v2.4: Trained with unet only. Small interference.
v2.5: Trained with unet and text encoder. More effective, but interfere with base model.
Breasts that sagging aside
