Tuesday, September 17, 2019

Humans Team Up with Computers to "Breed" Images

Here's a piece of digital art made by ArtBreeder, a website that generates images by machine learning, and then lets you "crossbreed" them to create new offspring.

Here's how Recomendo describes the process: "Using deep learning (AI) algorithms it generates multiple photo-realistic “children” mutations of one image. You — the gardener — select one mutant you like and then breed further generations from its descendants." 

"You can also crossbreed two different images. Very quickly, you can create infinite numbers of highly detailed album covers, logos, game characters, exotic landscapes."

The software currently doesn't "understand" the meaning of writing, but only the appearance of typographic letterforms, so the system churns out images that resemble evocative album cover designs.  

You can also create landscapes that look almost plausible, or combine dissimilar environments and see what results.

Some images appear to morph organic textures with humanoid forms, like this "feather-boa yeti." You could start with an image like this as reference, and then elaborate it with your own old-school sketch process.

Because the judgment of flesh-and-blood humans assists the computer in shaping the evolution of these images, the process yields different results than a generative adversarial network acting alone.

If you want to play with the software, it's free at ArtBreeder.
Thanks, Dan


juliem said...

In the mid 2000's I worked for Affinnova (now owned by Nielsen). They had a unique market research application based on "breeding" ideas. For example the color, shape and text on a soda bottle. A random sample of people would be invited to take the survey. Each person would be shown four images per page and asked to pick which ones they preferred. The algorithm would use these picks to fine tune what was offered in the next set. The user would probably see about 20 pages and then be asked questions about their top picks. Other people coming in later to the survey would see more refined results based on what the earlier participants chose. In the background this required a lot of statistical work and a lot of design work as all the parts had to match up in any way that they could be combined. One of my jobs was to check that potential image combination matches would look correct since a bad image could impact whether an idea was selected or not.
Anyway, it is amazing to see how easy it is to these images breedings today!

Luca said...

I had a look at the site and i've to say i find the results in the galleries a bit disturbing and unpleasant, for some reason.

Bill Marshall said...

Judging from what you have posted as examples, I have to agree with Luca. Can't put my finger on it, but downright CREEPY!


Pierre Fontaine said...

I actually really like the images. I particularly like the first image posted here. It reminds me a lot of a Jon Berkey painting.
Of course, these images wouldn't exist without the seed images but I'm still intrigued by this technology. Do I consider it art? No but I do think it could be a fantastic tool in the right hands.

James Gurney said...

Pierre, that's a good way to look at it: a fantastic tool for generating ideas. Computers are becoming more like intelligent partners with behaviors that we recognize as almost human, or often superhuman.