China open-sourced a model that reconstructs any scene in 3D from a regular video, in real-time. one camera. no LiDAR. 10,000+ frames without falling apart. just walk around with your camera and watch the entire world get rebuilt in 3D at 20 fps. → runs at ~20 FPS on... a single GPU → Stable over 10,000+ frames → Beats optimization-based methods on benchmarks → Works on drone footage, driving videos, indoor walkthroughs 100% open source.show more

Now imagine what a spying agency have if now we see this publicly

I feel like you could implement a real life brain dance (cyberpunk 2077) from this

Putting aside the military applications of this, for making video games or even sets for movies this is going to be ridiculous. Something that would've otherwise needed to be hand built from scratch now can be done pretty easily with footage. AI is moving things forward.

We had several platforms doing this 4 years ago Polycam and MidiPort, 3Dsnap This is not novel

Wow, I should record my whole life and unlock as much of the map as possible

这玩意怎么不早点开发出来啊啊啊啊，这样我就可以把童年里的那些巷子房子记录下来了

I can imagine the real estate implications of this are huge

Is this using gaussian splats ? And does the AI also generate accurate hitboxes? So far it seems pretty early stage but I hope iteration will continue... Something like this would be interesting to see at large scale.

And you guys just now finding about it - Niantic been done ts already as well as much more. Look into their research.

China is prepared for AI. But why are they giving all of it for free??

Hey @grok can you look through that open source code for any backdoors or anything else hidden in it.

If this holds up outside the demo reel, the interesting part isn't the reconstruction itself, it's what happens to game asset pipelines once a phone video becomes a usable 3D environment. That collapses a step that used to take a modeling team weeks.

Didn't we have AR kit like 10 years ago?

Love that it's open source. Can't wait to see what the community builds with this.

Alright team, feed it the Scientology building speed runs

This would be excellent for construction, imagine a homeowner being able to quickly map their house and visualize remodels. You could see locations or core utilities when making changes.

@grok is this gaussian splatting?

This is like saying look china finally got to Tesla level

With this available it is ever so important to check, if someone is filming around system relevant buildings, banks and co. With this you could just go around with a cam and in a few minutes you have a complete 3d modell of something. Impressive and scary af at the same time.

Real-time 3D from single video, no LiDAR, 20fps on one GPU. Insane. Repo link?

Now imagine a camera mounted on let's say... Vaccum cleaners, drones, cars... connected to chinese servers... You just mapped in 3D the western world.

Gunner make home insurance quotes and claims much easier to manage without an agent or loss adjuster 🤔… just saying… $LMND

this is a fucking cyberpunk braindance and you cant tell me otherwise

I’m watching these vision models closely. Turning normal camera input into something useful still feels like one of the biggest shifts coming.

This is a cherry-picked research demo. The second you try running this open-source pipeline on regular video footage with fast motion blur or complex reflections, the tracking will break and your 3D mesh will look like a distorted glitch.

Now you see how Teslas work

If you watch Corridor Crew - you know there is nothing new here. I don't remember how it's called, but it exists for quete few years now, recreates objects based on light and shadows. It's limitation is highly reflective objects like mirrors.

Is that using some kind of skew Gaussian splatting? Crazy performance 😳

looks like it still needs some work

@elonmusk had been doing for a while in Tesla's.

Amazing technology, reconstructing 3D scenes from video. How does it handle complex environments?

wait so this basically makes traditional 3D scanning obsolete?

If the given tables are saying the truth and this does work as efficiently woa boy we are in for a ride