Yep. People have been doing this kind of stuff for computer games for decades. It's actually not that difficult. It's not clear what novel problem is being solved here.
Yeah but those traditional procgen techniques don't use AI, and this one does use AI. They solved the problem of them not being AI enough for the AI era. AI!
In Mario 64 there is a staircase you can run up forever, granted it looks the same no matter how long you have Mario run up the stairs, but that certainly fits "big but uninteresting 3d world."
I know 'interesting' is subjective, but your comment is demonstrably false. Just type "mario 64 staircase" into youtube, and look at the hundreds (thousands? millions?) of videos and many millions of views.
Consider the patterns generated by cellular automata.
Both tend to stay interesting in the small scale but lose it to boring chaos in the large.
For this reason I think the better approach is to start with a simple level-scale form and then refine it into smaller parts, and then to refine those parts and so on.
I wonder if they also have a strategy for deleting generate tiles, otherwise the infinite is limited to the size of available memory. I also wonder if with their method can exactly recreate tiles that have been deleted. Or in other words, that they have a method for generating unique seeds for all tiles. The paper does not give much technical details. If the seed has a limited size and there is a method for generating seeds for each 2D coordinate, I wonder if it is possible to make a non-repeating infinite world. I think it is not possible with a limited size seed.
This could be a great way to make backrooms horror environments!
I've dreamed of a NeRF-powered backrooms walking simulator for quite a while now. This approach is "worse" because the mesh seems explicit rather than just the world becoming what you look at, but that's arguably better for real-world use cases of course.
> The code is being prepared for public release; pretrained weights and full training/inference pipelines are planned.
Any ideas of how it would different and better compared to "traditional" PCG? Seems like it'd give you more resource consumption, worse results and less control, neither of which seem like a benefit.
The description in the linked YouTube video for some reason has more info than the github repo:
> We tackle the challenge of generating the infinitely extendable 3D world — large, continuous environments with coherent geometry and realistic appearance. Existing methods face key challenges: 2D-lifting approaches suffer from geometric and appearance inconsistencies across views, 3D implicit representations are hard to scale up, and current 3D foundation models are mostly object-centric, limiting their applicability to scene-level generation. Our key insight is leveraging strong generation priors from pre-trained 3D models for structured scene block generation. To this end, we propose WorldGrow, a hierarchical framework for unbounded 3D scene synthesis. Our method features three core components: (1) a data curation pipeline that extracts high-quality scene blocks for training, making the 3D structured latent representations suitable for scene generation; (2) a 3D block inpainting mechanism that enables context-aware scene extension; and (3) a coarse-to-fine generation strategy that ensures both global layout plausibility and local geometric/textural fidelity. Evaluated on the large-scale 3D-FRONT dataset, WorldGrow achieves SOTA performance in geometry reconstruction, while uniquely supporting infinite scene generation with photorealistic and structurally consistent outputs. These results highlight its capability for constructing large-scale virtual environments and potential for building future world models.
Is it just me, or some of the places it generates are just not realistic? Like a small area of some kind which is a dead space, and there is a giant window into it.
I don't think generating virtual space is the issue.
It's about generating interesting virtual space!
You reminded me of this https://book.leveldesignbook.com/process/layout
And Valve I think used to have a series on level design, involving from big to small and "anchor points", but I seem to have misplaced the link.
Yep. People have been doing this kind of stuff for computer games for decades. It's actually not that difficult. It's not clear what novel problem is being solved here.
Yeah but those traditional procgen techniques don't use AI, and this one does use AI. They solved the problem of them not being AI enough for the AI era. AI!
Do you have some particular piece of software or tech demo or game in mind with interesting very large generated 3D worlds?
In Mario 64 there is a staircase you can run up forever, granted it looks the same no matter how long you have Mario run up the stairs, but that certainly fits "big but uninteresting 3d world."
> big but uninteresting 3d world.
I know 'interesting' is subjective, but your comment is demonstrably false. Just type "mario 64 staircase" into youtube, and look at the hundreds (thousands? millions?) of videos and many millions of views.
Valheim and No Man's Sky are ones I've played recently.
Minecraft surely fits those criteria.
” The generated scenes are walkable and suitable for navigation/planning evaluation.”
Maybe the idea is to create environments for AI robotics traini ng.
Consider the levels generated in any roguelike.
Consider the patterns generated by cellular automata.
Both tend to stay interesting in the small scale but lose it to boring chaos in the large.
For this reason I think the better approach is to start with a simple level-scale form and then refine it into smaller parts, and then to refine those parts and so on.
(Vs plugging away at tunnel-building like a mole)
Or at least coherent.
I wonder if they also have a strategy for deleting generate tiles, otherwise the infinite is limited to the size of available memory. I also wonder if with their method can exactly recreate tiles that have been deleted. Or in other words, that they have a method for generating unique seeds for all tiles. The paper does not give much technical details. If the seed has a limited size and there is a method for generating seeds for each 2D coordinate, I wonder if it is possible to make a non-repeating infinite world. I think it is not possible with a limited size seed.
This could be a great way to make backrooms horror environments!
I've dreamed of a NeRF-powered backrooms walking simulator for quite a while now. This approach is "worse" because the mesh seems explicit rather than just the world becoming what you look at, but that's arguably better for real-world use cases of course.
i'm thinking a new version of LSD dream emulator could be really interesting.
> backrooms horror environments
True, it sounds (and looks) a lot like https://scp-wiki.wikidot.com/scp-3008
This is cool. And could be fun in games. Not sure I get the point otherwise... The thought that came to mind was "Architectural slop".
It is only a paper as of now:
> The code is being prepared for public release; pretrained weights and full training/inference pipelines are planned.
Any ideas of how it would different and better compared to "traditional" PCG? Seems like it'd give you more resource consumption, worse results and less control, neither of which seem like a benefit.
The description in the linked YouTube video for some reason has more info than the github repo:
> We tackle the challenge of generating the infinitely extendable 3D world — large, continuous environments with coherent geometry and realistic appearance. Existing methods face key challenges: 2D-lifting approaches suffer from geometric and appearance inconsistencies across views, 3D implicit representations are hard to scale up, and current 3D foundation models are mostly object-centric, limiting their applicability to scene-level generation. Our key insight is leveraging strong generation priors from pre-trained 3D models for structured scene block generation. To this end, we propose WorldGrow, a hierarchical framework for unbounded 3D scene synthesis. Our method features three core components: (1) a data curation pipeline that extracts high-quality scene blocks for training, making the 3D structured latent representations suitable for scene generation; (2) a 3D block inpainting mechanism that enables context-aware scene extension; and (3) a coarse-to-fine generation strategy that ensures both global layout plausibility and local geometric/textural fidelity. Evaluated on the large-scale 3D-FRONT dataset, WorldGrow achieves SOTA performance in geometry reconstruction, while uniquely supporting infinite scene generation with photorealistic and structurally consistent outputs. These results highlight its capability for constructing large-scale virtual environments and potential for building future world models.
cant wait for the new diablo :)
With a quarter the size of the development team, 'cause productivity!
It looks more like the Stanley parable.
Is it just me, or some of the places it generates are just not realistic? Like a small area of some kind which is a dead space, and there is a giant window into it.
Not only not realistic but also not explicit: not so much as a peachy bottom in sight.