Benten is modeled after Benzaiten, a Buddhist deity among the Japanese Seven Lucky Gods. By depicting her in a chainmail bikini, riding a flying motorcycle, and carrying a high-tech bazooka, Takahashi aimed to make her resemble the stereotypical member of an outlaw motorcycle club. By giving her a tomboyish and violent temperament, Takahashi contrasted the character with Benzaiten, who is typically characterized as a clever, graceful, and artistic figure.
Moved the version notes to "About Version"
Description
Not all bikinis are armored bikinis now.
V2.1 (V2.1+ Pony): She can now wear a bikini that is not armored. I moved some prompts around and added trigger prompts for normal wear. I guess as sort of good news, the airbike gens is a bit better now (on Illu checks, but I haven't tried on Pony). Still wonky but if you want, give it a go.
If you have any issues with short hair, try adding a "single hair bun" prompt. On the other hand, if the chain hairband shows up with the long hair version, put "hairband" in the negs. It should help keep it in check.
FAQ
Comments (4)
tried it and gives me only wrong proportions and fuzzy images. I suspect something bad happened during the training...
Hmmm, I'm assuming you are talking about the Pony version 2.1. Tried different checkpoints? Otherwise, I can try making farther adjustments.
@Farfegnugen Yes, I tryed on Pony. But I see this is happening also on Illustrious, for example:
https://civitai.com/images/58526065
as you see, the body is strange... and if you look better, feel a bit uncorrect in most of the images posted...
Unfortunately, today the CivitAI generator is out of service and I cant perform experiments...
The other LORA about Ran and Oyuki were OK. Onlt Benten is strange...
@settima_ai Yeah, I had some troubles with Benten in Version 2, particularly with the Pony version. In retrospect, I should have shelved it and came back to it later. My objective with models is to keep them flexible enough to avoid the clothing bleed through that I've seen in a few, but also keeping them fairly consistent. Unfortunately, during training, it mostly did not want to be either.
I might just return for a V3 in April and try a different checkpoint and more data modifications. If anything, I will make the training data available.