Terrain Diffusion: Perlin noise’s learned, infinite successor

What it does

Terrain Diffusion generates planetary-scale heightmaps and climate data in real time using a hierarchical stack of diffusion models. It treats generative AI like a procedural function: you get infinite, deterministic terrain that is randomly accessible at any coordinate without generating the whole world first. The project ships with pre-trained models for 30 m and 90 m per-pixel resolution, plus a GUI explorer, an API, and even a Fabric Minecraft mod.

The interesting bit

The core trick is InfiniteDiffusion, a drop-in replacement for MultiDiffusion that makes sampling lazy and unbounded. Built on the infinite-tensor framework, it gives you O(1) random access and seed consistency, which means you can stream terrain like Perlin noise but with learned fidelity. There is also a self-contained annotated demo that generates an infinite panorama with Stable Diffusion v1.5 entirely independent of the terrain pipeline.

Key highlights

• InfiniteDiffusion provides deterministic, O(1) random access to unbounded diffusion outputs.
• Two pre-trained models target different use cases: 30 m/px for playable worlds and 90 m/px for expansive realism.
• Includes real-time streaming of elevation, temperature, and precipitation data.
• Can condition generation on hand-drawn or procedurally generated coarse maps, including exports from Azgaar’s Fantasy Map Generator.
• Ships with a Minecraft mod and a SIGGRAPH 2026 paper behind it.

Caveats

• macOS is CPU-only, and CPU inference is described as significantly slower.
• Customizing world generation beyond the pre-trained hierarchy requires retraining the coarse model or editing the synthetic Perlin-noise base maps directly.

Verdict

Worth exploring if you build open-world games, simulation tools, or just want to see what happens when diffusion models behave like deterministic noise functions. Skip it if you are looking for a polished end-user tool; this is clearly a research codebase with modding hooks.