Simulating
Creative Intelligence
We are building a substrate to train and evaluate AI models to be
capable design partners across expressions of taste and consumption.
001 - The Problem
A story isn't consumed in one place.
It's met by different people, in different moments, on different terms.
SECTIONTHE PROBLEM
Every surface
a story lives on
a story lives onhas its own inhabitants,
its own pace,and its own
and its ownrules of engagement.
002 - THE SOLUTIONWhat AI needs to become
What AI needs to become
a real design partner.
SECTIONSOLUTION
Understand the task
What is being asked, where it lives, and who it’s for.
Produce the design
With the right structure, tone, and constraints
Iterate on it
Refine without resetting, adapting across contexts
003 - THE BENCHMARK
See the full resultsSee the full resultsThe Graphic Design Benchmark
Five categories. Forty-nine tasks. Four frontier models put through every step a designer takes — from reading a layout to producing one.
1 / 49
Understanding
Generation
Aspect ratio classification
989 layouts
Predict canvas aspect ratio from nine categories from visual content alone.
Key finding
GPT-5.4 near-solves this (93.9%); all other models below 25%.
| Model | Accuracy↑ | Macro-F1↑ |
|---|---|---|
| Gemini-3.1-Flash-Lite | 0.236 | 0.105 |
| Gemini-3.1-Pro | 0.245 | 0.085 |
| GPT-5.4 | 0.939 | 0.679 |
| Claude Opus 4.6 | 0.093 | 0.179 |
GraphicDesignBench · Use ← → to navigate
Solved Partial Unsolved
004 - THE RESEARCH
View AllView AllWhat it takes to get there
A snapshot of our work to train AI to be a design partner:
SECTIONTHE RESEARCH
