facebookresearch

vggt: Reconstruct 3D Scenes from Images

VGGT is a feed-forward vision model that estimates cameras, depth, point maps, and point tracks from one or more images. It suits researchers and developers who need fast 3D scene reconstruction without a traditional multi-stage pipeline.

View on GitHub Compare
Stars
14.5k
Forks
1.6k
Stars, last 30 days
+118

Our take

Health score

26/100

How it is scored
Activity
19
Community
44
Issues
21
Pull requests
20

+118 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

14,632

Forks

1,573

Change since Nov 14, 2026: +1 stars·0 forks

Overview

VGGT (Visual Geometry Grounded Transformer) infers several kinds of 3D scene information directly from images, including camera parameters, depth maps, point maps, and tracks. It addresses the need to combine multiple geometry-estimation steps by producing these predictions in one model, from a single view or a set of views.

The project is aimed at computer-vision researchers and developers building reconstruction workflows. Its outputs can also be exported to COLMAP format for use with tools such as Gaussian splatting pipelines.

Key Features

  • Predicts camera extrinsics and intrinsics from image views.
  • Produces depth maps, 3D point maps, and point tracks.
  • Supports single-view and multi-view reconstruction.
  • Provides Python APIs and demos for image or video inputs.
  • Exports predictions in COLMAP format, with optional bundle adjustment.
  • Includes training code and an example for fine-tuning on a custom dataset.
  • Can connect exported reconstructions to Gaussian splatting workflows such as gsplat.

Use Cases

  • Researchers can evaluate feed-forward 3D reconstruction methods on image collections.
  • Computer-vision developers can generate camera and depth estimates as inputs to downstream geometry workflows.
  • 3D practitioners can create COLMAP-format reconstructions from scene images for Gaussian splatting.
  • Users exploring a single photograph can try the model's zero-shot single-view reconstruction capability, while treating results as an exploratory estimate.

Getting Started

git clone https://github.com/facebookresearch/vggt.git
cd vggt
pip install -r requirements.txt

See the README for model loading, demos, exports, and package installation details.

Alternatives

  • LAM: LAM reconstructs and animates a single-image 3D Gaussian head avatar, while VGGT estimates general scene geometry and camera properties from one or more images.
ProjectLanguageLicenseStarsStatus
vggtPythonOther14.5kNot checked yet
LAMPythonApache-2.01.1kQuiet

Considerations

  • The pretrained model is large, and practical inference is best suited to a CUDA-capable GPU. The README's example includes a CPU fallback, but does not promise comparable speed.
  • Model predictions are estimates; reconstruction quality can vary with the input and downstream settings. Visualization may take substantially longer than inference.
  • The README distinguishes checkpoint terms: the original checkpoint is non-commercial, while the separate VGGT-1B-Commercial checkpoint is described as permitting commercial use subject to its terms. Check the repository license and the chosen checkpoint's terms before use.

Found this useful?

Share it with someone who would like vggt.

Comparisons