Grounding rule-specified entities to actual visual objects in generated scenes, even when those objects haven't appeared yet.