This might be a vague question, but what kind of intuition or knowledge do you need to work with these kind of things, say if you want to make your own model? Is it just having experience with image generation and trying to incorporate relevant inputs that you would expect in a 3D world, like the control information you added for instance?
I think this is very interesting because you seem to have reinvented NeRF, if I’m understanding it correctly. I only did one pass through but it looks at first glance like a different approach entirely.
More interesting is that you made an easy to use environment authoring tool that (I haven’t tried it yet) seems really slick.
Both of those are impressive alone but together that’s very exciting.
I love this! Your results seem comparable to the counter strike or minecraft models from a bit ago with massively less compute and data. It's particularly cool that it uses real world data. I've been wanting to do something like this for a while, like capturing a large dataset while backpacking in the cascades :)
I didn't see it in an obvious place on your github, do you have any plans to open source the training code?
Yup, definitely similar! There are a lot of video-game-emulation World Models floating around now, https://worldarcade.gg had a list. In the self-driving & robotics literature there have also been many WMs created for policy training and evaluation. I don't remember a prior WM built on first-person cell-phone video, but it's a simple enough concept that someone has probably done it for a student project or something :)
I also thought those wooden guard rails looked pretty spot on how they would look on 2C-B. The only thing that's missing is the overlay of geometric patterns on even surfaces.
Appreciate this article that shows some failures on the way to a great result. Too many times, people only show how the polished end-result: look, I trained this AI and it produces these great results. The world dissolving was very interesting to see, even if I'm not sure I understand how it got fixed.
Thanks! My favorite failure mode (not mentioned in the post - I think it was during the first round of upgrades?) was a "dry" form of soupification where the texture detail didn't fully disappear https://imgur.com/c7gVRG0
Is this a solo/personal project? If it is is indeed very cool.
Is OP the blog’s author? Because in the post the author said that the purpose of the project is to show why NN are truly special and I wanted a more articulate view of why he/she thinks that?
Good work anyway!
Yes! This was a solo project done in my free time :) to learn about WMs and get more practice training GANs.
The special aspect of NNs (in the context of simulating worlds) is that NNs can mimic entire worlds from videos alone, without access to the source code (in the case of pokemon) or even without the source code having existed (as is the case for the real-world forest trail mimicked in this post). They mimic the entire interactive behavior of the world, not just the geometry (note e.g. the not-programmed-in autoexposure that appears when you look at the sky).
Although the neural world in the post is a toy project, and quite far from generating photorealistic frames with "trees that bend in the wind, lilypads that bob in the rain, birds that sing to each other", I think getting better results is mostly a matter of scale. See e.g. the GAIA-2 results (https://wayve.ai/wp-content/uploads/2025/03/generalisation_0..., https://wayve.ai/wp-content/uploads/2025/03/unsafe_ego_01_le...) for an example of what WMs can do without the realtime-rendering-in-a-browser constraints :)
Amazing project. This has the same feel as Karpathy’s classic “The Unreasonable Effectiveness of Recurrent Neural Networks” blog post. I think in 10 years’ time we will look back and say “wow, this is how it started.”
Mostly 1xA10 (though I switched to 1xGH200 briefly at the end, lambda has a sale going). The network used in the post is very tiny, but I had to train a really long time w/ large batch to get somewhat-stable results.
Really cool. How much compute did you require to successfully train these models? Is it in the ballpark of something you could do with a single gaming GPU? Or did you spin up something fancier?
edit: I see now that you mention a pricepoint of 100 GPU-hours/roughly 100$. My mistake.