GAE Paper: Geometry-Native Latent Space Cuts FVD by 23% and Halves Camera Error

yshan2u · x · 2026-09-23

The GAE paper argues 3D-inconsistent generation is a representation problem: it reparameterizes geometry foundation model features into a compact latent space, jointly decodable to appearance, depth, cameras and point maps, then generates there with conditional flow. Controlled comparisons show FVD drops 12.7%/23.1% on RealEstate10K/DL3DV and camera-trajectory error halves. Code is open-sourced.

Related event: GAE: A Geometry-Native Latent Space for 3D-Consistent Generation(4 posts)→

Original post →

More from Research

Research channel →