>some neural networks are currently struggling with drawing 3 people in the same pic
Illustrious and noobAI struggle with just about anything that isn't tagged hundreds of times on danbooru, unless the user has the time and patience to build datasets and train loras himself to make up for that. Even when the AI knows what something looks like you can't get much variation. It can draw an alarm clock and do simple stuff like change the color and size of the alarm clock, but can it make a character-themed alarm clock that actually looks good without training a lora specifically for that? It also has a very poor "understanding" of physics, and so many other problems that make it boring to use after a while, and I find looking at the endless supply ero art that japanese artists continue to make more interesting. It's great if you need more naked pics of specific anime girls that you've already seen all the existing ero art of though, or need to see them performing specific lewd acts there aren't many pics of, like shitting. I hope some mad scientists manage to make far smarter and more creative AI someday, one that knows exactly how physics work and can deliver exactly what you want and better no matter how complicated or weird the prompt is.