Preliminary attempt at horizontal pupils.

Usage: horizontal pupils

Highly recommended with ADetailer or similar.

I've tried to improve transferability by training at a moderately low rank. By using a low SNR Gamma (=2) and a high res of 1024 w/ batch size 2 I've hoped that the network would only focus on the pupil detail. Yet the success rate is perhaps only every second image. Using a higher strength than 1.0 can improve the result a little, but will affect style.

I imagine that the most optimal way would be to train a concept slider LoRA (e.g. in the likes of Ostris or Baulab), but alas I don't know how to do that.