All the photos that we ask the AI systems have the same problem: the hands come out with an infinite number of fingers, or only with three or four.
Something is wrong with Artificial Intelligence, it seems that it doesn’t know how to do hands, and that is a problem if we want to do decent realistic work.
The examples made by Midjourney are really terrible in that regard. He even manages to make the fingers float, as if they had been torn from the hand. Creepy.
Problems are due to various reasons:
– In the images that have served as training, the hands are less visible than the faces.
– Many times the hands appear behind the bodies, or above the shoulders, so the AI is not very clear about the context of where they should appear or with how many fingers.
– In training photos, the hands are often holding things, so they don’t show perfectly.
How can we fix the problem?
The only way to get the AI to draw good hands would be to train them with perfect shots of hands doing everything, and then many other images showing those same hands in context with the body, but for this you would need millions of specifically created photos. for the subject, and it is not something easy to achieve.
It would be possible, for example, to ask the world to send photos of their hands in different positions, but few would do it if there is no prize in return (a month of free use of an AI of this type, for example).
