You can definitely do that now for images as long as people can wait two seconds. But also normally MUDs have most locations already described so they could be pre-generated.
The thing about it is that SDXL understands intent much better so to get the really good visuals you would need to wait for that or DALL-E 3 which isn't quite as fast as the really fast diffusion models.
But still within 10 seconds or so I think you can do that for many things. So I am pretty sure that will be a thing.
I am working on something combining something like a simplistic game engine in an LLM agent and planning to include an action for generating an image.
Actually I think AI Dungeon already demoed something like that for their upcoming release.