The problem
Illustrating an entire book with AI is easy; keeping the same character consistent across 200 pages is not. And manually checking 550 images doesn’t scale.
What was built
A pipeline that generates, redoes, and links images to their positions in the book, with verification against the reference document (not guesswork) and prompts written in pure positive phrasing — negation doesn’t work in image prompts. APIs were benchmarked before selection: only one delivers an alpha channel, another always returns JPEG, and pure PNG costs 2x more and takes 5x longer.
The client approves through a dedicated portal with a PIN — nothing hosted on third-party services.
What was learned
A consistent image library is a matter of method, not luck: a canonical reference, batch verification, and a “redo” path that doesn’t break what’s already been approved.