I think that kind of multi-modal work is ongoing but not as advanced as text-based LLMs are today.

All these kinds of generators are LLM-via-text-proxy in the sense that people are using LLM's excellent text generation properties to generate via scripting interfaces in various tools.