单人科学家发明无限地形生成器，融合扩散模型与多尺度细节

Imagine you could hold an entire planet in your hands, a world with mountains that never existed, coastlines stretching past the horizon in every direction forever. A solo scientist invented something incredible here. One, it is a truly infinite terrain generator. This panel shows a region that is roughly the size of a country like Congo, millions of square miles of land. And if you would zoom in to one of these regions over and over, it just keeps going and going.

Amazing. Now, many of the examples are shown within Minecraft, but it can generate fully continuous worlds depending on the game engine being used. But it gets better. Two, this generator also learns. More on why that is amazing in a moment. Now, wait, wait, wait. We already have terrain generator programs, a ton of them. So, is this really new? Well, there are existing methods that build terrain out of noise. These can generate infinitely in every direction, but they aren't that organic.

They are a bit too uniform, too repetitive. It just generates new stuff with no plan for the whole planet, no large-scale coherence. So, these methods don't really learn. This is the price of speed. Or, there are AI-based methods. These can learn. Yes, you can feed them the statistical distribution of real terrain from Earth and have it generate something similar without copying. Ooh, that is amazing. But unfortunately, they are quite inefficient.

Why? Well, because every newly generated area depends on every other area in the world. That's how you get your coherence. It has to know about everything. Ooh, but it's a nightmare in In of speed. This takes forever. So, learning or speed, choose one. Story of the last 40 years. But, this new research essentially fuses these two into one technique that has the advantages of both. Is that even possible? After 40 years of using noise to create virtual worlds, does anyone think there is a better way?

Let's have a look. Dear fellow scholars, this is Two Minute Papers with Dr. Karoly Zsolnai-Feher. This key formula is brilliant. First, it uses diffusion to create these terrains. So, much like the image generator AI systems of today, it starts out from noise, and it slowly reorganizes it to an image. But, this concept is now adapted to terrains. Okay? That is a cool building block, but that is not new. Now, here comes the brilliance.

One, this formula tells us what a new region R should look like. It says, "Well, 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 spot, and ignore everyone else. Now, this this is genius. Why? Well, by asking only the neighbors, it 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. Hmm. Now, that is an amazing property. For instance, it lets you teleport millions of miles instantly. How cool is that? Love it. But, we have a problem. The problem is that real terrain can span huge height differences, like an ocean trench to Mount Everest. Huge variation. However, the part that really makes terrain look like terrain is only a few feet tall.

Ridges, river banks, and all kinds of textures. Diffusion techniques can't deal with that. They either focus on the small terrain variations or large mountains, but not the two at the same time. So, what do we do? Now, genius idea number two. This is called the Laplacian re-extraction denoising for height maps. The trick says, "Do not denoise the heights as a raw signal." No, sir. Get this. Imagine photographing your friend standing at the foot of the mountain.

Hmm, if you frame the mountain in the shot, the person becomes a tiny little dot. Well, then, of course, focus on the person then. Well, if you do that, then you don't see any of the mountain. So, what do we do? Well, here is the brilliance of the Laplacian trick. Are you ready? Okay, so first, you take an image of the mountain, a perfect image. Then, you take a separate photo of the person on their own scale. So, you see them properly.

And then, you put the two together onto one photo. Yep, now you retain all the detail about both. So good. And this is what this tap does mathematically. And that is also how this technique generates terrain on multiple scales. So, finally, the mountains and the creeks have an even fight. So, finally, an efficient technique that can also learn about Earth or any other kind of data and generate new planets. And get this, as I keep reading the paper, it just gets better and better.

Now, hold onto your papers, fellow scholars, because this was trained in 2 weeks and was run on a 4-year-old consumer GPU interactively. The code and Minecraft mod are available for free. The power of open science. What a time to be alive. And this was written by a solo scientist. The paper was published at SIGGRAPH, the most prestigious conference in computer graphics. This is incredible. This is a bit like showing up at the Olympics alone, without a team, and winning a gold medal.

Yep, an independent scientist. That is an amazing achievement. And he just gives it all away to all of us for free. Thank you so much and huge congratulations. I use Lambda to reproduce AI research papers often in minutes. It's also great to train your own models or fine-tune an existing one. Run inference or text to image or video, easy peasy. Running a deep seek chatbot or agent, super fast, super reliable. Lambda gives you powerful Nvidia GPUs to run your own experiments.

I test ideas from the papers I cover and moments later, results.

[snorts]

Love it. Seriously, try it out now at lambda.ai/papers.