Research autoencoder with a geometry-native latent space for 3D-consistent video and world generation.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
GAE (Geometry-Native Autoencoder) is a research project from HKUST and Tencent ARC Lab (IEG), with additional authors from HKU and UT Austin. It distills multi-level features from a frozen geometry foundation model (DA3) into a compact latent, compressing 3,072-channel features to a 128-channel state. Flow matching is then run directly in this latent space, and the same generated latent can be decoded into RGB, depth, camera poses and point-cloud geometry.
The project reports controlled comparisons in which only the latent is swapped while the generator and training protocol stay fixed. Against baselines such as the Stable Diffusion VAE, the Wan2.1 video VAE, RAEv2 and raw DA3 features, GAE lowers FVD by 12.7% on RealEstate10K and 23.1% on DL3DV, and halves camera-trajectory error on RealEstate10K. Two variants are described, GAE-64 for compact generation and GAE-128 for higher-quality reconstruction. The demo generates 81-view 672x378 sequences from one observed frame and a metric camera path, and also supports text-to-image with depth and point clouds.
It is a roughly 1B-parameter research release (2026) with a project page, paper and code links. A YouTube creator reported that weights under 12 GB plus training and evaluation scripts were released, but the provided pages give no repository, Hugging Face or license details, so these are unverified here.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Generates consistent 3D worlds with video, depth map, camera trajectory and 3D reconstruction simultaneously. Weights under 12 GB plus training and eval scripts released.
Closest neighbours by meaning, not just by category.
Model · Tencent ARC Lab
Video world model from Tencent ARC Lab that turns an image or text into camera-controlled explorable scenes with 3D-aware memory.
Research
Research framework for globally consistent 3D reconstruction from long RGB videos using chunked priors and test-time adaptation.
Model
14B autoregressive world model that generates interactive, camera-controlled video worlds in 3 denoising steps.
Model · World Labs
World Labs' world model that generates and reconstructs explorable 3D worlds from images, text, video or 3D input.
Model
Open framework and dataset for turning video diffusion models into real-time, long-horizon interactive world models.
Research
Research method for long-context video generation using hierarchical latents and coarse-to-fine rollout.
