Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Design through Quality-Diversity Data Synthesis and Language Models

Adam Gaier, James Stoddart, Lorenzo Villaggi, Shyam Sudhakaran

TL;DR

Problem: Generative design in engineering requires diverse, high-quality data and strict constraint adherence. Approach: Synthesize diverse, high-performing data via Quality-Diversity (MAP-Elites), fine-tune a language model on that data to generate high-level designs, and refine them into constraint-satisfying layouts with Wave Function Collapse (WFC). Findings: Datasets built with QD substantially improve fidelity to natural-language prompts and overall validity compared with randomly sampled data, enabling reliable, text-guided architectural layouts via TileGPT. Significance: This data-efficient, interactive framework merges evolutionary data generation with symbolic constraint solving to make generative design more accessible and controllable in architecture and related engineering domains.

Abstract

Two fundamental challenges face generative models in engineering applications: the acquisition of high-performing, diverse datasets, and the adherence to precise constraints in generated designs. We propose a novel approach combining optimization, constraint satisfaction, and language models to tackle these challenges in architectural design. Our method uses Quality-Diversity (QD) to generate a diverse, high-performing dataset. We then fine-tune a language model with this dataset to generate high-level designs. These designs are then refined into detailed, constraint-compliant layouts using the Wave Function Collapse algorithm. Our system demonstrates reliable adherence to textual guidance, enabling the generation of layouts with targeted architectural and performance features. Crucially, our results indicate that data synthesized through the evolutionary search of QD not only improves overall model performance but is essential for the model's ability to closely adhere to textual guidance. This improvement underscores the pivotal role evolutionary computation can play in creating the datasets key to training generative models for design. Web article at https://tilegpt.github.io

Generative Design through Quality-Diversity Data Synthesis and Language Models

TL;DR

Problem: Generative design in engineering requires diverse, high-quality data and strict constraint adherence. Approach: Synthesize diverse, high-performing data via Quality-Diversity (MAP-Elites), fine-tune a language model on that data to generate high-level designs, and refine them into constraint-satisfying layouts with Wave Function Collapse (WFC). Findings: Datasets built with QD substantially improve fidelity to natural-language prompts and overall validity compared with randomly sampled data, enabling reliable, text-guided architectural layouts via TileGPT. Significance: This data-efficient, interactive framework merges evolutionary data generation with symbolic constraint solving to make generative design more accessible and controllable in architecture and related engineering domains.

Abstract

Two fundamental challenges face generative models in engineering applications: the acquisition of high-performing, diverse datasets, and the adherence to precise constraints in generated designs. We propose a novel approach combining optimization, constraint satisfaction, and language models to tackle these challenges in architectural design. Our method uses Quality-Diversity (QD) to generate a diverse, high-performing dataset. We then fine-tune a language model with this dataset to generate high-level designs. These designs are then refined into detailed, constraint-compliant layouts using the Wave Function Collapse algorithm. Our system demonstrates reliable adherence to textual guidance, enabling the generation of layouts with targeted architectural and performance features. Crucially, our results indicate that data synthesized through the evolutionary search of QD not only improves overall model performance but is essential for the model's ability to closely adhere to textual guidance. This improvement underscores the pivotal role evolutionary computation can play in creating the datasets key to training generative models for design. Web article at https://tilegpt.github.io
Paper Structure (22 sections, 2 equations, 9 figures, 1 table)

This paper contains 22 sections, 2 equations, 9 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Wave Function Collapse
  4. Quality-Diversity
  5. Language Models
  6. Method
  7. Dataset Generation
  8. Language Model Training
  9. Layout Generation
  10. Step 1: Design Initiation via Prompt
  11. Step 2: LLM-Driven Site Design Formation
  12. Step 3: Translation to Permissible Tile Sets
  13. Step 4: Detailed Design Completion through WFC
  14. Step 5: Finalizing a Valid Design
  15. Experiments
  16. ...and 7 more sections

Figures (9)

  • Figure 1: Algorithm flow of the proposed generative design approach, TileGPT. (1) A dataset of paired designs and attributes is generated with the MAP-Elites algorithm, which is used to (2) fine-tune a GPT model to produce designs with given attributes. (3) Given a natural language description a simplified design with the described attributes is generated by the GPT model, and (4) given to a constraint satisfaction algorithm, which refines it into a detailed site plan.
  • Figure 2: Mutation of a WFC genome. Fixed tiles are encoded into the genome, and set at the start of a WFC rollout, influencing the development of the final design.
  • Figure 3: Possible WFC cell states and their simplifications for tokenization. Designs are evaluated using the WFC cell states, but generated using the reduced set of LLM cell states.
  • Figure 4: Layout Generation in TileGPT. (1) A site description is provided to the model, which (2) produces a high level layout. (3) This layout is converted into preconstraints for the WFC algorithm, which (4) generates detailed geometry. The 2D geometry can be then be extruded (5) into a form suitable for use with commercial design software.
  • Figure 5: TileGPT architecture. Text prompts are encoded through a frozen text encoder and are combined with previous tiles in GPT2's cross attention mechanism.
  • ...and 4 more figures