WorldFoundry
English中文

VGGT

VGGT from Meta AI (FAIR) / University of Oxford (VGG): 3d-reconstruction and 3dgs; Runner parity pending.

Edit this pageGitHubMarkdown

VGGT (Visual Geometry Grounded Transformer, CVPR 2025) is a feed-forward network that infers cameras, depth maps, point maps, and 3D point tracks from one to hundreds of views in a single pass — no per-scene optimization. In this catalog the in-tree two-stage 3DGS pipeline is registered, distinct from the vggt-omega and vggt-world entries.

Official paper teaser: VGGT reconstructs 3D from one to many views faster and cleaner than DUSt3R.
Official paper teaser: VGGT reconstructs 3D from one to many views faster and cleaner than DUSt3R.
Method overview: DINO tokens plus camera tokens go through alternating global and frame attention, then camera, depth, point-map, and track heads.
Method overview: DINO tokens plus camera tokens go through alternating global and frame attention, then camera, depth, point-map, and track heads.

Project page: vgg-t.github.io. Code: facebookresearch/vggt.

The paper is Paper. Project page: vgg-t.github.io. Code: facebookresearch/vggt.

Build your run

Pipeline
VGGTPipelineworldfoundry.pipelines.vggt.pipeline_vggt
TaskInteractive Video interactive-video
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  vggt \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Use vggt for catalog geometry inference. Choose it when the artifact is a reconstruction, not a generated video.

Per the CVPR 2025 paper, a large transformer alternates frame-wise and global attention over image tokens, with task-specific heads for cameras, depth, pointmaps, and tracking. Weights are apache-2.0 (facebookresearch/vggt); the recorded route builds 3DGS from VGGT geometry in two stages. It is published by Meta AI (FAIR) / University of Oxford (VGG).

Catalog tasks include 3d-reconstruction, 3dgs, and point-cloud. Documented outputs include point cloud.

Not VGGT-World and Infinite VGGT.

Runtime contract

Optional input_path, interactions, fps, height, and width.

CLI task profiles: interactive-video. WorldFoundry binds VGGTPipeline.

Family status: Runner parity pending.

Integration is Runner parity pending. For the verified scene-export route use vggt-omega; use this entry when you want the two-stage 3DGS pipeline over base VGGT geometry.

In-tree VGGT two-stage 3DGS pipeline is registered; distinct from vggt-omega and vggt-world catalog entries.

Running in WorldFoundry

Launch worldfoundry-eval run vggt --pipeline.task-profile interactive-video to write geometry. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

The environment profile is not recorded.

Recorded weight repositories include facebook/VGGT-1B. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh vggt. Add --download to fetch missing public checkpoints.

Status: Runner parity pending.