Why tie it to a model? Just have the model as a parameter or config option.

Also, it didn't even dawn on me how easy it would be to use a MCP to tie together the main text-based LLM to a pure text-to-image generator.