Image-generating AIs, like Midjourney, are getting more and more powerful, to the point where they can sometimes fool humans. However, there are still many subjects that they cannot reproduce.
At first glance, it might look like completely normal unicycle photos. However, when you look closer, several anomalies are obvious: there are not always pedals or saddles, and some devices are completely crazy. It’s normal: the images were created on Midjourney by the English journalist Luke Bailey.
On Twitter, he explains that he realized that the artificial intelligence dedicated to generating images was having difficulty recreating certain objects, such as unicycles. And this is important news for detecting the fake images that are increasingly on the web.
Midjourney can’t do everything
We wanted to replicate the experience by having Midjourney create realistic unicycles. The result is, for us too, inconclusive. The feet of some cyclists are not in the right place, and unicycles even have two wheels, which has nothing to do with the vehicle requested. It’s a zero point.
It’s not the only device that Midjourney misses. Tests carried out by Numerama have shown that the artificial intelligence also struggles to model bicycles, and still does not know how to make the letters of the alphabet (a problem that the previous version of the AI already had). She also doesn’t know how to make braces, or “grillz” (the gold teeth of rappers, editor’s note), or dinosaurs, and she is also in trouble with certain animals and their skeletons.
Finally, Midjourney fails to properly portray people eating, be it burgers, couscous, or even drinking coffee. We let you judge how Emmanuel Macron ingests spaghetti.
At the beginning of March 2023, version 5 of Midjourney amazed us by finally managing to make realistic hands. While previous versions of the AI could never do normal fingers, this one does them much better. Today, Midjourney’s creations are impressively realistic, to the point where it’s now easy to be fooled by a fake photo — like the one with the Pope in a puffer jacket.
That being said, Midjourney still has major shortcomings. The fact that she doesn’t know how to ride unicycles well, however, won’t solve the problem of AI-generated fake news — there’s no often of unicycles in the photos. And above all, maybe version 6 of Midjourney will manage to make spaghetti correctly.
A track can explain the weaknesses of Midjourney: the lack of data. The AI is trained on a large corpus of images. The AI is therefore particularly good at recognizing and recreating what appears very frequently in photos: landscapes, humans, pets, etc. The hands are more rarely the target of an image; it is usually the face of the people photographed. This explains the time Midjourney took to get up to speed.
For the rest, it’s the same thing: shots of unicycles are less common than those showing cars. Ditto for dinosaurs versus cats, and so on. The issue of “input” data, to feed AI systems, remains key.
If you liked this article, you will like the following ones: do not miss them by subscribing to Numerama on Google News.