what you want as a "stabilizer" is a good preference-optimized finetune, it can be a merge crap but usually merges work worse. you don't want to collapse the output distribution of a model with a lora because it will mess a lot of things up, especially if you are trying to use multiple loras.
what you will get out of a "good" preference-optimized finetune is a certain, defined "plastic" look of flux, piss tint and aco seeping through on pony, and the like.