, we direct levearge E(3)-equivariant features for geomtry-consistent supervision, inherently ensuring geometric feature consistency regardless of input orientation.
• Apply Point Cloud Dropout (PCD) technique🧶: PCD mitigates over-reliance on geometric inputs and improves multimodal generalization. Also, it is a critical training technique for multimodal CAD generation.

Performace (text-to-CAD & pc-text-to-CAD)🔥

Qualitative Results

Bibtex🤗