✦
Summary
GAE is a geometry-native model for generating appearance and 3D structure from a shared latent representation. It compresses multi-level geometry features into a structured state and performs flow matching directly in that state, decoding both visual and geometric outputs.
About GAE Geometry-Native Autoencoder
It gives world-generation developers an alternative to reconstructing geometry after generating appearance-only video. Core capabilities include Geometry-native latent space, RGB decoding, Depth output, and Camera-pose output.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include 3D geometry reconstruction, Structured 128-channel state, Flow-matching generation, and Public implementation.