I lost count of how many otherwise-good images I had to discard because the eyes were useless. I tried the tricks the basic how-to guides recommend (symmetry, stopping at 80%, only using light upscaling) but it was pretty well hit-and-miss.
It's also very hard to get it to create two different people with different descriptions standing next to each other. You can do it with names, but even then it often believes you want one person who looks like both of them. Also, when you ask for variations on an image, the variations are invariaable worse than the image and you have to get variations of them.
I'm frustrated that there isn't a "but with" option, or at least if there is I haven't found it. "I want image 2 but with blue eyes" or "I want image 3 but with a completely black robe" would be good. You can use a web-based image as a starting point, which could conceivably work but I haven't tried it.
It does do some amazing things, though, I won't say the most impressive one it produced because it's a spoiler for the book, but it was along the lines of "[person] who doesn't know [person]".
Unfortunately, I made the mistake of showing the images to my wife. She took an instant dislike to the featured character on the basis of this and now refuses to read the draft.
I still have 30 minutes of rendering time left this month, so maybe I'll try a different approach — one that doesn't show people's eyes.
