Works       Bio      Sketchbook

Generative Tiles

GANning traditional geometric tile patterns
How would a machine learning algorithm design tiles?
Trained on a dataset of traditional Indian and Islamic tile patterns, the StyleGAN model interprets color, geometry and relational patterns to generate mischevious, wild and unconventional patterns.

This project was undertaken as a musing ~ how would a machine (aka A.I. ) design architectural furnishings? It attempts to participate in now long standing debate and conversation around collaboration of man and machine in creative practices. To what extent can machines take over design processes?

The training set comprised of about 300 images which were collected from the web. Each of the images represented traditional Indian or Islamic tile patterns. These patterns are characterized by combinations of repeated squares, circles or other polygons. Experimenting with composition, some of the most intricate patterns of traditional tile art were a part of this dataset.


Can AI design tile patterns that can be physically manifested in our environments? The dissatisfying thing about this design move was acclimatizing myself to the AI’s ‘taste’. I didn’t automatically love the patterns it generated, it took a lot of observing, looking and re-looking in order to start seeing the value behind them or questioning why they are the way they are. On the other hand, if these patterns were made by a person, I could ask them why and how they made it and they would be able to give some form of an answer. Can AI really replace human designers given the fact that it itself cannot explain the designs it makes?

While the initial results (the first four images on top) did not instinctively look like a tile pattern, the AI got better at it (the last four images above). The patterns were gradually starting to look more geometric and in correspondance with the training data. But I wonder if the initial generations show more of the AI’s mind...they have more character and potential for development and manipulation.
︎ ︎