Meta's VGGT Rebuilds 3D Scene Reconstruction in a Single 0.2-Second Pass
Meta and Oxford's 1B-parameter transformer reconstructs full 3D scenes from one to hundreds of images in under a second, no optimization required.
Takeaways
- Meta and Oxford released VGGT-1B, a 1B-parameter transformer that infers full 3D geometry in one forward pass
- Outputs camera intrinsics, extrinsics, depth maps, point maps, and 3D point tracks from 1 to hundreds of views in under a second
- Won CVPR 2025 Best Paper, beating DUSt3R, MASt3R, and matching VGGSfMv2 without optimization (paper)
- Architecture: 24 blocks alternating frame-wise and global attention, DINOv2 tokenizer, camera and register tokens per frame
- Commercial-licensed checkpoint VGGT-1B-Commercial available for production use, excluding military applications
- Spawned ecosystem of SLAM, streaming, and robotics variants built on the pretrained backbone (GitHub)
Meta AI and Oxford’s Visual Geometry Group have released VGGT, a 1 billion parameter transformer that predicts cameras and scene geometry from a single image, a small image set, or hundreds of views. The paper received a CVPR 2025 Best Paper Award, and the project repository includes inference code, pretrained weights, and a browser demo.
VGGT compresses feature matching, camera estimation, depth prediction, and point tracking into one feed-forward model. Developers can use its output directly for fast reconstruction or pass the predicted cameras into bundle adjustment when an application needs higher pose accuracy.
One pass replaces a chain of solvers
Classical multi-view reconstruction usually detects and matches image features, runs structure from motion to estimate cameras, refines those cameras with bundle adjustment, and applies multi-view stereo to recover dense geometry. Each stage introduces separate dependencies, parameters, failure modes, and optimization time.
VGGT maps RGB images to the main outputs expected from that stack:
| Output | What it represents | Common use |
|---|---|---|
| Camera intrinsics | Per-view parameters such as focal length and principal point | Projection, rendering, and calibration |
| Camera extrinsics | Each camera’s orientation and position in a shared coordinate frame | Structure from motion, AR, and neural rendering |
| Depth maps | Per-pixel depth estimates with confidence values | Geometry recovery and occlusion handling |
| Point maps | Per-pixel 3D coordinates expressed in a common frame | Point clouds, surface estimation, and scene alignment |
| Point tracks | Corresponding scene points across multiple images | Tracking, pose refinement, and temporal reconstruction |
Attention links the camera views
VGGT encodes each image as patch tokens using a pretrained vision backbone based on DINOv2. Its 24 transformer blocks pair frame-wise self-attention, which processes evidence within each image, with global self-attention, which exchanges information across the full image set.
The model adds one camera token and four register tokens to every frame. The first frame receives a distinct token set, while later frames share another set. This asymmetry anchors predictions to a canonical coordinate system based on the first view. Changing the first image can therefore change the reported coordinate frame, so applications comparing separate runs may need to align their outputs.