Tim made a few good points about my AI art illustration from yesterday's post on frilled lizards with annoying pet peeves. Replying to him is actually worth a blog post, methinks.
Tim made a larger point about the content of the post, illustrated with a link to a cartoon that is worthy of another blog post, perhaps tomorrow.
Anyway, dealing with the issue of AI art, it certainly has come a long way in the year or so since I started using it. As long as I ask for no more than two characters, it does a fine job, giving me something I like on the first try almost every time these days. My "old Southern man and his cat" series is working really well.
However, AI art still can't handle any more than two characters. I tried a dozen or so variations on the image above with a prompt something like this: "an old southern man and his cat looking annoyed at a frilled lizard which is furiously exhibiting deimatic behavior." That was the best I got before I ran out of patience and went with it.
The cat is supposed to be a gray tabby. The lizard is supposed to be frilled. The old man's hands are indeed deformed, but at least he was brandishing a broom, giving the scene the appropriate comic feel. Each time I corrected the AI, it would introduce new errors. Usually, the cat's expression was completely out of place. Sometimes the lizard was unthreatening. It was really frustrating.
I keep experiencing this same limitation - 2 characters, no more. For Thanksgiving, I tried the old, Southern man with his cat and a happy family on the porch, but almost every time, the people in the background looked like mutants. I tried a scene from the Arthurian legends, but ended up with a knight fighting a lady while the characters in the background reacted with indifference or faced in the wrong direction.
If you want surreal scenes or goofy illustrations and you only care about the mood and general characteristics of the subject, AI art does fine. Ask for a panda floating through the air, hanging from a bouquet of balloons and you get what you wanted. Ask for a surreal scene of fiery destruction and you get a decent illustration of the SEC this year.
If you want to precisely illustrate something from a story or a multi-person event, forget it. Adobe Premiere's beta of their next version has the ability to use AI to extend a video a couple of seconds past its end point, but the results there have been disappointing as well.
I don't see how this will change without the AI translating the still image or the video into 3D wireframes and applying biomechanics to them. At that point, it's not large language modeling any more, it's a modern video game.