I was really surprised to learn during one of our Information Strategy lectures that AI generally struggles with generating pictures of human hands. First, I wanted to try that out and second, I wanted to know why.
I used two AI tools to test this anomaly, Bing Image Creator and ChatGPT-4 with the DALL·E 3 beta. I used the following prompt: “Create a realistic-looking picture of a professor in a university lecture hall. The professor is standing in front of the class and is holding a presenter in the one hand, and a coffee cup in the other hand.” At first sight, you don’t really see the issue but when you look more closely, you see how weird some hands look. In picture 1, the woman is missing a finger of her right hand, and her left hand is looking especially unnatural. The professor’s presenter in picture 2 is floating in the air above his hand and the woman in picture 3 appears to have three hands, two of which are holding the coffee cup and another one is holding the presenter. Overall, all these generated images look very good and natural in my opinion, except for the hands.
But why is generating images of human hands such a problem for AI? Firstly, the models are 2D image creators that do not understand the three-dimensional nature of a human hand. (Hughes, 2023). Secondly, their training data mostly focused on other parts of the human body, such as the face (Hughes, 2023). Therefore, the AI tools have especially big problems with creating images of hands if you provide a context in which the hands must appear, such as holding specific objects in my case.
I’m curious to see how this topic evolves in the future and how long it will take AI tools to get better at generating human hands. I look forward to your comments!
References
Hughes, A. (2023, February 4). Why AI-generated hands are the stuff of nightmares, explained by a scientist. BBC Science Focus. https://www.sciencefocus.com/future-technology/why-ai-generated-hands-are-the-stuff-of-nightmares-explained-by-a-scientist
