GAE-GeometricAutoEncoder: Video Generator — What It Does and What You Need to Run It
GAE-GeometricAutoEncoder is an open source video generator from ARC Lab, Tencent PCG. It solves the image-to-video task: from a single image it creates a video sequence and also outputs a 3D point cloud. The developers position the tool as a way to achieve 3D-consistent world generation through a geometry-native latent space.
What it does
The tool takes an image and generates video. As described by the author, three artifacts are produced on output: an MP4 video, a trajectory visualization, and a point cloud in ply format. This means that along with the flat video sequence you can get a spatial representation of the scene.
It is based on the GAE-D64-1B model, available on Hugging Face. The authors describe the approach as learning a geometry-native latent space for 3D-consistent world generation. The repository tags list 3D reconstruction, depth-anything-v3, VAE, video generation, world generation, and world model.
What you need to run it
The code is written in Python, the library is PyTorch. As described by the author, Python version 3.10—3.12 and torch 2.5.1 are required. The DA3-GIANT backbone is pulled from Hugging Face.
The largest model weights file takes up 3.5 GB, all model weights files together — 5.2 GB. This is worth keeping in mind when planning disk space and downloads.
The code license is the License Terms of Learning a Geometry-Native Latent Space. It permits free use, copying, modification, publication, distribution, sublicensing, and sale of copies of the software, but for academic purposes only. Any non-academic, commercial, or production use is prohibited under any circumstances. All copies or substantial portions of the software must include the copyright notice and the permission text.
The repository was created on September 16, 2026, the last code change was on September 24, 2026.
Who it suits
The tool is aimed at researchers and students working with video generation and 3D reconstruction in an academic setting. The ability to get a point cloud together with video makes it interesting for tasks where you need to analyze the spatial structure of a scene, not just a flat image.
It is not suitable for commercial development, production, or any non-academic projects: the license explicitly prohibits such use. It is also worth noting that the model weights take up more than five gigabytes — for weak machines and quick experiments this can be noticeable.
GAE-GeometricAutoEncoder is a research tool with a clearly defined scope. It combines video generation with obtaining a 3D point cloud and works on the image-to-video model. The academic license and the considerable size of the weights define its audience: it is more of a laboratory than a production tool.
Fact sheet
Repository · Model on HuggingFace
| Task | image-to-video source |
|---|---|
| Code license | License Terms of Learning a Geometry-Native Latent Space: academic use only source |
| Largest weights file | 3.5 GB source |
| All weights files | 5.2 GB source |
| Export formats | ply per the author’s description source |
| Python | 3.10 per the author’s description source |
| Library | pytorch source |
| Language | Python source |
| Last code change | 2026-09-24 source |
| Repository created | 2026-09-16 source |
| GitHub stars | 328 source |
|---|---|
| Forks | 10 source |
| Downloads per month | 0 source |
| Model updated | 2026-09-24 source |
Values are collected automatically from official sources and were checked on 2026-09-26. Each one links to its source, and values taken from the developer’s pages also carry a verbatim quote — hover over the note. Pricing and versions are shown as of the check date and change most often; verify on the vendor’s site before buying.
