It is something I’m interesting in exploring. There are multiple challenges in AI map generation. I think it will be a lot of work to get right. In particular, maps are made up of many parts, and their image names alone do not describe the parts of the image, and often time are colorful and don’t even describe the map in a specific way.

For instance, a tavern map might be named “Weary Tavern” but how can the AI model know which parts are tables, chairs, stairs, bars, and so forth, so that they can be produced from a prompt?

I’m not sure how tagging parts of images for training works in stable diffusion but it’s on my list to check out when I have the bandwidth.