There's not that many western styles so I decided to try my hand at one. Suffices to say is very very very , so prompt accordingly, even though 90% of the dataset is scenes it can do pinups with almost no problem.
This was a very big learning experience; I must have trained around 20 times to try to get as close as the source material as possible, with a lot of different settings. In the end I couldn't 100% achieve was I was aiming for, but I got pretty close so for the time I'm sharing and moving on. I might update in the future as new techniques develop.
There are two flavors, the heavier 128 one is closer to the dataset style. It needed the extra data to better overcome the anime faces I think. Nonetheless, the smaller one does nice generations.
I don't think is necessary, but the activation tag is `btfrm`. I recommend weights from 0.8 to 1, some things are more overcooked/overfit than others so it might need tweaking here and there. Lower weights make some pretty good mix between anime and the style.
Some characters were trained and work with their names, but need some help sometimes with their eyes, hair, etc. I left examples in the 128 gallery.
I also made a LoHa that I will publish separate as I might update that one later down the road too.
I have a ko-fi now where you can get commissions for characters/face LoRAs, Early Access to the LoRAs that I'm training but are not finished or donate to my electricity bill. If you want to help me in a different manner, there's my wife DA where you can subscribe and pics on return.
And I forgot some thanks to @LazyTrainer and RocketBoo(I can't figure your user sorry) for all their suggestions when I was having problems and @guy907223982 for inspiration and talking be about his training on other games, go check them out.
Metadata is there but don't hesitate if you want to ask about the training.
