This is a LORA trained on the style of Kan Liu. There's no trigger tags or anything, so just prompt normally and have fun. The XL version usually does a good job staying out of the way on composition details and sticking to just handling style stuff. Unlike the V1 and v2 models, you can even push the faces out of the typical Kan Liu range and into the wild PonyXL range by simply prompting for the expressions you're familiar with from the model. This might be a negative for some, but I personally enjoy the flexibility I got out of v3.The 1.x versions are a bit different, mostly down to stiffer composition and worse faces and hands. You can still have fun with it on SD1.x anime models, and v2 does some wonderful stuff when prompted for ink, shadows, hydrokineses, and chromatic aberration, but I can't really say as many nice things about it as the XL version.Anyway, check the version descriptions. Minutia about each version is included there.

Kan Liu style LORA

This version chases the same goals as the first two, but pose-locking issues from the 1.x versions have been kinda accidentally solved. I'm not going super serious with this round of training, so this might be the only one I release for at least a while. I do still want to solve some of the issues below, though. Anyway, just try to have fun with it if it's your cup of tea. That's why most of us are training and genning stuff, after all.Known biases:Hair styles - Light, often adds topknot hairstyles, but you'll usually be able to prompt out of it.Clothing - Heavy; adds some often nice but usually unasked-for embellishments to clothing.Skin tones - Heavy; in my testing, "dark skin" rarely works. "Tan" and "tanlines" still mostly work, but this version tends to keep things really close to the shades present in the dataset at normal weights. Try reducing weight a bit if this gets in the way for you.Faces - Inconsistent. <s>If you take it past about 0.7 weight, you might need to inpaint or use Adetailer to fix things. It'll still usually do alright for close-up faces, but that's like praising a model for genning an image at all.</s> Y'know, I've had a pretty similar distribution (compared to PonyXL by itself) of hosed and good faces from this lora between 0.6 and 1.0 that I'm not entirely sure if the lora's actually causing the problems here.As for the example gens, yeah they're probably a bit too blended with other style elements to be considered a clean representation of the style. I can post some new example gens if anyone wants me to drive a bit closer to the source style.

Files with SHA256: ed3d48b00bb0590d96c22fdb1473b6db92b15e5c4f6ad918453f6a4c813d5242