I think this is as good as it gets for a concept like this with Z-Image.

training notes: 32/32, 150 images, qwenvl3 very detailed captions, 0.0002

Start with strength around 0.8 and adjust higher if you don't see correct details and lower if you're seeing "problems'