facebookresearch/vggt-omega

[CVPR 2026 Oral] VGGT Omega

What it solves

VGGT-Ω provides a way to estimate camera poses and depth from a set of images, allowing for the reconstruction of 3D scenes from 2D visual data.

How it works

The project uses a pretrained transformer-based model (available in 1B parameter versions) that takes images as input and predicts pose encodings, depth maps, and depth confidence. These predictions are then converted into camera extrinsics and intrinsics to generate a 3D point cloud and visualize the scene as a GLB file.

Who it’s for

It is designed for researchers and developers working in computer vision, 3D reconstruction, and spatial AI.

Highlights

  • End-to-end inference: Predicts both camera parameters and depth from images.
  • Text-alignment support: Offers a specific checkpoint for text-aligned embeddings.
  • 3D Visualization: Includes a Gradio demo that visualizes depth-unprojected point clouds and predicted cameras as a GLB scene.
  • Scalable memory usage: Provides detailed GPU memory benchmarks for processing varying numbers of input frames.

Related

  • Project
    MIT-SPARK/VGGT-SLAM

    A real-time dense feed-forward scene reconstruction system for 3D mapping and localization, featuring optional open-set 3D object detection.

  • Project
    microsoft/MoGe

    MoGe is a model for recovering accurate 3D geometry, including metric depth and point maps, from single open-domain images.

  • Project
    facebookresearch/mvdust3r

    A single-stage 3D scene reconstruction tool that creates 3D point clouds and camera poses from sparse RGB views in about 2 seconds without requiring pre-known camera poses.

  • Project
    ByteDance-Seed/Depth-Anything-3

    A model that predicts spatially consistent 3D geometry, depth maps, and camera poses from single or multiple images using a unified depth-ray representation.

  • Project
    facebookresearch/VLM3

    VLM³ is a framework that enables standard Vision Language Models to perform 3D vision tasks like depth estimation and camera pose estimation using only text-based outputs and data scaling.