Sakana AI and NYU Release Dream-Cubed: Minecraft Generation Models Trained on Billions of Blocks

Sakana AI, in collaboration with New York University, released Dream-Cubed, a project aiming to achieve controllable high-fidelity 3D world generation in Minecraft through a large-scale dataset and a family of generative models. The core motivation is to apply the success of scaling language models based on discrete tokens to Minecraft's block structure, offering a new path for 3D generative models.

Confirmed

• Large-scale dataset: A voxel-level Minecraft dataset containing billions of blocks from procedurally generated terrain and high-quality human-built maps.
• Model family: The team trained discrete and continuous 3D diffusion models to generate playable Minecraft worlds at block resolution.

Why it matters

• Analogy to language models: The authors note that language models benefit from words as natural compositional units, and Minecraft blocks share this discrete compositional property. Dream-Cubed validates the potential of scaling large 3D generative models using this structure, providing a new research direction for controllable 3D world generation.

2026-07-29 ~ 2026-07-30 · 6 related posts

• [source] Sakana AI and NYU release Dream-Cubed, a Minecraft dataset built from billions of cubes — SakanaAILabs · 2026-07-29
• Sakana AI Introduces Dream-Cubed: 3D Generative Models for Minecraft Trained on Billions of Blocks — hardmaru · 2026-07-30
• Sakana AI and NYU Release Dream-Cubed: AI Generates Playable Minecraft Worlds Trained on Billions of Cubes — tkasasagi · 2026-07-30
• Generating Controllable Minecraft Worlds by Training Diffusion Models on Billions of Cubes — tkasasagi · 2026-07-30

2 near-duplicate retellings: SimonGColton · SakanaAILabs