AI Minecraft World Generator by Sakana AI

What Sakana AI built as a Minecraft world generator

Dream-Cubed is a dataset and a set of AI models that generate Minecraft worlds at the resolution of single blocks. Sakana AI built it with New York University and announced it on July 29, 2026. The paper is on arXiv as 2604.22847, and its title begins "Dream-Cubed: Controllable Generative Modeling."

Two groups should care. Players and map creators get worlds they can edit and play straight away, not a 3D picture. Researchers get the full data, code and trained models to build on.

Why a Minecraft AI can treat blocks like words

A Minecraft world is already split into named blocks, the way text is split into words. Each block is one thing: sand, stone, water. AI researchers call this kind of 3D grid "voxels."

Language models learn from text that comes cut into words. In the same way, a model can read a patch of Minecraft terrain as a sequence of block "tokens" and learn which ones tend to sit next to each other.

Sakana AI also points to a gap. Image, video and text generators have moved fast, while the interactive 3D worlds of video games have had far less research. Minecraft is the best-selling video game in history and already a favorite testbed for AI research.

A Minecraft dataset of 30 billion blocks and 15 biomes

The training data holds over 30 billion blocks, drawn from game-generated terrain and from hand-built maps. The terrain is cut into cubes of 32×32×32 blocks called chunks. The hand-built maps come from a professional map creator who gave permission to use them.

Generated terrain is close to unlimited, but uneven. A biome is a region type like desert or ocean, and some types show up far more often than others.

Villages are the hardest case: they hold the most complex structures and are also the rarest. Sakana AI set up collection pipelines that hunt for specific biomes so rare ones still get enough samples.

Hand-built maps are the opposite problem. They are scarce, use many more block types, and contain far more detailed buildings. Sakana AI states that it releases the data and models with the map creator's agreement.

How the Minecraft diffusion model generates terrain

Dream-Cubed generates terrain with a diffusion model, a generator that learns by scrambling data and then restoring it. During training, real data is scrambled step by step, and the model learns to undo each step.

To create something new, it starts from fully scrambled input and repairs it until a clean result appears. Many image generators work this way, and Dream-Cubed applies the same idea to blocks.

Training at native block resolution without compression

Dream-Cubed feeds raw blocks to the model with no compression step, trading speed for control. Image diffusion models usually squeeze each picture into a smaller code first, using a helper model called a variational autoencoder (VAE). That makes training much cheaper.

Here, each 32×32×32 chunk goes straight into a 3D Diffusion Transformer with about 280 million parameters. Sakana AI accepts the lost efficiency because the model then sees every single block. That is what later lets a user pin individual blocks in place and have the model respect them.

Two diffusion models for Minecraft blocks: MD4 and DDPM

A block is a category, not a number: "stone" is not halfway between "sand" and "water." Sakana AI tried two ways around this, MD4 and DDPM, on the same Transformer.

• MD4 (masked diffusion): blocks are hidden behind a special [MASK] token during training. To generate, the model guesses which real block belongs under each mask and fills it in.
• DDPM (noise in embedding space): each block's in-game name, such as "stone," is turned into a 16-dimensional number vector by a frozen text model. The model removes noise from these vectors, and each cleaned vector is matched to the nearest real block.

Both models learned to produce believable chunks for a chosen biome. The deciding difference is what MD4 can do beyond single chunks, covered next.

Inpainting structures and generating worlds of any size

MD4 can fill in around blocks a user places and can extend a world without limit, with no extra training. Inpainting means filling in the missing part of something while keeping the rest. Since MD4 already generates by filling in masked blocks, any block a user places is simply treated as already filled, and the model has to build around it.

In practice, a creator can lay down a few blocks in a spiral, a slope or even a volcano shape and let the model complete a natural-looking chunk around them. The same trick works on terrain cut from a real game world.

Outpainting stretches this past the edge of one chunk. The model slides its window so each new chunk overlaps the last one, keeps the shared blocks fixed, and fills the rest. Repeating this stitches together a world of any size.

How the AI world generator scored against real Minecraft

The generated terrain held up against real Minecraft in a player study, and the training data mix shaped where it was weak. Sakana AI tested this three ways: by changing the data mix, by asking players, and by checking its automatic score against those players.

Minecraft biome balance drove the generated map results

A model gets good at the biomes it sees often, so the data mix decided which terrain came out well. To show this, Sakana AI built three versions of the dataset with 1 million chunks each and trained on each one.

With the game's natural mix, oceans flood the data and villages almost vanish, so the model struggles with villages. Doubling the village share in the Boosted mix brought a significant improvement.

Patch size also mattered. A patch is the group of blocks the Transformer reads as one unit, and models using size 4 did noticeably worse on villages than those using size 2. Sakana AI reads this as a sign that finer detail matters for complex structures.

Minecraft players preferred generated chunks in all 3 models

Players chose the generated chunk over the real one more often than not, for every model tested. Each person saw two sets of chunks from the same biome and picked the one they thought was better. One set came from a model, the other from real Minecraft terrain.

MD4 at patch size 2 was preferred in 67.1% of trials; the other two, at 57.1% and 55.2%, sit just above an even split. Sakana AI adds its own caveat: the result may come partly from how it cleaned and picked the data. It reads the numbers as the output being at least on par with real terrain, not as proof that it is better.

Checking the FID metric against Minecraft player votes

Sakana AI's automatic quality score agreed with players more closely when the gap between two models was large. The score is FID (Fréchet Inception Distance), a standard measure for image generators that compares generated pictures with real ones; lower is better. No such score exists for 3D block worlds, so Sakana AI rendered chunks as images and measured FID on those.

That only helps if the score tracks what people actually like, so Sakana AI checked it against the player votes. When two models scored almost the same, the link between FID and the votes was weaker.

When the FID gap was larger than 15 points, the model with the lower (better) FID won 66.1% of the time. Sakana AI calls this a limited but significant link.

Human-authored Minecraft maps cut to 171 block types

For hand-built maps, Sakana AI cut the thousands of block types they use down to 171. Each map is also treated as its own extra "biome," the same kind of label the generated terrain carries. Sakana AI adds that learning from data that holds human creativity raises new problems, including ethical ones, and presents this part of the work as a first step.

What the Sakana AI release includes for Minecraft research

The release is a full research kit: paper, code, datasets and trained model weights, with no app or in-game button. Four pieces are public:

• Technical blog at pub.sakana.ai, with 3D samples of generated chunks and worlds
• Paper on arXiv (2604.22847), with the full method and results
• Code on GitHub (SakanaAI/DreamCubed) for training, evaluation and visualization
• Data and models on Hugging Face: raw and processed datasets, the curated human-built maps, and the trained model weights

For a player who just wants a new world to explore, this is not a finished tool yet. For anyone who can run the training code, the data, code and weights are all in one place.

Summary: a playable Minecraft world generator with caveats

Dream-Cubed is an AI Minecraft world generator whose output is real game terrain, ready to edit and play. That comes from one decision: training on raw blocks with no compression. The same decision makes inpainting and worlds of any size possible with the MD4 model.

The headline result, that players preferred generated chunks in all 3 models, needs its caveat. MD4 at patch size 2 won 67.1% of trials, while the other two models won 57.1% and 55.2%, and Sakana AI itself says data curation may account for part of the win. The safest reading is Sakana AI's own: the generated terrain holds up against the real thing.