Prompt-to-Geometry: Exploring the Translation of Natural Language into Real Generative Design 

Voiceover

Student Presenter(s)(s)

Aidan Oh, Andrei Vince

Abstract or Description

Our research explores the use of large language models (LLMs) to generate procedural architectural geometry from natural language prompts. The project investigates how text-based instructions can be translated into Python scripts that drive 3D design tools such as Rhino and its parametric plugin, Grasshopper. We began with foundational research into cellular automata (rule-based systems that generate complex patterns from simple inputs) and the role of artificial intelligence in architectural design. Building on this foundation, we developed custom scripts to generate diverse architectural typologies—ranging from houses and cellular-automata-based towers to New York-style skyscrapers and Suprematist-inspired forms. These scripts use parametric logic to create unique and customizable geometries.

A central focus of the project is automating the generation of multimodal datasets for training and evaluating generative AI models in architecture—bridging the gap between code, language, imagery, and 3D geometry. Our final product is a structured dataset that includes prompts, parameter files (stored in JSON format), and images generated through an automated workflow. This pipeline produces random architectural structures in Rhino3D, captures and renders them realistically, and then uses the images to generate natural language prompts capable of describing and regenerating those designs.

This work contributes a replicable framework for building text-to-geometry datasets in architecture and points toward a future where AI tools can meaningfully participate in early-stage design workflows, enabling more iterative, multimodal, and collaborative design processes.

Mentor

Sabrei Gokmen