InfiniteDiffusion: Bridging Learned Fidelity and Procedural for Terrain Gen.
11 hours ago
- #terrain-generation
- #diffusion-models
- #procedural-generation
- InfiniteDiffusion is a training-free algorithm that enables diffusion models to perform unbounded, stateless, and lazy generation while maintaining high fidelity and random access.
- It transforms any diffusion model into an infinite, logically stateless array indexed by seed and coordinates, supporting O(1) random access, determinism, and parallelism.
- Terrain Diffusion is the first learned procedural terrain generator built on InfiniteDiffusion, handling massive dynamic ranges and large-scale features like continents.
- InfiniteDiffusion breaks the trilemma of content generation by achieving infinite extent, stateless generation, and learned realism simultaneously, unlike diffusion models, procedural noise, or auto-regressive methods.
- It outperforms auto-regressive methods in random access, determinism, error handling, parallelization, and statelessness, while being training-free.
- Terrain Diffusion operates interactively on consumer hardware, running 9 times faster than orbital velocity in Unity and integrated into Minecraft as an open-source mod.