Two Minute Papers (Dr. Károly Zsolnai-Fehér) covers a paper by a solo scientist that delivers a truly infinite terrain generator. Many of the demos run inside Minecraft, but the method can also generate fully continuous worlds depending on the game engine. The key claim is that it breaks the 40-year-old tradeoff in virtual world generation: previous methods could be either fast (noise-based) or able to learn from real data (but painfully slow), never both. This technique fuses the two.
Károly opens with the pitch: hold a whole planet, with mountains that never existed and coastlines stretching past the horizon forever. A solo scientist built a truly infinite terrain generator. One shown panel covers a region roughly the size of Congo, millions of square miles of land, and zooming into any region repeatedly just keeps producing more world. Many examples are shown inside Minecraft, but the method can generate fully continuous worlds depending on the game engine used. And crucially, this generator also learns.
We already have plenty of terrain generators, so is this really new? Existing noise-based methods can generate infinitely in every direction, but the output is not organic. It is too uniform and too repetitive, generating stuff with no plan for the whole planet and no large-scale coherence. These methods do not learn. That is the price of speed.
The alternative family of methods can learn. You feed them the statistical distribution of real terrain from Earth and they generate something similar without copying it. The catch is that they are quite inefficient: every newly generated area depends on every other area in the world. That global dependency is exactly how you get coherence, since the model has to know about everything, but it is a nightmare for speed. So for the last 40 years the choice has been learning or speed, pick one.
The new work fuses both sides. First, it uses diffusion to create the terrains, much like today's image generator systems: start from noise and slowly reorganize it into an image, only here the concept is adapted to terrains. That is a cool building block, but not new. The brilliance is the formula for what a new region R should look like: ask a bunch of overlapping windows that touch region R, run denoising on each, and take their weighted average. In other words, blur together the opinions of the neighbors who can see this patch and ignore everyone else.
By asking only the neighbors, the method decouples the cost of the query from the size of the world. In simpler words, as the world you generate grows, the technique does not get slower. That is an amazing property, and it enables things like teleporting millions of miles instantly.
A problem remains. Real terrain spans huge height differences, from an ocean trench to Mount Everest. But the part that really makes terrain look like terrain (ridges, riverbanks, all kinds of textures) is only a few feet tall. Diffusion techniques cannot handle both: they either focus on the small terrain variations or on the large mountains, not the two at the same time.
The answer is Laplacian re-extraction denoising for height maps. Do not denoise the heights as a raw signal. Imagine photographing your friend standing at the foot of a mountain: frame the mountain and the person becomes a tiny dot; focus on the person and you lose the mountain. The trick is to take a perfect image of the mountain, take a separate photo of the person at their own scale, and then put the two together into one photo, retaining all the detail about both. That is what this step does mathematically, and it is how the technique generates terrain on multiple scales. Finally, the mountains and the creeks get an even fight.
The result is an efficient technique that can also learn about Earth (or any other kind of data) and generate new planets. It gets better as you read the paper: it was trained in two weeks and runs interactively on a four-year-old consumer GPU. The code and Minecraft mod are available for free, which Károly calls the power of open science. The paper was written by a solo, independent scientist and published at SIGGRAPH, the most prestigious conference in computer graphics, which he compares to showing up at the Olympics alone, without a team, and winning a gold medal. The video closes with a Lambda sponsor segment about renting GPUs to reproduce AI research papers.