Controllable Generative Modeling in Minecraft by Training on Billions of Cubes
13 hours ago
- Dream-Cubed is a large-scale dataset of Minecraft chunks (procedurally generated and human-authored) for generative modeling.
- It studies discrete and continuous 3D diffusion models for biome-conditioned chunk generation.
- Discrete masked diffusion enables inpainting, outpainting, and block-conditioned generation as a free byproduct.
- Data balancing ensures rare biomes like villages are adequately represented through targeted collection.
- Models are trained at full block resolution (32^3) using a Diffusion Transformer, without compression.
- Human evaluation shows generated chunks are often preferred over real ones, with FID metrics validated against human judgment.
- The dataset, code, and model weights are released to support further research on interactive 3D worlds.