The AI has to learn from the few examples it has.The quality of the data is also a factor.Let’s say the AI has a ton of pictures of hands.But the pictures are blurry or the hands are in weird positions.This makes it hard for the AI to learn what a hand should look like.And finally, the margin for error is super low.If you draw a hand wrong it’s still recognizable as a hand.But if the AI messes up it’s just a blob.So, if you want to create a post-apocalyptic giraffe astronaut, AI art can get you there.But if you want to make a woman holding a cell phone, it might take a few more tries to get it right.
Understanding AI Art and Hands
Creating a post-apocalyptic giraffe astronaut is no problem for AI art models, but when it comes to a woman holding a cell phone, things get a bit more difficult. I spoke to two people who have worked with generative art models to understand why AI art struggles with hands.
The data size and quality, the way hands act, and the low margin for error are three big reasons. The AI has to learn from the few examples it has, and if the pictures are blurry or the hands are in weird positions, it makes it hard for the AI to learn what a hand should look like. Additionally, if the AI messes up, it’s just a blob, whereas if a human messes up, it’s still recognizable as a hand.
Pattern recognition has allowed AI and people to draw decent apples, but the processes differ. People learn how things look by living in the world and recognizing patterns, whereas an AI is like a person trapped in a museum from birth. All the machine has to learn from are the pictures and the little placards on the side.
So, if you want to make a woman holding a cell phone, it might take a few more tries to get it right. you need a lot of computing power.Second, they’re trying to find better ways to annotate the data.They want to annotate hands and faces in a way that will help the AI learn more.That means they need to figure out how to annotate hands in a way that’s useful to the AI.
AI art is generally not very good when it comes to creating hands, due to the scarcity and complexity of data available to train the AI. For example, popular datasets like Flickr HQ and celebrity faces have 70,000 and 200,000 images respectively, but they don’t have annotations for details like eyeglasses or pointy noses. Even datasets specifically for hands, like the one with 11,000 images, don’t have the same level of detail that is required for the AI to accurately create hands. Furthermore, the hands in art museum images are not annotated to show how they work, making it even harder for the AI to learn. As a result, the AI is unable to accurately make hands, and this jankiness is seen in other AI art like horses which can have three, five or six legs.
The AI is also limited by its lack of bias, which humans have when looking at hands. We expect them to have five fingers and to be able to do certain things, but the AI doesn’t know this. For example, when asked to create a person with exactly five freckles, the AI was unable to do this accurately. This is because it needs more data and better annotations to understand hands.
The latest AI art generator, Midjourney version 5, has made some progress in this area, but it’s not perfect yet. The AI is still limited by its lack of data and bias, and it is spending a lot of time on things that are not always noticed. To fix this, the AI needs more computing power to look at more images, and better ways to annotate the data so it can learn more. If you want to train more than 100 images, this would take tremendous resources to retrain the model itself. The other solution might be to invite more people to rate the images generated by the model, so that the model can be fine tuned to generate images that are convincing to people. This would require a lot of engineering to get people to label so much data, but it would help to train the models to do what people like, such as recognizing patterns or large amounts of something, rather than just the hand, teeth, and abs.