Ed: “There are still frequent problems with fingers”
The story I heard was that AI gets confused because there are too many examples of wrongly-drawn hands on the web. AI can only “learn by example”. If it sees a significant proportion of incorrectly drawn hands on the web, and has no way to be taught right from wrong. And then, we can expect AI to generate a similar percentage of incorrectly drawn hands. Effectively, “garbage in, garbage out”.
Many of us have been looking at web pages for 25 years. Most of us have probably never remember seeing a web page and exclaiming “that hand is wrong”. Yet as soon as we judge created by AI, we notice. Could it be that we subconsciously assume that AI should be superior, and apply a higher standard to the work it produces than to the work produced by mere humans?
Maybe even, if I popped into the National Gallery and had a close look at paintings I’ve briefly glanced at before and study them in more detail, I would find similar examples of badly drawn hands amongst them. But I don’t have time to do that right at the moment.
Instead I did a web search, and found this:
Why can’t AI draw realistic human hands?
Eray Eliaçık
January 19, 2023
One observation it makes is that “most cartoon hands only feature three digits and a thumb. Since we’re used to it, our brains don’t even register it”.
I never knew that, and was prepared to believe it to be true — but wanted convincing. The cartoons I saw at an early age would mostly have been Beano and Dandy. But an image search for Beano and Dandy Cartoons completely disproved that idea. Every hand I checked had the correct number of fingers :). Perhaps Beano and Dandy were two examples of cartoons that were “different to most” but it seems too much of a coincidence. Perhaps the statement is correct for the USA market and UK cartoonists are more strict on accuracy.
What about the backgrounds?
Two possible reasons. Sites have to share resources between everybody that is using them, so putting less computation power into generating good backgrounds means possibly more attention can be given to getting the main person right. Certainly it seems that asking for just one person is more likely to give good results than asking for several. And one user has recommended specifying “blank background” to improve rendering of the foreground.</div>
But I also wonder whether that’s a technique that AI might have learned from famous artists in museum pictures found on the web. There you will often find huge pictures with a few people in the foreground painted in very high detail, and lots more people in the background each painted with just a few brushstrokes to give an impression of a lot of people with very little detail each
Ed likes this