InfiniteDiffusion introduces a training-free algorithm that adapts diffusion models for lazy, unbounded generation, effectively merging the high fidelity of learned models with the infinite extent and deterministic access of procedural noise. The accompanying Terrain Diffusion framework demonstrates this capability by generating realistic planetary terrain at speeds nine times faster than orbital velocity on consumer hardware, utilizing hierarchical models and constant-memory tensor manipulation.
Background
Procedural generation has long relied on mathematical noise functions like Perlin noise for infinite, deterministic world-building, while diffusion models excel at photorealistic content but are typically limited to fixed-size images. This work addresses the industry need for scalable, high-fidelity assets in open-world environments without the computational overhead of traditional auto-regressive methods.
- Source
- Lobsters
- Published
- Jul 13, 2026 at 03:56 AM
- Score
- 8.0 / 10