Lumalabs AI | Hacker News Reader

Lumalabs AI

lumalabs.ai

69 pointsdownboots16 comments

The speed with which Luma transitioned their product from being NeRF-based to 3D-Gaussians-based following the release of the 3D Gaussian Splatting paper was really impressive in my opinion. It's one thing to implement a brand new research paper, it's another to implement it in a polished manner and ship it as a working product in just a few weeks!

I'm not super familiar, but I think Polycam and KIRI Engine also added Gaussian splatting support in a similar timeframe. Unfortunately, it doesn't look like Luma has an Android app yet.

Am I old or is technology moving too fast these days? Or maybe before I have no idea what is possible and what isn't so I accept incredible stuff as normal but now the "impossible" becomes trivial and my world view is challenged?

Anyway, incredible stuff. I remember the days where a 3D model took a week to make in low res. Maybe the singularity isn't that far away...

What's the current best open source version of this kind of thing? Awkward to get running and a UI that's a half working notebook is fine, but are there decent models/setups to go from video/images to point cloud and renderable scene (with Gaussian splatting).

I believe they are using the original Gaussian Splatting code from Inria under the hood: https://github.com/graphdeco-inria/gaussian-splatting

Aha thank you, I was also stuck with the stage of getting point clouds but there is a video tutorial linked in that page that covers everything from images to the end product.

This level of fidelity will make the metaverse a compelling product.

Saleem

Is it possible to 3D print an object from a 3D Gaussian based scan?

The short answer is no.

The long answer is also no, but maybe eventually yes.

This is like a raw point cloud from a 3D scanner, but arguably even worse. A raw point cloud can be transformed into a solid object representation with enough work, but it's often easier to just re-draw the object by hand, cross-referencing the scan to make measurements when needed. With enough AI magic, maybe someday it will be possible for this to automatically work reasonably well, but not for the foreseeable future.

You could run it through marching cubes and get a mesh, whether or not that mesh would be printable is a different question entirely.

It might actually be better to slice the point cloud directly. Go layer by layer and trace the boundary of the shape. I’d wager someone will make that work in the near future.

> whether or not that mesh would be printable is a different question entirely.

In point of fact, I believe that was the question :P

And that is the fundamental problem. The answer is “no, it’s not printable” in almost all cases without significant manual work.

At the moment, the geometry produced is not great. Everything looks great but the actual underlying geometry is messy - for now.

Can do that with iPhone with LiDAR

I’ve used iPhone lidar to 3D scan, the results are no where near this level of fidelity.

Not this well you cant. Lidar is super low res. Luma scans look way better than my lidar scans.

Reply on news.ycombinator.com