WonderWorld: Interactive 3D Scene Generation from a Single Image | Hacker News Reader

WonderWorld: Interactive 3D Scene Generation from a Single Image

kovenyu.com

188 pointslnyan18 comments

If you click on the image of "Link" (I know he is not really) in the "Interactive Viewing" section then you can see that in front of him (out of view) is a bunch of noise. I think it is interesting that it would predict randomness above just predicting nothing being there.

This is awesome tech.

Super impressive, and I can see it being useful in many cases already. Especially making interactive experiences in combination with position tracker of a user in a room. As you move around the room your perspective changes.

In a more creative approach I could imagine creating fake windows using flat-screen TVs in this approach as well. As you move around the room the perspectives would change as well, giving an illusion of the windows being real. Of course this would only work for a single person at a time but it would be quite interesting to experience. It should not be too difficult to hack it together as a solo dev.

This is like 1997's Blade Runner game camera (and from the movie too):

https://youtu.be/DRx2Leb2yDE?t=1680

Does anyone know if there are variants of this that output voxels? It feels like a more concrete representation of the space versus Gaussian splats.

It isn’t released yet, but the thing akin to this that Roblox is working on (to be open-weights) most likely is voxel-based.

Source on Roblox working on something like this?

"This internal AI project will power generative creation on our platform. Our 3D foundational model will be open source and multimodal, and it will power 3D generation through text, video, and 3D prompts. We see a powerful future where Roblox experiences will have extensive generative AI capabilities to power real-time creation integrated with gameplay. We’ll provide these capabilities in a resource-efficient way, so we can make them available to everyone on the platform."

https://corp.roblox.com/newsroom/2024/09/rdc-2024-robloxs-ne...

This is AMAZING!

I hope this is released for public use at some point. I'd love to run it through some of my older photos to see what it does with them.

It feels like "3D" is a stretch given the approach they're using. Obviously the result is pretty cool, but I suspect anything built using this tech is going to have a very distinct feel (almost like sprite based video games).

This is incredible. You could build entire games this way.

Would they be any good though, that remains to be seen.

Imagine Google street view data put to use in combination with this. You would essentially have an open world game of any city on earth.

Yeah, I was thinkin' exactly the same thing. Even if not for games (although that would be nifty) just imagine the additional "depth" and "sense of presence" this would bring to Google's Street View. Street View was already pretty slick, but pile things like this on top of what they've already got there? Just ... Wow!

Or, you could confuse the heck out of GeoGuessr players :)

can wait for the code /api

Very cool!

This is the future I was promised. Take my money please.

How very supercalifragilisticexpialidocious!

Reply on news.ycombinator.com