VGGT

VGGT took the CVPR 2025 Best Paper Award for showing that one feed-forward transformer can directly infer every key 3D attribute of a scene: extrinsic and intrinsic camera parameters, depth maps, dense point maps, and 3D point tracks, from anywhere between one and hundreds of input views in seconds, with no iterative optimization. A joint effort between Meta AI and Oxford's Visual Geometry Group, the 1B parameter model uses specialized prediction heads per output and can export COLMAP-format reconstructions for downstream pipelines. The code runs best with bfloat16 on Ampere-class NVIDIA GPUs, a May 2025 optimization cut memory use by roughly 2 to 3x, training and finetuning code landed in July 2025, and a VGGT-Omega follow-up arrived in May 2026. Licensing is split: the code ships under Meta's VGGT License, which allows commercial use with acceptable-use restrictions such as no military applications, while the original weights are non-commercial and a separate VGGT-1B-Commercial checkpoint is granted through an application form. The repository stands around 14.1k GitHub stars.