wtf is vggt-omega?

VGGT-Omega is a research project from Meta AI and the University of Oxford's Visual Geometry Group, accepted as an Oral paper at CVPR 2026. It is an AI model that looks at a set of photos, or a video, of a scene and figures out where each camera was positioned when the picture was taken, along with the depth of every point in the images. In plain terms, it turns a handful of ordinary photos into an understanding of 3D space, including camera angles, distances, and a point cloud you can view as a 3D scene. The project ships pretrained model checkpoints, though you need to request access to them on Hugging Face before downloading, and that request goes through an automated approval process rather than a person. Two versions are offered: one that works at 512 pixel resolution, and a smaller 256 pixel version that also aligns its output with text descriptions. To use it, you clone the repository, install the Python dependencies, and load a checkpoint. A short code example shows how to feed in image file paths and get back predictions for depth, camera position, and other internal representations. There is also an interactive Gradio demo you can launch locally, which lets you upload images or a video and see the resulting 3D point cloud and camera positions rendered as a downloadable 3D scene file. The README includes a detailed table of GPU memory requirements, since processing more input frames uses substantially more memory. Running the largest configurations needs a high end GPU such as an NVIDIA A100 with dozens of gigabytes of memory. The README does not spell out the license terms directly, pointing instead to a separate LICENSE file, and it notes the release is meant to support the open source research community rather than serve as a production ready tool. The authors provide an academic citation for anyone referencing this work in a paper.