My first try at LoRA merges. I trained separate slider LoRAs on lips / hips / waist / breasts and merged them together. By adjusting the weights I got rid of the color shift except in the extremes where the character becomes paler / tanner.
Without a character LoRA I find weights from -10.0 to 10.0 are okay-ish, but with LoRAs and to prevent too much color shift or artifacts, -8.0 to 8.0 work great.
As always with sliders, no specific prompt needed.
Note: It's been trained with "woman", but it also increases the "glamorous" look of men (muscle, tan).
Description
FAQ
Comments (6)
thanks! it works great!, using 8.00, but there is one thing. Body types in the training data have a very strong effect. So anytime I use 8.00 that body type is always the same, that very tall, long, skinny, and lanky women; unless I prompt very hard to remove it.
Yeah, unfortunately that's what z-image "thinks" of the women that fall in each sub-lora I made for this merge so as this is a merge, the effect is amplified :(
This is amazing. Great job on merging the loras properly.
Care to share your workflow for the merge?
Thanks, it's not a workflow it's a python script using the diffusers library :)
How do you guys train sliders for ZIT. When I try to train a slider lora in AIToolkit it throws an error.
ooooomg what is + do , what is - do , clarification please !