I always thought we should be training models to edit images via commands instead of generating them from scratch if we want precision, but maybe it won't be needed after all.