# Text2CAD: Generating Sequential CAD Models from Beginner-to-Expert Level Text Prompts

**Authors:** Mohammad Sadil Khan, Sankalp Sinha, Talha Uddin Sheikh, Didier Stricker, Sk Aziz Ali, Muhammad Zeshan Afzal  
**arXiv:** [2409.17106](https://arxiv.org/abs/2409.17106)  
**Date:** 2024-09-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.17106v1/

## Key points

- Text2CAD is an AI framework that generates text-to-parametric CAD models using user-friendly instructions.
- The framework utilizes Mistral and LLaVA-NeXT to create text prompts based on natural language instructions from the DeepCAD dataset.
- An end-to-end transformer-based auto-regressive network is proposed within the Text2CAD framework for generating parametric CAD models from input texts.
- Performance evaluation metrics include visual quality, parametric precision, and geometrical accuracy, showcasing the potential of the framework in AI-aided design applications.
- Expert-level instructions (L3) are included in annotations for users requiring precise geometric descriptions and relative measurements for CAD modeling tasks.
- The Text2CAD transformer architecture autonomously deduces all intermediate design steps to transform natural language descriptions into 3D CAD models.
- Experimental analysis demonstrates superior performance compared to traditional two-stage baseline methods commonly used in similar tasks.

## Summary

, , , , 

In the realm of Computer-Aided Design (CAD), prototyping complex models can be a time-consuming task due to the lack of intelligent systems for generating simpler intermediate parts. To address this challenge, we present Text2CAD, an innovative AI framework that generates text-to-parametric CAD models using user-friendly instructions suitable for designers of all skill levels. Our framework introduces a data annotation pipeline that utilizes Mistral and LLaVA-NeXT to create text prompts based on natural language instructions from the DeepCAD dataset, which includes approximately 170,000 models and 660,000 text annotations ranging from abstract CAD descriptions to detailed specifications. Within the Text2CAD framework, we propose an end-to-end transformer-based auto-regressive network that can generate parametric CAD models from input texts. Our model's performance is evaluated based on various metrics such as visual quality, parametric precision, and geometrical accuracy. The results demonstrate the potential of our framework in AI-aided design applications. Additionally, we highlight the inclusion of expert-level instructions (L3) in our annotations for users who require precise geometric descriptions and relative measurements for their CAD modeling tasks. By generating multi-level instructions over a span of 10 days, we ensure accuracy and reduce the likelihood of hallucinations often associated with minimal metadata approaches. The Text2CAD transformer architecture is specifically designed to transform natural language descriptions into 3D CAD models by deducing all intermediate design steps autonomously. Through our experimental analysis, we showcase superior performance compared to traditional two-stage baseline methods commonly used in similar tasks. In conclusion, this paper presents Text2CAD as a groundbreaking AI framework for generating parametric 3D CAD models through textual descriptions. We provide insights into our data annotation pipeline leveraging both Language Models (LLMs) and Vision-Language Models (VLMs), introduce an end-to-end transformer-based autoregressive architecture for CAD model generation from text prompts, discuss related work in the CAD domain, present experimental results demonstrating the effectiveness of our approach, acknowledge limitations within our framework, and conclude with future research directions.