The work was trained entirely on one RTX 3090 Ti.
Introducing InfiniteDiffusion, my independent paper accepted to #SIGGRAPH2026!
I have one RTX 3090 Ti. No funding, advisors, or team. By day I'm a new grad SWE at Walmart.
The paper has two main contributions: - InfiniteDiffusion: a new approach to infinite generation with diffusion models. - Terrain Diffusion: the world’s first learned procedural terrain generator.
Here’s why this matters, and how they are connected. 🧵
Many users expressed excitement over InfiniteDiffusion for its novel lazy infinite generation without global state, congratulating the solo researcher on the SIGGRAPH 2026 acceptance and open-source release.
No Digg Deeper questions have been answered for this story yet.
siggraph genuinely has the coolest stuff
Introducing InfiniteDiffusion, my independent paper accepted to #SIGGRAPH2026!
I have one RTX 3090 Ti. No funding, advisors, or team. By day I'm a new grad SWE at Walmart.
The paper has two main contributions: - InfiniteDiffusion: a new approach to infinite generation with diffusion models. - Terrain Diffusion: the world’s first learned procedural terrain generator.
Here’s why this matters, and how they are connected. 🧵
Code (+ Mod & API): https://github.com/xandergos/terrain-diffusion Project Site: https://xandergos.github.io/terrain-diffusion/ Paper: https://arxiv.org/abs/2512.08309 Minecraft Mod: https://modrinth.com/mod/terrain-diffusion
Respect!
Introducing InfiniteDiffusion, my independent paper accepted to #SIGGRAPH2026!
I have one RTX 3090 Ti. No funding, advisors, or team. By day I'm a new grad SWE at Walmart.
The paper has two main contributions: - InfiniteDiffusion: a new approach to infinite generation with diffusion models. - Terrain Diffusion: the world’s first learned procedural terrain generator.
Here’s why this matters, and how they are connected. 🧵
Unlike autoregression, InfiniteDiffusion is deterministic, randomly accessible, and effectively stateless: - Worlds are shareable by seed - Teleportation is O(1) instead of O(n) - It works in multiplayer and across distributed clusters - Zero persistent storage required
The result is that InfiniteDiffusion offers the same contract that made procedural noise the backbone of world generation for 40 years: any region can be efficiently queried independently, at any time, in any order, and the result is guaranteed to be identical regardless of what has been queried before.
By inheriting that contract, diffusion can power the same class of applications, just now with infinite, learned worlds.
Previously, infinite AI-generated images relied on auto-regression: extending an image one tile at a time. Quality is good, but it has some practical issues: - Teleport 100 tiles away, and you have to generate all 100 tiles. - Output depends on sampling order, so it's not reproducible. - The process is inherently sequential, so it doesn't parallelize. - The full history must be stored forever, so it doesn't scale infinitely.
InfiniteDiffusion resolves all four, and it's training-free.
The flagship application is Terrain Diffusion: the first learned procedural terrain generator.
It generates an infinite number of continents in every world, with continents spanning millions of square kilometers. Using a novel technique, Laplacian Denoising, every world spans Earth's full elevation range, from -11,000m (Mariana Trench) to 9,000m (Everest), while maintaining precision within fractions of a meter.
And it’s fast. On a single RTX 3090 Ti, Terrain Diffusion generates terrain 9x faster than a satellite can fly over it, and requires only 1.5GB of VRAM.
@xandurglar congratulations and thank you for your contributions to open source and open science 🤗 https://huggingface.co/collections/xandergos/terrain-diffusion
Terrain Diffusion's most popular implementation is as a Minecraft mod, which you can play with right now on any modern GPU.
If you want to play around with Terrain Diffusion in your own applications, the Python code is open-source, and includes an API you can use to experiment with it in your own games.
Traditional noise is hand-crafted, so you’re restricted to whatever structures you can come up with a formula for. InfiniteDiffusion removes that restriction - now you can generate anything you can find training data for.
What’s nice about InfiniteDiffusion is that because it has the same contract as procedural noise, you can swap out individual parts of your pipeline depending on your needs.
@dmarcos @yacineMTB The repo includes an implementation of EDM2 that you can experiment with. I built the entire thing on one GPU so you don’t need to make any changes. That being said, the model from the paper is built for quality. You can comfortably get away with making it significantly smaller.
@xandurglar This looks cool! Excuse my ignorance, but what is the benefit over traditional noise/biom based generation?
@fabianfranz @yacineMTB Nope. Just YOLO. I’ve never been in any academic circles.
@xandurglar @janusch_patas I do have the feeling that you won’t have to work for Walmart much longer. 😃
@xandurglar @yacineMTB Congratulations! How fitting - I am about to publish my infinite context paper.
How did you get (back) into the acdemic circles?
Did you just submit your paper or did you have academic contacts proofread it before?
@huggingface @xandurglar Terrain generation was one of the last creative domains holding out against AI. Open source models just broke that barrier. The ripple effects for game dev and simulation will be massive.
@xandurglar This is fascinating, I had no idea there were folks approaching terrain generation this way. The fact that it's deterministic and stateless checks all the necessary boxes to be extremely functional for a game.
@xandurglar @ronjart Is it only 2d generation with heightmap or complete 3d generation ?
@DFinsterwalder @xandurglar @janusch_patas I’m actually depressed that he works at Walmart after getting his degree lol. Walmart exists off of leeching from Medicaid/SNAP with shit wages. Not the place a graduate deserves
@xandurglar @fabianfranz @yacineMTB to publish on arXiv do you need a referral?
