Why are hands difficult for ai images?
In: 250
An AI is not a mind. In essence it is a pattern-matching algorithm that is trained on images and gains a model of what things tend to accompany each other in a lot of images. For example the AI doesn’t know what a tree is but it can learn that images of leaves tend to be clustered in certain ways associated with a trunk and with a particular orientation to the ground, etc.
Since it is all a big combination of matched patterns without any real understanding of what the patterns are, hands are a significant problem. Fingers tend to appear in images near other fingers, but their orientation with regard to each other varies significantly. Furthermore the number of fingers which are visible in a given image can vary as well, since many positions of the hand block the view of some of the fingers. The number of fingers is also very precise; only 5 per hand, no more and no less.
With that in mind it is very easy to spot something odd when AI is creating an image of hands. Nobody is going to notice if a tree has an unusually high or low number of leaves as there isn’t particular number to hit. Also the orientation of leaves can vary significantly and the tree will still look acceptable, but not so for fingers.
Weirdly enough hands are difficult for your unconscious mind too. One of the easiest ways to determine if you are dreaming is to look at your hands.
Because hands are difficult for humans to draw.
Most artists find drawing hands difficult because of all the small joints that have to be positioned just right to look real.
As a result, human made drawings often have badly drawn hands or the hands are hidden.
AI is trained on images made by humans so if humans struggle with something, it will too.
Teeth, when smiling, is tough for AI too, the results are quite disturbing! Toooo many teeeth
Teeth, when smiling, is tough for AI too, the results are quite disturbing! Toooo many teeeth
We have some ideas about why (mentioned by others), but we do not really know and it is an active area of research.
– source: I am PhD Student studying generative AI models since 2017.
We have some ideas about why (mentioned by others), but we do not really know and it is an active area of research.
– source: I am PhD Student studying generative AI models since 2017.
We have some ideas about why (mentioned by others), but we do not really know and it is an active area of research.
– source: I am PhD Student studying generative AI models since 2017.
At the end of the day, computers are still just fancy calculators. AI is just mathematics, specifically statistics. You feed an AI a tonne of images, some are a single hand with 5 fingers, some are two hands together with 10 fingers, maybe some fingers are bent or hidden behind something, maybe there are multiple people’s hands in the image. The images are flat, technically the AI doesn’t even understand 3D geometry and camera angles, or how joints and rotation works, so you’ve confused the shit out of it. Computer says, well on average an image of a person has 7 of these pointy fleshy things going in this direction so I’ll draw that, even though physiologically it makes no sense.
And when it screws up, it’s really obvious, because we humans do understand 3D geometry, and the basic biology of having exactly 5 fingers per hand, and the different ways they can bend. Hands are one of those things that you can’t really think about statistically, which is the only way the AI knows how. Unlike say, eyes, which are more or less always the same shape and in the same position on a face, or a nose which can be big or small or anywhere in between without going into the uncanny valley.
At the end of the day, computers are still just fancy calculators. AI is just mathematics, specifically statistics. You feed an AI a tonne of images, some are a single hand with 5 fingers, some are two hands together with 10 fingers, maybe some fingers are bent or hidden behind something, maybe there are multiple people’s hands in the image. The images are flat, technically the AI doesn’t even understand 3D geometry and camera angles, or how joints and rotation works, so you’ve confused the shit out of it. Computer says, well on average an image of a person has 7 of these pointy fleshy things going in this direction so I’ll draw that, even though physiologically it makes no sense.
And when it screws up, it’s really obvious, because we humans do understand 3D geometry, and the basic biology of having exactly 5 fingers per hand, and the different ways they can bend. Hands are one of those things that you can’t really think about statistically, which is the only way the AI knows how. Unlike say, eyes, which are more or less always the same shape and in the same position on a face, or a nose which can be big or small or anywhere in between without going into the uncanny valley.
Latest Answers