Sakana AIDreamCubed

DreamCubedDP2

DreamCubedDP2 is a pretrained Discrete MD4 checkpoint trained on a balanced natural-biome dataset.

Research release

The checkpoint generates block-based worlds

The released model uses two-block patches. The project supports biome-conditioned regeneration while keeping part of an existing chunk fixed.DreamCubed pretrained model list and generation workflows

Check the generated structure

Practical guidance. Compare the same source chunk and target biome across runs. Check whether the retained structure still connects to the newly generated terrain.

Specifications & access

DreamCubedDP2 specifications
DetailValue and conditions
StatusResearch release
AccessDownloadable
Model identifiersSakanaAI/DreamCubedDP2
Published variantDiscrete MD4; patch size 2.DreamCubed pretrained model list and generation workflows
Weight termsApache 2.0 in the checkpoint metadata.DreamCubedDP2 official model card