I'll probably rebuild the dataset and retrain afterwards, removing this version. But I will keep it for now, as it fits my workflow perfectly.
Well... This LoRA is shitty AF in terms of stability and strength cuz of experimental dataset. But! I really like the way it smooths/blends out the brushstrokes/colors/skin in other styles. So I decided to upload this version.
So yeah, here's my main recommendation: Don't use it, just skip it and be happy :^)
As a standalone style it is quite weak. But if you're a masochist like me who enjoys combining different artist tags, finding interesting combinations, then suffer. Let's cry together.
Here's some information that might be helpful:
Like STYLE 001 used as a SUPPORT LoRA, meaning you need to add artist tags or other LoRAs for better quality and stability.
Was trained to work with semi-realistic, realistic illustrations.
There is an activation tag: nhl-002, but you can leave it out. The difference is minimal.
More useful tags: realistic, painterly, nose, lips .
Trained on NoobAI-XL vpred 1.0 . So it may not work well on other illustrious models. I recommend using ΣΙΗ. My favorite merge at the moment.
Weight from 0.5 and up to 1. Here you need to adjust based on your own preferences. I've settled on 0.8. But look at your workflow, it's different for everyone.
If you are using ReForge and ΣΙΗ, Try to use RES Multistep/Euler A Schedule type: Beta. I don't know how it is in other UIs, so I can't say anything.
