AVM builds consistent 3D scenes with no 3D at all, remembering views by camera position

jonstephens85 · x · 2026-11-20

The author calls AVM the most interesting release this month: it builds a 3D scene with no explicit 3D components, instead generating a stream of views and remembering past ones keyed by camera position. Walk away and come back, and the room still matches. No code released yet.

Related event: Meta Unveils AVM, a World Model Generating Walkable 3D Scenes Without 3D Data(4 posts)→

Original post →

More from Multimodal

Multimodal channel →