v1: Trained on 3 episodes from the first OVA, only one outfit.

v2: Trained on the first 5 out of 6 episodes from the first OVA plus the special episode, three outfits (unfortunately it sometimes struggles to accurately generate the white outfit), and a more consistent dataset with better tagging redone from scratch. Also, different settings make this version more stylistically accurate at the cost of some flexibility, so I kept v1 if you like the cleaner and more modern look of it.

Many of the images had dark lighting, so some images may come out looking darker. Adding "night" to the negative prompt can eliminate that, though the colors may sometimes look a little saturated.

ADetailer (or FaceDetailer for ComfyUI) is highly recommended. Annoyingly though it tends to add black lines to her face even when they don't make sense, adding "blush" to the negative prompt sometimes helps with that.