The quality of hand generation by neural networks is gradually improving, but the problem of anatomical accuracy still remains. Models often fail to correctly “count” fingers. Instead, they recognize a pattern similar to “a certain number of similar elements placed next to each other.”*)
When a neural network tries to fill an area that, based on its learned patterns, should be a hand, it may generate four, six, or even more fingers. This behavior is explained by the statistical nature of training: the model has seen many images of hands, but rarely with fully detailed anatomy.
A neural network does not understand how a hand should look; it merely guesses the structure based on incomplete or obscured fragments and thousands of variations from the training dataset. Therefore, during generation, it is forced to “improvise” anatomical details, which leads to errors.
*) How do neural networks “guess” structure?
Modern generative models (DALL·E, Stable Diffusion, MidJourney) operate on transformers:
- The image is broken down into tokens (small fragments).
- The model predicts the next token based on statistical likelihood.
- If a part of the image resembles a hand, it chooses a combination of tokens that most frequently appeared during training.