EvoCAD: Evolutionary CAD Code Generation with Vision Language Models

TL;DR

EvoCAD combines vision language models and evolutionary optimization to generate CAD code, significantly improving topological correctness.

cs.CV 🔴 Advanced 2025-10-14 8 views
Tobias Preintner Weixuan Yuan Adrian König Thomas Bäck Elena Raponi Niki van Stein
CAD evolutionary algorithm vision language model topological analysis generative AI

Key Findings

Methodology

EvoCAD generates CAD code using vision language models and evolutionary optimization. It starts by generating an initial set of CAD codes from user prompts, then evaluates them using vision and reasoning language models. Selected codes undergo crossover and mutation optimization, resulting in topologically correct CAD objects.

Key Results

  • EvoCAD achieved 87.2% topological correctness on the CADPrompt dataset, significantly outperforming 3D-Premise and CADCodeVerify.
  • In spatial metrics, EvoCAD's Hausdorff distance was 0.177, better than other methods.
  • Through evolutionary optimization, EvoCAD surpassed baseline methods after the second generation.

Significance

EvoCAD addresses the issue of topological correctness in CAD code generation by combining vision language models and evolutionary algorithms, enhancing efficiency in automated design and manufacturing.

Technical Contribution

EvoCAD is the first to apply evolutionary algorithms to CAD code generation, introducing new topological metrics for precise semantic similarity evaluation.

Novelty

EvoCAD is the first method to combine vision language models and evolutionary algorithms for CAD code generation, significantly improving topological correctness.

Limitations

  • EvoCAD performs limitedly on non-watertight objects as topological metrics cannot be computed.
  • The quality of initial CAD code generation depends on the performance of language models.

Future Work

Future research could explore EvoCAD's application to more complex CAD objects and the potential integration of other optimization algorithms.

AI Executive Summary

EvoCAD is a CAD code generation method that combines vision language models and evolutionary optimization to address shortcomings in existing methods regarding topological correctness. Using GPT-4V and GPT-4o for evaluation, EvoCAD performs exceptionally well on the CADPrompt dataset, significantly improving topological correctness and spatial metrics.

The method generates CAD code through evolutionary optimization, using vision language models for object description and reasoning language models for evaluation, ultimately producing CAD objects that align with user prompts. Experimental results show EvoCAD surpasses existing methods across multiple metrics, particularly in topological correctness.

EvoCAD's technical contributions include the first application of evolutionary algorithms to CAD code generation and the introduction of new topological metrics. Despite limitations on non-watertight objects, it holds broad application potential in automated design and manufacturing. Future research could explore its application to more complex objects.

Deep Analysis

Background

Computer-aided design (CAD) plays a crucial role in product development and engineering. Traditional methods rely on manual design and verification, which are inefficient. The emergence of generative AI tools offers new possibilities for CAD code generation, especially in research combining vision language models and evolutionary algorithms.

Core Problem

The core problem in CAD code generation is ensuring the topological correctness of generated objects. Existing methods often encounter errors when handling complex structures, affecting design accuracy and efficiency.

Innovation

EvoCAD introduces a new CAD code generation method by combining vision language models and evolutionary algorithms. This method improves topological correctness through evolutionary optimization and introduces new topological metrics, addressing shortcomings in existing methods.

Methodology

  • �� Initialization: Generate initial CAD code set from user prompts.
  • �� Evaluation: Use vision language models for object description and reasoning language models for evaluation.
  • �� Crossover and Mutation: Select suitable codes for crossover and mutation optimization.
  • �� Generation: Produce topologically correct CAD objects.

Experiments

Experiments were conducted using the CADPrompt dataset, comparing EvoCAD's performance with 3D-Premise and CADCodeVerify. Evaluations were performed using GPT-4V and GPT-4o, with multiple metrics set, including topological correctness and spatial metrics.

Results

EvoCAD achieved 87.2% accuracy in topological correctness, significantly outperforming baseline methods. In spatial metrics, EvoCAD's Hausdorff distance was 0.177, showing better geometric alignment performance.

Applications

EvoCAD can be applied in automated design and manufacturing, particularly in complex structure design requiring high topological precision. Its method can enhance design efficiency and reduce the need for manual verification.

Limitations & Outlook

EvoCAD performs limitedly on non-watertight objects, unable to compute topological metrics. The quality of initial CAD code generation depends on language model performance, potentially affecting final results.

Plain Language Accessible to non-experts

Imagine you're designing a complex LEGO model. EvoCAD acts like a smart assistant that automatically generates the design blueprint based on your description, ensuring each LEGO piece fits perfectly. It uses a method called evolutionary algorithm, similar to trying different LEGO combinations until finding the perfect design.

ELI14 Explained like you're 14

Imagine you're playing a design game where you need to create a perfect building model. EvoCAD acts like a super-smart game assistant that automatically generates the building design based on your instructions, ensuring every part fits perfectly. It uses a method called evolutionary algorithm, like trying different designs until finding the perfect one. Isn't that cool?

Glossary

Evolutionary Algorithm

An optimization algorithm that simulates the process of biological evolution, using selection, crossover, and mutation to optimize problems.

Used to optimize the CAD code generation process.

Vision Language Model

A model combining visual and language processing capabilities, able to understand and generate image descriptions.

Used to evaluate CAD object visual descriptions.

Topological Feature

Features describing the structure of an object, such as the number of holes.

Used to evaluate CAD object topological correctness.

Hausdorff Distance

A metric measuring the maximum distance between two sets.

Used to evaluate CAD object spatial alignment.

CAD Code

Code used to generate and describe three-dimensional design objects.

Representation of objects generated by EvoCAD.

Open Questions Unanswered questions from this research

  • 1 How to achieve topological metrics on non-watertight objects? Current methods cannot solve this, requiring new algorithms.
  • 2 How to improve the quality of initial CAD code? Language model performance limits generation effectiveness.

Applications

Immediate Applications

Automated Design

Designers can use EvoCAD to quickly generate CAD models of complex structures, enhancing design efficiency.

Long-term Vision

Intelligent Manufacturing

EvoCAD can be used in intelligent manufacturing systems to automatically generate product models that meet design requirements.

Abstract

Combining large language models with evolutionary computation algorithms represents a promising research direction leveraging the remarkable generative and in-context learning capabilities of LLMs with the strengths of evolutionary algorithms. In this work, we present EvoCAD, a method for generating computer-aided design (CAD) objects through their symbolic representations using vision language models and evolutionary optimization. Our method samples multiple CAD objects, which are then optimized using an evolutionary approach with vision language and reasoning language models. We assess our method using GPT-4V and GPT-4o, evaluating it on the CADPrompt benchmark dataset and comparing it to prior methods. Additionally, we introduce two new metrics based on topological properties defined by the Euler characteristic, which capture a form of semantic similarity between 3D objects. Our results demonstrate that EvoCAD outperforms previous approaches on multiple metrics, particularly in generating topologically correct objects, which can be efficiently evaluated using our two novel metrics that complement existing spatial metrics.

cs.CV cs.AI cs.NE