December 10, 2025
Noise, but make it infinite
Terrain Diffusion: A Diffusion-Based Successor to Perlin Noise
AI promises endless worlds; gamers cheer, devs ask “is it fast”
TLDR: An AI technique claims endless, coherent game worlds as a successor to Perlin noise. Comments split between hype over the Minecraft demo and doubts about real-time speed and control, with veterans skeptical it’ll replace classic methods while others dream of non‑Earth planets and flashy textures.
A solo researcher just dropped Terrain Diffusion, pitching an AI-powered successor to the old-school Perlin noise that built your favorite game maps. The promise: truly infinite, seed-consistent worlds in real time, using a clever “InfiniteDiffusion” trick and a stack of models to blend planet-scale context with local detail. The demos on Minecraft and the open-source code had hype rolling in—one fan cheered that it’s “scalable to extremely large maps.”
Then the thread split. The Perlin purists showed up, saying this doesn’t feel like the right way to solve terrain and that they’d rather have pure code than training data. Speed became the main battleground: a master’s grad said they hit the same wall—pretty worlds, but too slow for live gameplay—while another pro sniffed that real terrain is usually guided by hand because it’s faster and more controllable. One practical take: great for textures, maybe, but not replacing classic noise without tight control tools.
Meanwhile, space nerds asked for lava moons and airless planets—can it do non‑Earth vibes? The memes flew: “Perlin hive vs Diffusion enjoyers,” jokes about AI painting continents like Bob Ross, and that eternal dev question: “cool demo, but can it ship?”
Key Points
- •The paper proposes Terrain Diffusion as a diffusion-based successor to Perlin noise for procedural terrain generation.
- •InfiniteDiffusion is introduced to enable seamless, real-time synthesis of infinite landscapes with seed-consistency and constant-time random access.
- •A hierarchical stack of diffusion models combines planetary context with local detail for coherent outputs.
- •A compact Laplacian encoding stabilizes outputs across Earth-scale dynamic ranges.
- •An open-source infinite-tensor framework and few-step consistency distillation support efficient, constant-memory generation over unbounded domains.