What about using Google image search?
There are some good ones, I’ll try…
Yes, it fills the whole scene with different content. Only the massive tree(?) in the background looks like some kind of copy and paste…
It’s probably some kind of magic tree that would be cleverly explained away with “look at” text.
I also spotted some off-brand Escher.
You could also experiment with “simple” maps like these:
Let the AI generate a simple map and then photoshop the buildings as symbols on the right spots. But I’m not sure if the AIs are able to generate such simple “cartographic” maps.
Trained from pixel art monkey 2, only day scenes, mostly exteriors:
Good colors but blocky, fuzzy details. If only we had more original scans…
Maybe I should first upscale with ESRGAN, and then train…
Yeah, the issue with maps is the AI wants to add a bunch of random things to your map that you don’t want, and it’s hard to specify what you do want…
It seems easier to just make it more blank terrain (trees, mountains, etc.), and then add the details. The MI Melee Island map only has little lights to show locations, so that’s very doable.
But some of them aren’t bad. Especially if you scale them down to the VGA resolution. One good example is the graphic for “The Secret Of The Pirate Flamingo”:
I would pay good money to play this game.
I like the mood of that one

Though maybe it looks a bit too realistic. Did you accidentally include photos from a trip to Venice in the training data?
.
Here are some useful resources for when the puzzles need to be made - puzzle dependency graphs for MI, MI2, Curse, and DOTT





































