V2 : Improved dataset tagging and training parameters. Impressive improvement, considering it took me 10 minutes. For some reason her skin got a little darker, absolutely no idea why considering I used the same dataset.

Artifacts were a problem on the first version, especially with complex prompts deviating from the training set. This version gives far superior results in those case (see comparisons), in addition to a - subjectively - improved image quality.

A few tag changes :

Main trigger : Andou Miyako

Appearance : Black hair, low twintails, rimless eyewear, grey eyes (main change !)

Outfit : Black serafuku, black skirt, white socks, loafers, (green neckerchief, sailor colar are mostly redundant)

Tags from v1 still work, just adjust eye color and trigger. You'll probably get better results with fewer tags compared to v1, as fidelity is better with this version.

________

I have been working on this one for a while. This is Andou Miyako, from Euphoria (fair warning for the pure and innocent people who ended up here : if you don't know what this game/hentai is about, I'd advise against looking it up. It's pretty dark and disturbing even by hentai standards.)

For fellow men and women of culture degenerates who wished we could have seen more of this character, here are the trained tags :

General appearance : AndouMiyako, long hair, black hair, low twintails, black eyes, glasses,

Usual outfit : black serafuku, black shirt, black skirt, green neckerchief, sailor collar, white socks, loafers,

Others : euphoria choker, white room, pubic hair

Redundant but sometimes useful for stability : school uniform, hair over shoulder, forehead, long sleeves,

Of course, trained over various poses, expressions, camera angle and nsfw situations, all tagged accordingly. Trained on IllustriousXL 0.1, tested with Wai NSFW Illustrious v14.0

Due to my tagging of the dataset, extensive description usually lead to better results. In particular : "forehead" usually improved fidelity, despite not being trained. Weird, but I recommend using it.