Swooping-go-Round

A certain LoRA vanished into thin air, and it so happened to be one of my favourite ones. This is my attempt at recreating that model, with the Monet touch.

This model aims to do swooping and round breasts with or without puffy/pointy nipples and/or large areolae. Also Emilie from Genshin Impact is here, since I was looking for an excuse to make a LoRA for her; this one is perfect.

V1--SD1.5/PD/Ill

I trained this using the multiple training method on the on-site trainer, but an error occurred when submitting it, thus it duplicated the SD1.5 model. Since it still cost me buzz for that extra model—and for some reason the training steps seem to differ from the original, I’m keeping it and calling it model B.

I’ve never dealt with the Illustrious model in depth, thus this is my first attempt at making an Illustrious LoRA model. The day I can figure out how to make Illustrious models work with Hyper LoRAs, or simply get the steps down to 8 or less without sacrificing details, is the day I start switching and committing more to that model. In short, this Illustrious LoRA isn’t for me; it’s for you~♡

As I’m more comfortable with PD models these days, the PD version of this LoRA is here as a given. There’s a possibility I might do a refined version of the PD model, locally, after I complete my other project. As far as training parameters go, there’s no difference between this and the Illustrious model. I did however, dumb down some of the settings for the SD1.5 models so as to not overbake it.

V2--SD1.5/PD/Ill

Tried to do a major overhaul, ensuring that only actual swoops were tagged as such, everything else that resembled it but had either a slight swoop or no swoop at all was relegated as being simply perky.

Fig-1: Monet's hand-drawn animation; round vs swoop. Only the swoop version was in the dataset, as the animation came after V2 was completed.

Fig-2: Monet's hand drawn image of a proper swoop. A similar image was used for the dataset, but this one was drawn after V2 was completed.

To combat the fact that V1 fought me hard to reproduce the swooping effect, I incorporated some of my own hand drawn stuff in the dataset (though not what you see above, as these came after V2 was made).

To capture some of the styles, especially seeing how readily the Ill version adopts it, I've called the trigger for the style Nibelung Valesti... yeah. Couldn't come up with a better name, and I didn't want to use a bunch of random letters like "azngs", which I'd be sure to forget. The Nibelung Valesti style tries to replicate a certain hentai art style. If you know, you know.

Review the trigger words on the side. You can combine some of the trigger words together, but I don’t recommend combining ALL of them, unless you know what you’re doing.

Recommended weight of 0.8 or less; weights of 1 seems to have an adverse effect on quality, at least on the PD checkpoints I use these days.