That original idea also sounds interesting to me, if you modified it to work in a semantic space where "beautiful" and "not beautiful" differed by more. Write a hypothetical review and it finds the closest match. It's a very cool project, thank you for sharing this with us.