In at least my limited experience, they're kind of bad. They can retrieve shapes that already exist, sometimes inaccurately, but they are less reliable at creating novel ones
Regular LLMs are quite bad at it (see simonwillison's blog post). However this paper [0] describes an apparently sound approach using Neural Radiance Fields (NeRFs), however their github repo [1] has been "code coming soon!" for months now, so you can't really use it.