damn,.... this is so incredibly cool use case for discrete diffusion model

original post :

Check out our best-selling Impact Case for #iPhone16, now with a kickstand for hands-free viewing! Get yours today. #CASETiFY

he’s late

Wow this is great

For those wondering: the model was trained from scratch (~23M params) on ~3k example buildings from Greenfield map. Each block type has a learned representation. A 3D denoising diffusion model (U-Net-based) learns to reconstruct plausible structures from noise, over 1k steps. Outputs are matched via nearest-neighbor search to valid Minecraft block types. Inference runs via ONNX + TensorRT in C++; the Java mod pipes in world context, triggers inference, and reads back block IDs—acting as a bridge between Minecraft and the model.

you can just do shit

The applications of discrete diffusion models are getting more impressive by the day. This is a great example of how the technique can be applied to real-world problems with impressive results!

@aidan_mclau

But no joke, this scaled up a lot... dude this could do 3d model generation, world generation (not just minecraft) and so much more....

@elonmusk We need to train a giant model for that to create entire Minecraft worlds with it. Can you provide capital or GPUs? It is a requirement for the future of humanity

@Frostgon dios mio.