Important: This LoRA works best as a Text-encoder-only LoRA. Use it like this: <lora:EnvyAttitudeAdjustmentXL01:0:0.7>
Note: the example pictures were created in ComfyUI, and I've weighted this lora at 0:0.7, 0:0, and 0:-0.45.
I stumbled on this kind of by accident while playing around with newer Kohya_ss features, specifically the fact that it now trains LoRAs that affect both the unet and the text encoder. As I was experimenting with it, I didn't particularly like the results I was getting using it the normal way, but I discovered that it had some interesting effects if I turned off the UNET part altogether. If I knew of a way to do it, I would just zero out the UNET stuff directly in the LoRA, but I'm not aware of any programs that are able to do that right now (I wish there were a python script or something that could manipulate every aspect of a LoRA's weight, but if that exists, I'm not aware of it).
Anyway, this LoRA gives characters a confident, punk look when weighted above 1, and a shy look when weighted in the negatives.