48 comments

[ 1.9 ms ] story [ 67.7 ms ] thread
First steps towards the Holodeck.
I'd call this 3DAssetGen. It's not a world model and doesn't generate a world at all. Standard sweat-and-blood powered world building puts this to shame, even low-effort world building with canned assets (see rpg maker games).
This is cool, but it seems much more like a 3d asset generation than the scene generation like World Labs.
My first thought was the comment in the thread from the other day about Zork and hooking up an AI image generator to that.

But, it looks like WorldGen has that slightly soulless art style they used for that Meta Zuckverse VR thing they tried for a while.

It’s a fun demo but they never go into buildings, the buildings all have similar size, the towns have similar layouts, there’s numerous visual inconsistencies, and the towns don’t really make sense. It generates stylistically similar boxes, puts them on a grid, and lets you wander the spaces between?

I know progress happens in incremental steps, but this seems like quite the baby step from other world gen demos unless I’m missing something.

Every environment appears to be a miniature golf course version of reality. Was this a deliberate choice?
Having the technical knowhow to have an ai generate 3d models, but then generatively compositing those assets together into environments in a way that would have seemed overly simplistic to gamedevs 3 decades ago…
It's funny, I don't know if I see a use for it, and this feeling surprises me. Just as procedural maps bore me, I feel this will be similar in any use case I can think of. What I like is the perceived care behind every action. After the initial "wow" of the care put into that research, I don't think it will end up being a "wow" that scales—I don't know if I'm making myself clear.
It's definitely a step forward from that 'Minecraft world' gen tech demo that had no persistence of vision.

I can see it being useful for isolated Unity developers with a concept and limited art ability. Currently they would be likely limited to pixel games.

They are not limited like that. You can get a game very far with those asset store assets
I loathe how meta.com makes my back button gray out in my browser. Stop trying to force me to stay, it is obnoxious.
The back button works fine for me (on Chrome). I can go back from the post to HN, and I can navigate to other pages on Meta and then go back.

What browser are you using? How is it even possible for a site to remove previous browser history in a tab?

It's funny, I clicked the link to the demo, but it 404s, then I tried googling Worldgen, and it turns out someone else has built the same thing in May and called it Worldgen as well. Looks like it does better at realistic 3D scenes compared to this.

[0] https://worldgen.github.io/index.html

It's weird, houses are almost all tall and too narrow.
I can see this working as a randomly generated map for some quick game, like the Worms games did in 2D.

But, having things feel strongly on a grid kind of ruins the feel. It's rare for every building to be isolated like that. I am guessing they had trouble producing neighboring buildings that looked like they could logically share a common wall or alleyway.

These look a lot like World of Warcraft. I wonder how much of their training data they got from it.
Can’t wait until entire triple A games are generated by a prompt. Hopefully in my lifetime.
I would simply spend $5 at an asset store for some blobby generic buildings, than orchestrating a 12-figure corporate debt bubble to build warehouses of rapidly depreciating rust that boils a lake in order to generate them, but I guess that's why I'm not a Business Genius.
so clearly soon I’ll be able to tell a story and watch it appear before my eyes
Does it in fact create a world that reflects the prompt? Probably not, except in a vague way.

Any world you can summon into existence with a few words is by the laws of information theory going to be generic. An interesting world requires thousands of words to describe.

Compare https://odyssey.ml/ another text conditioned world generator
Personally, I think diffusion model world gens like Odyssey are the future, and not polygon-gen tools like TFA.
This just seems like an engineered pipeline of existing GenAI to get a 3d procedurally generated world that doesn't even look SOTA. I'm really sorry to dunk on this for those that worked on it, but this doesn't look like progress to me. The current approach looks like a dead end.

An end-to-end _trained_ model that spits out a textured mesh of the same result would have been an innovation. The fact that they didn't do that suggests they're missing something fundamental for world model training.

The best thing I can say is that maybe they can use this to bootstrap a dataset for a future model.

The people who worked on it did what they could to satisfy the demands of their higher-up’s, who frequently are out of touch with the technical landscape.

Being kind to them and understanding the environment they work in won’t improve their lives, but it will expand our understanding of the capability of particular large companies to innovate.

What’s SOTA in this area right now?
Not sure what is going on but seems like meta is lacking behind other startups and other frontier models in this space. They invested most in Meta reality labs in the last decade more than any other company and they come up with such poor rendering while the competitors are making pretty cool real world demos. Meta should stop thinking of these as research projects and actually spend time building real products with proper 3d rendering.
it won't be long now till we see a vr star trek holodeck type thing.