# Hierarchical Neural Coding for Controllable CAD Model Generation

**Authors:** Xiang Xu, Pradeep Kumar Jayaraman, Joseph G. Lambourne, Karl D. D. Willis, Yasutaka Furukawa  
**arXiv:** [2307.00149](https://arxiv.org/abs/2307.00149)  
**Date:** 2023-06-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.00149v1/

## Key points

- Groundbreaking paper introducing a novel generative model for Computer Aided Design (CAD)
- Model represents high-level design concepts as a three-level hierarchical tree of neural codes
- Unprecedented flexibility and precision in CAD design by controlling generation through specifying target design using a code tree
- Implementation of vector quantized VAE with "masked skip connection" to extract design variations at three levels
- Use of two-stage cascaded auto-regressive transformers to generate code trees from incomplete CAD models and complete them according to intended design
- Superior performance on conventional tasks and innovative interaction capabilities on conditional generation tasks demonstrated in experiments
- Advancements in learning-based methods enabling generation of engineering sketches and solid models, parsed with solid modeling kernel for editable parametric CAD files
- Gap in providing automated system for specifying parameters and equations within CAD models, highlighting need for more efficient tools
- Hierarchical Neural Coding for Controllable CAD Model Generation enhances efficiency and precision of CAD modeling through advanced generative techniques

## Summary

Hierarchical Neural Coding for Controllable CAD Model Generation is a groundbreaking paper that introduces a novel generative model for Computer Aided Design (CAD). The model represents high-level design concepts of a CAD model as a three-level hierarchical tree of neural codes, ranging from global part arrangement to local curve geometry. By controlling the generation or completion of CAD models through specifying the target design using a code tree, this model offers unprecedented flexibility and precision in CAD design. The paper details the implementation of a vector quantized VAE with "masked skip connection" to extract design variations as neural codebooks at three levels. Additionally, two-stage cascaded auto-regressive transformers are employed to learn how to generate code trees from incomplete CAD models and then complete them according to the intended design. Extensive experiments showcased superior performance on conventional tasks such as random generation while also enabling innovative interaction capabilities on conditional generation tasks. Furthermore, recent advancements in learning-based methods leveraging CAD modeling sequence history and sketch constraints have enabled the generation of engineering sketches and solid models. These generated sequences can be parsed with a solid modeling kernel to obtain editable parametric CAD files containing 2D engineering sketches or 3D CAD shapes. This progress opens up new possibilities for modifying existing CAD models while constraining certain properties or auto-completing designs interactively. Despite these advancements, there remains a gap in providing designers with an automated system for specifying parameters and equations that define profile positions and sizes within CAD models. The current industry standard still relies heavily on manual input, highlighting the need for more efficient and user-friendly tools in the field of computer-aided design. In conclusion, Hierarchical Neural Coding for Controllable CAD Model Generation represents a significant step forward in enhancing the efficiency and precision of CAD modeling through advanced generative techniques. The integration of neural codes and transformer networks opens up exciting possibilities for automating complex design processes while empowering designers with greater control over their creations.