AI in Art faces the same challenges as Human Painters
Hands are still the very rock we both split
It’s amazing, how closely AI algorithms follow the perils and pitfalls every painter of the past and present faced in their artistic practice.
For example, many of you have noticed, that the hands (the wrists and fingers to be precise) appear to be the toughest part of the human body in terms of accurate representation on canvas.
I often pay particular attention to this fact in the stories about portraits in the Smart Art — Art History Escape app.
Indeed, even the best painters have struggled with their proper depiction and often used to come up with some tricks to execute them properly (saying hi to Anthony van Dyck and his plaster generic hand models he is said to apply to every portrait) and often charge a substantial premium for portraits with hands not “in pockets” (like fabulous Pompeo Batoni did).
Some of you might remember my first advice from “How not to waste money with your first Old Master? 4 things that make a painting worth investing”:
“The first things you should train yourself to look at are the wrist, the fingers, the hand itself. How was it drawn? Does it look exactly like a real human’s hand or merely resemble some flabby glove at maximum? You would probably be surprised, but «hands» are the most challenging part of a human body in terms of correct naturalistic depiction. And their proper execution is a clear sign of a great master.”
And now imagine, looks like AI generative models which are now trained to provide you with every possible drawing, painting, or illustration you would ever dream about, fall into the same trap good old painters did.
They just fail to deliver hands properly!
It’s amusing how I bumped into this “discovery”.
A couple of days ago I was working on a video for my youtube channel with video essays about enigmatic historical figures, mysteries, facts and legends of the past [natural curiosity drove me to create this side project somewhat related to my main art history and collecting niche].
I was doing a video about Kate Warne, America’s first female detective, and found myself in a situation when there was not enough historic illustrative material to brighten up the 20-minute footage of my talking head.
So, I referred myself to one of the most advanced models readily available online.
And each and every time I submitted a description of the desired outcome (like, “a lady detective in her office, studying papers, 19th-century drawing, retro style”), I got fascinated.
First of all, because the AI really worked much better than one would suppose. I got perfectly acceptable stylised images of a lonely person roaming in the forest, a man reading a note in his office, a meeting of a secretary with the lady detective and so on, yet one detail, if present in the composition, was ALWAYS a failure.
Yeah, you got the idea, these were the hands.
Well, it is known that these models are trained by feeding the readily available images online. Drawings, photos, paintings, anything that can be used to recreate the desired style, manner, and subject matter.
Amusingly, almost always the faces, their emotional expressions, and the postures were delivered on a rather acceptable level (taking into account it takes the model about 10 seconds to come up with a single illustration).
I’m not even talking about the non-human objects, like trains, interiors, and exterior settings — these too were rendered with a good level of accuracy.
Yet, the hands of the sitters — this was just a nightmare. Rather funny one sometimes, though.
Four to six fingers, enormous, or, on the opposite, minuscule sizes, weird bending of fingers, and just complete nonsense instead of elegant palms.
Try to play with it yourself!
So, it’s curious, right, how the artificial intelligence mind of a neural network repeats the path of a human neuron and struggles with adequate hand representation even when the other parts of the picture seem to look more or less valid.
Is there a real connection to art here? Or is it just the style that I input as a reference — a 19th-century drawing?
Or is it just my bias and it’s just another example of the “correlation does not imply causation” principle?
Would be interesting to see how the technology develops in the near future.
