I trained an AI to imitate my own art style. This is what happened.
Alright, let’s talk about artificial intelligence. AI is everywhere today, and if you think we just DO NOT NEED another AI article, you’re probably right. But before you close this tab, please hear me out. This one is different.
First of all, I’m not a developer or software engineer. I did not create another AI. I’m a designer / digital artist with a little bit of coding skills. However, I’m very passionate about all things technology. When I was researching AI for a university project, I came across some interesting image-generating scripts. There was especially one that caught my attention: a method called Neural Style Transfer (NST).
What NST does is basically this: You define a content image and one or more style images. The Script then tries to transfer the art style it learned from the style image(s) to your content image, creating something pretty new and probably really wild.
When I learned about this, I was pretty excited. Not by most of the generated pictures I saw on the web, I thought styling your latest selfie with van Gogh’s Starry Night is kind of boring. But rather by the possibilities this stuff creates. I have worked with generative art before but NST seemed to be on another level entirely. And there was this one question I just couldn’t stop thinking of:
Can I train an AI to copy my very own illustration style?
As I said, I’m a digital artist. For over 15 years I’ve been doing digital abstract art. Mostly personal, sometimes commercial. Here are some examples of my work (for more visit my portfolio at https://eightvisions.net):
So I wanted to answer that question. I’ve been looking for a way to do NST myself and found Somshubra Majumdar’s Neural Style Transfer repo on github. I figured out that this stuff needs a good NVIDIA GPU with lots of computation power, which I do not have. Luckily there are some companies who offer these GPUs for rent. I decided to go with Paperspace*, created an Ubuntu VM and configured it as needed.
*Affiliate Link! If you sign up on Paperspace with this link, you will get a 10$ credit and I will earn a small commission. Thank you for using it!
And then, the fun started.
After setting up the script and everything, I did some first tests to see how this works in general. This certain NST script is using a single style image and runs in a pre-defined number of iterations. Every iteration will make the final image clearer, more saturated and will smoothen out some rough areas in the details. There are also many advanced options you could use like masking or multiple style images.
After doing some test I found out that it pretty much makes no sense to run the script for more than 25–30 iterations because after that number you will only see minor enhancements. That is especially the case when you train it with something very abstract like my illustrations. I just do not need that much clarity in the finale image.
Lets do an alphabet!
I was looking for an interesting use case for this, and during my experiments I found out that applying this stuff to typography results in some very interesting visual effects I’ve never seen before nor would I be able to do by hand at all:
That’s why I decided to generate an alphabet. The basic rules would be the following:
- 26 characters (A-Z)
- the content image is just b/w with a letter in a sans-serif hairline-typeface on it, to not influence the “work” of the AI too much
- each letter gets assigned to an unique style image that is one of my hand made illustrations from the last years (to be more precise, it is a clip of my artworks showing many details).
- 20 iterations from the AI
- be curious what will happen
The Results
Now, here’s what the AI made for me:
And here are some gifs, showing the process of the said 20 iterations:
Pretty interesting, right?
If you would like to see all the letters, check out the full project on my portfolio: https://eightvisions.net/work/the-alphabet/
Conclusions
Last but not least, there are some thoughts I’d like to share.
So, can an AI actually copy my own art style?
No, not yet. What I’ve learned over the years doing this is not a technique or a software skill. Its a feeling for composition, use of colors and creating harmony within a total mess. I do not think that an AI can do that TODAY. I have no idea where technology will lead us, so maybe one day it might be possible.
Do I personally like the AI-generated results?
Yes! It’s a new take on generative art, it kind of visually melts the elements together in a way I’ve never seen before. Its special, different, sometimes ugly, sometimes beautiful. You never know what you’ll get out of it. But wait a second, this raises another question:
Are we as humans even allowed to judge art done by machines?
I don’t know. That’s a tough one and you could probably write a book about that topic. For this project, I just claim my right to judge it as I trained it on my own art style.
Finally I would like to send out a big THANK YOU to Somshubra Majumdar and team for their work on the NST script and for sharing it with the world. And of course to the people who supported me while doing this project.
Update April 2020:
One year later, its great to see that this project still gets some attention. Huge thanks to everybody who shared this or mailed me about it. Really happy to hear all your thoughts.
I have also done another project with NST. If you’re interested, you can find it here: https://eightvisions.net/work/the-essence-of-art/. That one is basically another experiment, randomly cloning 10 hand made illustrations on top of each other.
Thanks for reading!
I’m a product designer and digital artist from germany. Check out my portfolio at https://eightvisions.net or say hello via E-Mail (michael@eightvisions.net) or Twitter https://twitter.com/eightvisions
