instruction following is good most of the time, but hands are still regularly coming out garbled and there's a small percentage of failed generation in the mix. style following is also a bit hit and miss, I'm producing watecolor mostly, but every now and then a generation is photographic
I manually remove failed gens, but the proces doens't scale to where I'd like if I have to manually verify images. and I have to have a few automatic rejection in the pipeline, like top border not white or image not squared.
OpenAI costs are surprisingly expensive. It's about $7 to generate a whole book (24 pages). There are 8-24 images allowed in a book, with a cover too. So there'll be 48 max pages in a book (incl blank pages).