Liepard LoRA trained on NovelAI with Kohya training scripts using the DAdaptAdam optimizer at 512x512 resolution. Has issues getting the fur patterns right, there should be yellow oval markings and her forearms and legs from the shin down should be yellow. Tail placement is unreliable, with the tail not showing up sometimes. Prompting for (tail) and (thin tail) can help with that. Should have wide compatibility with model mixes based on NovelAI, which is most of them. Trained on an all female dataset.
Description
Initial release
FAQ
Comments (5)
Another great choice! Very underrated 🐱
Out of curiosity, why are you training all of these on yiffymix 2.1 rather than the newer versions? I'm trying to get into making loras myself and id like some advice if youre able to give it
I’m using YiffyMix 2.1 to generate the example images because I like the way it looks with my current prompting style. For the actual training, I’m using the leaked NovelAI model from way back because the vast majority of model mixes have ancestry that traces back to that base model. I highly recommend using proper base models for LoRA training instead of model mixes to get more widely compatible and stable results.
@KomodoStar i see, where can i find the NovelAI model? also, is it a general purpose model, ie can it be used to train non-furry characters as well? Because I used AOM3 for a shygal lora that turned out pretty alright, but if theres ways to improve it ill happily remake it again
@spunchebop The NovelAI model was leaked a while back, so you'll have to sniff around for a torrent of it if it's still up. It's general purpose, more skewed towards anime girls than furry from what I've seen. If you get good LoRA results using a model mix, then keep using it. If it works, it works ❤️. AOM3 is itself mixed into a bunch of other models, so LoRAs trained on it will inherit that compatibility.