What's this about?
GAE compresses a scene into a representation that jointly reconstructs appearance, depth, cameras and point maps before a standard flow model generates from it.
What is a geometry-native latent?
It is a compact hidden representation organised around 3D structure rather than only visual appearance, helping different viewpoints refer to the same world.
Why does this matter?
author-reported preprint; 38 HF upvotes.
- Latent choice alone improves visual and geometric measures in controlled comparisons.
- Camera-trajectory error is reportedly halved on RealEstate10K.
- Benchmark consistency does not yet prove stable, interactive world physics.