I was being asked how I make my models, I had to draw it out so it'd be easier to understand.
This was made fast and dirty so don't let your OCD take over, the cubes are not aligned, the words are not proper.
By model that can fix it, I mean your best trained base model.
For me, I extracted my SDnew, which could only do really ruined versions of furry characters, then applied it next to the lora i mixed to make SDnew, so both the extracted SDnew and the lora worked together to make SDnewR.
Check images constantly before rebuilding the lora back in.
This lets you almost refine a consistent model without much alteration, unless you want a lot of alteration.
You can train lots of concepts then merge them all back in as a lora with supermerger at levels of your choosing, even getting to view and test each image with the loras before rebuilding the models.
Really lets you refine which parts were lacking without dealing with the AI picking random parts to learn.
I'm sure other people have wrote something similar, but this is just for those who've not found this method yet and so others can understand my process.