# Sakana AI and NYU Train a Diffusion Transformer to Generate Editable Minecraft Worlds

Dream-Cubed learns from billions of blocks with a 280-million-parameter 3D diffusion model, treating each cube as a token for controllable, inpaintable terrain.

- Published: 2026-07-30T05:37:30.825Z
- Canonical: https://polylog.news/ai/2026-07-30/sakana-ai-and-nyu-train-a-diffusion-transformer-to-generate
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Sakana AI](https://pub.sakana.ai/dream-cubed/), [arXiv](https://arxiv.org/abs/2604.22847), [Polylog editors](https://polylog.news)

Sakana AI and New York University have released Dream-Cubed, a system that generates playable, controllable Minecraft worlds by learning directly at the native block resolution. Each 32-by-32-by-32 chunk is fed into a roughly 280-million-pa…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-07-30/sakana-ai-and-nyu-train-a-diffusion-transformer-to-generate (subscription information: https://polylog.news/pricing).