Sunday, February 9, 2020

Abstract Paintings—by Humans or A.I.?


Here are two groups of abstract paintings divided by a black line. Half of them are made by humans, and the other half are made by a machine-learning algorithm. Can you tell which is which?
------
Related Previous Posts:
Abstraction Generator
Is it Cheating to Base My Art on Computer-Generated Images?
The Other Abstract Movement

8 comments:

Sam Bleckley said...

Sure.

The cheap answer is that all the AI paintings are perfectly square (and one of the human paintings is hanging on a real wall).That'd be easy to fix, though.

There are a couple subtler signals that I think would work fine to tell the difference even if they made some non-square pictures:

- ML images are notoriously bad at broad compositional problems (like ensuring that a picture of a clock has exactly 12 numbers, not 7 or 18). Almost all the human images (minus the acrylic pour in the bottom row) have either a single focal point or a clear overall direction of movement. It's part of what makes them cohere as a whole (some of them even have a hierarchy: one focal point spread across 3 implied "panels"). The computer paintings have very little directionality, and any number of "main objects". Miro or Kandinsky might be good training targets, because they have lots of little objects that are simple and clear, so the computer might more easily hide the lack of compositional logic.

- ML images have no ties to the physical constraints of paint. In the ML images, things fade smoothly into murky grey; edges appear and disappear apparently at random. In the human paintings, the paint is really dominant; there are drips, dribbles, hard edges, and evidence of the brush even in the softest edges. It's pretty costly to make a ML system create high-res images, but it's cheap and easy to make use of the almost-infinite detail of the real world.

The real question is: if you didn't prompt us, and just stuck an ML painting in with a group of human paintings, would we pick it out as odd? I *think* the answer is still yes, but just like with generated human faces, I get less certain every year.

JR said...

The topic is provocative, the post modestly entertaining, and the terms "A.I." artificial intelligence strike me as quite inadequate

(https://twitter.com/search?l=&q=%22artificial%20intelligence%22%20from%3Arogerschank&src=typd)

(https://twitter.com/search?l=&q=AI%20from%3Arogerschank&src=typd)

What about 'machine learning'? I don't know, really.

Roberto Quintana said...

Great answer, Sam.
My initial response was that a Human could have done all of them, and since the programs are all created (or initiated) by Intelligent Meat, they are all the products of a Human Brain. So which are actually painted by a machine?
I saw what Sam saw, The lower half all have an overall design quality that reminds me of an horizon or focus, while the Upper half seem to be more random or haphazerd, (not that those qualities deny a human hand).
The other thing I noticed that Sam points out, is the paint handling and manipulation. The upper half tend to be darker or muddier overall, the fades and blends tend to fall-off like a photoshop tool, (Not that I couldn’t achieve that effect if I wanted too!)
So, with that all said, and since I have a 50/50 chance of being right, I’ll throw-in w Sam and say the Upper half are made by Machine and the lower half are made by Meat. -RQ

James Gurney said...

Yes, you guys have it right: the bottom set was created by the brain-hand-meat-machines, and the upper ones by a GAN.

Sam and Roberto, you're good visual sleuths. It is getting harder to tell computer generated from photographic or human-created images, and I suppose it will get harder still in the future. I've been reading Russell Stuart's book and Max Tegmark's book on the subject-- very interesting stuff.

JR, you're right that the term "machine learning" would describe the process more accurately. This is indeed a supervised narrow task. But the term "A.I." is broad enough to cover such narrow tasks, and is not reserved for General Intelligence.

Drake Gomez said...

Just as the Photorealist painters turned the "threat" of photography around on itself and began to imitate the look of photos--including such idiosyncrasies as lens flare, depth of field, lens curvature, and so on--will there come a time when painters will imitate by hand the look of AI or ML created images (such as the ultra-smooth gradations or haphazard compositions that others have noted)? Or are there painters already doing this? I can hardly wait to hear what name the critics come up with for this style!

Capt Elaine Magliacane said...

Well I LIKE the ones on the bottom, the ones up top seem muddy and dull... So do I like AI paintings or Human paintings? That is the question.

Roberto Quintana said...

So much for the Turing Test…
This all raises an even more interesting question: “Is It Art?”!
Some folks would say none of the paintings on this post are ‘Art’.
But the human creators/painters would disagree, ‘Their’ paintings are Art!
(The GAN would have to defer to it’s Meat-Over-Lords.)
Which, as Drake points out, brings me to the interesting part: What is ‘AI/Machine Art’?
Not what can a machine produce that looks like human art, but what would an intelligent, autonomous, self-programmed computer create, totally on its own, without human direction or input (other than turning it on and putting a paint-brush in its output port).
Would it be an interpretation of its own program? Just squares and lines, ala Mondrian or Kandinsky? Or squares and lines like Chuck Close, only a portrait of one of its ancestors, a washing machine? What would an intelligent machine do, on its own, that was new and creative, and dare I say, expressive? -RQ

Carol said...

Either "found" or created, it takes a human to appreciate and judge the results. And obviously not all abstracts are created equal.