The goal is to show high profile, round implants in your generations.

For Klein9Base - initial attempt. Seems to work well around 0.5 to 1 and above in strength. Haven't had a chance to see if it works well with other LORAs.

For V3 Zbase: Much better than the previous versions. Use this one. This is technically a LOKR, used V1's dataset. Works well with and without clothes, use captions of covered and clothed to help guide the model. V3 works at strengths up to 2 pretty well, depends how much you want to go into "ask your surgeon for a refund" territory.

For V2: More of a quick attempt to see what a larger dataset without captioning gets me while trying out the base model for LORAs. I will try V1's dataset/captioning at some point for a 1:1. Strength is same as the distilled model's lora. Seems overall to do a lot better but still not perfect.

For V1:Trained on 25 1024x1024 image dataset, captioned by LLM then edited by hand. Largely avoids faces. Used the first gen adapter on. For the V1 lora, I have found strength seems to be best between 0.40 - 0.8. Despite having plenty of areolae in its dataset, it does struggle, especially with nipples. It'll get it right eventually.