This is essentially a distillation on the bigger model; you'd wind up surfacing a lot of artifacts from the host model, amplifying them in the same way repeated photocopying introduces errors.

https://dailyai.com/2025/05/create-a-replica-of-this-image-d...

There are not enough samples in that book to generate new "infinite" data.