A simple LoRA model trained with a small set of images to recreate a Venti-inspired character design that blends feminine and androgynous features.
This model focuses on maintaining Venti’s recognizable aesthetic while adding a softer, more gender-neutral appearance.
Description
not recommend
FAQ
Comments (4)
Lots of v.bad, but what makes a v.good?
Decided to test the 8 current versions. V0.0, V.Bad1, V.bad6, and what I would assume will become V.Bad7 are the best currently on this model. Honestly, I would upgrade V.Bad1 and V.Bad6 into alt versions. Example. V0.5 (was V0.0), V0.5A (was V.Bad1), V0.5B (was V.Bad6), V0.5C (was V.Bad7?).
V.Bad4 is defiantly the worst here. Capable of alright image's if you lower strength but defiantly not "intended."
V.Bad3 and V.Bad5 have furry tendencies when select words are being prompted.
V.Bad2 also went furry after lowering the strength with some words in prompt.
V.Bad2 and V.Bad6 had the best Egyptian themed pictures. V.Bad2 is fairly good minus furry tendency. Its on the cinematic side which can have some appeal.
Personally, I think v0.0 and V.Bad7 (or whatever it gets named) are the best here, V.Bad1 is also very good followed by V.Bad6. Ill be running more tests with these version in particular going forward.
I think none of the current versions are working well. Some versions suffer from significant noise, like background artifacts or replicated clothing. I will continue testing until I get a stable build that avoids overfitting or any other issues. Honestly, I didn't put the newest version at the top because I didn't want anyone to waste their Buzz on it. Instead, I will put the version I think is better at the very top.
@QianYue_ they seemed to work just fine on my model, but testing these in bulk as i was is expensive time wise. So; ts possible my limited samples just didn't have artifacts.
I do want to note that my checkpoint is not your preferred so my opinions are bias.
Edit: medel I'm using is my chaos illustrious v3. #notAnAd