---
post_type: "project"
link: "https://github.com/amirhouieh/grid-reveal"
title: "noui"
blurb: "Early machine learning experiments asking whether interfaces could be generated and personalized instead of hand-designed."
tags:
    - Experiment
    - Machine Learning
year:
    - 2020

stack:
    - StyleGan2 
    - Pix2Pix
    - Python
order: 1322
---
noui was a small set of experiments around a question I kept coming back to: could an interface adapt to the person and the context, instead of being drawn once by hand for everyone? Two parts of the design problem interested me — interfaces that personalize themselves, and automating the design process.

The first experiment trained a StyleGAN2 to draw website layouts, using a dataset of the 1000 most-visited sites. To build that dataset I wrote a small tool, [gridr](http://gridr.amir.cloud/), [open source on GitHub](https://github.com/amirhouieh/grid-reveal). The second used a Pix2Pix GAN to turn a wireframe into a UI, trained on 1000 random webpages — see the [Colab](https://colab.research.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/generative/pix2pix.ipynb). The results were rough and I knew it; a cleaner dataset in one domain would have helped. It was an early try at software shaping itself to its user, which is the thread I've followed since.