VGGT

VGGT (Visual Geometry Grounded Transformer) is an open-source 3D computer vision model from Meta AI with more than 14,500 stars on GitHub — winner of the CVPR 2025 Best Paper Award — that reconstructs 3D scenes directly from images in a single feed-forward pass.

Overview

Classic 3D reconstruction pipelines chain together feature matching, camera estimation, and iterative optimization, and can take minutes per scene. VGGT replaces that pipeline with one transformer: feed it one, a few, or even hundreds of images of a scene, and within seconds it directly infers everything needed to describe the scene in 3D — camera parameters, depth maps, point maps, and 3D point tracks. Because the model was trained at scale, its predictions are accurate enough that they can be used directly, and they also serve as excellent initialization for downstream optimization when maximum precision is required. The repository provides the model code, pretrained checkpoints, and examples for running inference and fine-tuning, making a state-of-the-art research result easy to experiment with. Released by Facebook Research (Meta AI), the project represents the current direction of the field: geometry learned by large models rather than assembled from hand-crafted stages.

Key Features

• Feed-forward 3D reconstruction – full scene geometry from images in seconds, no iterative optimization required.
• All-in-one outputs – camera parameters, depth maps, point maps, and 3D point tracks from a single model.
• Flexible input – works with a single image, a handful of views, or hundreds of photos of a scene.
• Award-winning research – CVPR 2025 Best Paper Award, from Meta AI’s research team.
• Pretrained checkpoints – downloadable weights and inference code to start experimenting immediately.
• Strong initialization – predictions can seed traditional optimization pipelines for even higher accuracy.
• Fine-tuning support – code included to adapt the model to your own data.

💳 Pricing

VGGT is free to download from GitHub, including the pretrained checkpoints. It is a research release from Meta AI, so check the license file in the repository for the exact terms of use before building on it commercially.

Who Is It For

Computer vision researchers, 3D graphics and robotics engineers, and developers building photogrammetry, AR, or scene-understanding applications who want modern learned reconstruction instead of a fragile classical pipeline.