GAE (Geometric AutoEncoder) is a research codebase from ARC Lab at Tencent IEG, with the Hong Kong University of Science and Technology, the University of Hong Kong and the University of Texas at Austin, accompanying the paper "Learning a Geometry-Native Latent Space for 3D-Consistent World Generation" (arXiv 2609.24981). Instead of encoding frames as images and adding geometry from outside the latent, it compresses frozen Depth Anything 3 features into a compact per-view latent that decodes jointly into RGB and geometry (depth, rays and point maps through the frozen DPT head). Stage 1 trains the codec, with a 64- or 128-channel latent, using token alignment to C-RADIOv2.5-B and a DINOv2-L structure loss. Stage 2 trains one conditional flow model over the frozen latent, controlled by clean reference latents, metric Plücker ray maps for camera motion, and text. The released weights, TencentARC/GAE-D64-1B on Hugging Face, turn one image and a prompt into a multi-view video with a consistent 3D point cloud, or a prompt into a single image, and the codec can reconstruct RGB, depth and point clouds from input video. The repository ships a Python API (GAE.from_pretrained), demo scripts, a Gradio Space app, a multi-GPU Ulysses image-to-video server that expects at least eight CUDA GPUs, and training, data preparation and evaluation scripts. The README says the released flow weights are a stronger continued-training model than the one used for the paper's generation tables, so those tables will not be reproduced exactly, and the original research checkpoint is not part of the release. It gives no benchmark numbers of its own and refers readers to the paper. The license is a custom Tencent permission notice that restricts use to academic purposes and prohibits non-academic, commercial or production use.