Style LoRA trained on on NoobAI-XL vpred 1.0 with images from the twitter artist Ino_mahe / イのマヘ💡:
https://x.com/ino_mahe
Trigger word is not needed, only for the v1 version of this lora (jz235)
It works on mentioned model and its mixes:
https://civarchive.com/models/833294/noobai-xl-nai-xl
Negatives and quality tags are optional, but are not necessary at all. If you don't like the sketchy look of most of inomahe's art, you can put sketch in negatives (either via. the negative prompt or via. negpip) to make it look more modern.
Description
No more trigger word needed! Just plug and play.
Larger dataset
Pre-downscaled images before training using a new better custom image downscaling script
sd-scripts does this already, but I wanted to achieve a higher downscale quality (I can upload it later)
Introduced tag-drop at rates of 17% w. protected tags enabled during training, for training the uncond of the model. This is to make the lora's style work better on things it is not trained on (like 1girl stuff).
Higher batch size/grad acc to generalize better during training (corresponds to batch size of 4).