Luma Is Doing the Sora Thing, Except They Actually Know What 3D Is

Luma is doing the Sora NeRF trick.

If you watched the Sora technical report drop in February and got excited about the part where the camera orbits a fixed scene and the geometry stays consistent — that was the bit. The one that made people say "it has a world model" and "it understands 3D space" and other things that are probably slightly overclaiming but not entirely wrong. The implicit-NeRF-without-calling-it-a-NeRF move.

Luma is doing that now, except Luma is a company that actually built a NeRF product, ships a NeRF iPhone app, and has been thinking about neural radiance fields since before most people could spell it. So when they do the trick, it's not an emergent property of a video model that ate the internet — it's people who know exactly what a radiance field is, deliberately building a pipeline around it.

The demo comes from their graphics engineer, who says out loud that this is a pipeline they're actively building. Not a research artifact. Not a one-off. A thing they intend to ship.

I'd been sitting on the prediction that someone would eventually close this loop — take a video diffusion model and wire it properly into a 3D-native workflow rather than treating video generation as an endpoint — but I hadn't seen anyone actually try it publicly until now.

The interesting question isn't whether this works. It clearly works. The question is what happens to the whole cottage industry of NeRF pipelines when the video model underneath them gets good enough that the NeRF step becomes vestigial. Luma is either building the bridge to that future or building the thing that makes the bridge unnecessary. Maybe both. Probably both.