Experimental LoRA. It messes up a lot of the time, but it's still a big improvement over the base model when it does work.

Works best when prompted with natural language (see sample images).