This is ADOM, the shirt designing A.I. that collaborated with us

Will A.I. or computational unified devices ever be able to produce art better than humans?

The answer to this depends on the eye that passes the judgement…qualifying and defining ‘better’ is a daunting task…Neural networks and machines are relentless at doing their tasks, and if configured to assemble artistic structures and styles, they will do very well at it, and already are capable of outputting hundreds more designs than their human counterpart in the same amount of time…

Right now we are still rooting for team human. 

In January we set off to program an A.I. hoping it would be capable of producing art in a similar style that represents our most popular graphics.

A variety of methods have been developed to tell a computer to analyze an image and then produce something with the same attributes or ‘style’ that the original had. Using GANs ( generative adversarial network ) or a method called Art Style-Transfer - We can now easily tell a computer to analyze a certain image or video or sound and then apply the similar elements to another canvas to produce a similar effect or 'style'. 

The project goal:
Collaborate with an A.I. to produce wearable art.

The method we used was Art-Style-Transfer built with:
- Tensorflow
- cudaNN
- Python
We used 1 GTX1080ti to for all data rendering / computing

What the A.I. ADOM did:
Below are the textures we input for the styles - We told ADOM to analyze these textures for each shirt design, then produce what it thought we would make. It took 10,000 iterations per texture rendered by ADOM.

A very simple explanation of an iteration: Let’s say we want an old analog scanner to learn what a photo of a dog is. Each time the scanner scans the photo of the dog it calculates the data that it observed, categorizes it accordingly and remembers the information for later. Improving itself. That process would be 1 iteration.

The higher the file size the longer or more computing power it requires. We ran 10 texture styles each many times. Each 1 style produced by ADAM took 4 hours. Everything ADOM analyzed and produced was at 2000px.

What Humans did:

We then took all of ADOM’s content, meshing it in PS. We modified the layout of the art and used the clone stamp tool to extend & glitch elements on the final canvas. Lastly we copied parts of larger textures to make 1 piece for each of the sleeves and spines.

Below is the final result of the two shirts:

We encountered a lot of trial and error setting up ADOM, but had a great time working together to produce this set. 

Currently a computer is nothing without a human, but once the spark is ignited, the capabilities are endless. In this case, our NN was used as a tool...and hopefully won't be replacing us anytime soon.   

Thank you.

If you would like to try and produce your own similar Art-Style-Transfer, there's hundreds of resources on GitHub and Reddit. 

Programmed and Produced by: 
Nate Khouli 
Jonathan Colby

The pieces are now up on our Shop. 
Human Being:


Write a comment

Comments are moderated