This repository supports our NeurIPS submission on generating CadQuery-based 3D models from natural language, building on the foundations of Text2CAD and DeepCAD. It includes data annotation, model training, inference, and evaluation pipelines across six open-source LLMs.

• data_annotation/
  Scripts for annotating CAD sequences using Gemini 2.0 Flash on top of the Text2CAD dataset. The full annotated dataset is available here: CadQuery.zip

• train/
  Training scripts for six open-source models (CodeGPT, Gemma, GPT-2, Mistral, Qwen).
  Finetuned models are available on HuggingFace.

• inference/
  Step-by-step pipeline for evaluating the finetuned models:

  • step1_generate_CadQuery: Use finetuned models to generate CadQuery code from natural language prompts.
  • step2_clean_run_CadQuery: Extract valid Python code from model outputs and execute it to generate STL files.
  • step3_rendering: Render STL files using Blender.
  • step4_gemini_eval: Evaluate rendered 3D models using Gemini 2.0 Flash.
  • step5_compute_metrics: Compute quantitative metrics such as Chamfer Distance and other geometric similarity scores.

We gratefully acknowledge the authors of Text2CAD and DeepCAD for their foundational contributions and datasets.

• Text2CAD
  Mohammad Sadil Khan*, Sankalp Sinha*, Talha Uddin Sheikh, Didier Stricker, Sk Aziz Ali, Muhammad Zeshan Afzal
  Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts

• DeepCAD
  Rundi Wu, Chang Xiao, Changxi Zheng
  DeepCAD: A Deep Generative Network for Computer-Aided Design Models