Artificial intelligence (AI)-generated imagery has moved on impressively in the last decade, from 2014 when the most cutting-edge technology could produce a picture that looked like it was taken on a Game Boy Color. Now it can even win state fair art competitions.
However, there are still a few things that all the main leaders in AI-generated images struggle with. As well as teeth (and remembering not to cram hundreds of them into a single horrific mouth), AI is pretty terrible when it comes to hands.
A recent example of this was posted to Twitter by creative director for mobile games Elizabeth Sampat. “My favorite part is Josh Groban looking confused,” she added.
As plainly demonstrated, the image generator used has not really nailed the concept of hands. But this is part of the reason why they are so bad – working on pattern recognition and producing patterns rather than having the concept of hands themselves.
The image generator may think it has nailed the pattern by putting an indeterminate amount of fingers clumped around a palm, or in this case sometimes through the palm, but it will not survive the scrutiny of a quick glance by humans, with our fancy concept of what hands are.
On top of that, it’s incredibly difficult for humans to draw hands, too.
“Due to the complex geometry of hands, there is no standard set of lines or shapes that AI can recognize as a hand,” one designer and AI expert, Jim Nightingale, explained in a blog post. “In order to generate realistic hands, AI needs to collate multiple different shapes and arrangements. There are about 30 points of geometric divergences in the human hand, from finger length and width to the metacarpals and wrist joints.”
Coupled with additional human scrutiny of human hands and facial features, AI has a hell of a task on its hands to impress. For now, we’d settle for hands with five fingers that don’t emerge from other fingers or the center of the palm.