Dreaming in Voxels: How AI is Generating Playable Minecraft Worlds

Sakana AI's Dream-Cubed, developed with New York University, generates Minecraft environments as arrangements of editable game blocks. The models learn from procedurally generated terrain and human-built maps, using individual cubes as the units of generation. The research announcement describes outputs that players can enter, modify and play.

This representation also supports filling in selected regions and extending a world beyond its existing boundaries. The contribution is a dataset and model family for producing usable block-based environments, giving researchers a way to study generation where the output must function inside a game.