I think you could completely automate this by training a neural network on local neighborhoods to generate a voxel probability distribution for the center voxel

I think you could completely automate this by training a neural network on local neighborhoods to generate a voxel probability distribution for the center voxel, so a randomly initialized world would converge to a world that has pieces that resemble the worlds that the network was trained on. So in 1D you would take a sequence of voxels ABCDE, and the network would learn to predict C from A,B,D,E. Similar idea without neural networks: https://github.com/mxgmn/WaveFunctionCollapse https://github.com/mxgmn/WaveFunctionCollapse I'm surprised no one applied this idea to Minecraft yet. There's even a 3D version: https://i.imgur.com/IEOsbIy.png https://i.imgur.com/IEOsbIy.png