InfiniteDiffusion Brings Infinite, Seed-Consistent Terrain Generation to Consumer GPUs

Researcher Alexander Goslin has introduced InfiniteDiffusion, a training-free algorithm accepted to SIGGRAPH 2026 that reformulates diffusion sampling to enable unbounded, seamlessly tiled terrain generation. The approach bridges the realism of diffusion models with the properties that made procedural noise functions like Perlin noise indispensable — infinite extent, seed-consistency, and constant-time random access — without requiring persistent global state or additional model training.

Built on InfiniteDiffusion, the accompanying Terrain Diffusion framework positions itself as the first learned procedural terrain generator, using a hierarchical cascade of diffusion models to couple planetary-scale context with local detail. A compact Laplacian encoding stabilizes outputs across Earth-scale elevation ranges, and the system runs on consumer GPUs at speeds described as nine times faster than orbital velocity for a 1024×1024 relief map. The framework and an open-source infinite-tensor library for constant-memory manipulation of unbounded tensors are both being released alongside the paper.