AI Can't Stop Generating 'Kissing Tangents'
Why does it look like that? - Among vague details and extra digits, AI generated images invariably have this one easily detectable mistake. Once you know, you start seeing it everywhere.
In a Reddit post, a parent seeks advice on a dollhouse for their child's birthday. When the progress pics arrived, they suspected they were being scammed. Most everyone agreed.
Many in the thread identified typical AI markers. Vague details and blurriness. But, I had noticed some geometric oddities no one was pointing out. The lines of the roof perfectly aligned with the saw. The other saw nestled right between the gaps in the frame.
Human artists try to avoid this. The visual effect can be distracting, and even look like a mistake to the eye. The Trouble With Tangents is an interesting read about how an actual artist compensates for this effect. For great comics, read this one.
Here's some more examples I collected:
Once you notice this, you'll start to see it all over. Even in AI video. It's possible for a real image to have kissing tangents but seeing multiple instances is symptomatic of the disease.
Under the hood, an image model is compositing a stack of relatively simple filters. These filters produce hard edges that create shapes. This allows AI to create things but also influence the arrangement of the resulting image.
I speculate that an image model is economical with its edges. If need be, a model will not balk at reusing a edge for two or three distinct purposes. It seems kind of lazy in that way.
The simplicity of the pattern makes it exciting for those who are interested in forensics at scale. A traditional computer program can detect this pattern millions of times faster than an AI can generate an image. Someone please build this!
A few years ago, it was easier to pin AI generated art. Today, two reasonable people can disagree about the authenticity of an image. Extra fingers and impossible details are less common.
But AI's flexibility is also its downfall. The generic approach to generation renders the same mistakes everywhere, across scales, styles, and media. It is unable to correct it because it is a fundamental part of how it works.
A line becomes two. A saw becomes a dollhouse.