GAE, a geometry-native autoencoder, reparameterizes features from a geometry foundation model into a compact latent space jointly decodable to appearance, depth, camera pose, and point maps, intended as a shared representation for both perception and generation. Swapping this latent into a fixed video generator and training protocol improves 3D coherence and visual quality, cutting FVD by 12.7% and 23.1% on RealEstate10K and DL3DV respectively and halving camera-trajectory error on RealEstate10K. The results argue latent-space design, not just generator architecture, is central to 3D-consistent world generation.
