Experimental LoRA. It messes up a fair amount of the time, but still an improvement over the base model.

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