WorldFoundry
English中文

Infinite VGGT

Method overview: Framework of StreamVGGT. Our model consists of three main components: an image encoder, a spatio-temporal decoder, and multi-task prediction heads. During training, we utilize full-sequence inputs to provide the.

Infinite VGGT from Tsinghua University: 3d-reconstruction and streaming-3d; Runner parity pending; Streaming 4D Visual Geometry Transformer.

Edit this pageGitHubMarkdown

Infinite VGGT (StreamVGGT) turns the feed-forward VGGT reconstructor into a streaming model: it processes incoming frames causally with cached tokens, so reconstruction extends indefinitely instead of being bounded by a fixed window. The in-tree pipeline is registered for streaming 3D reconstruction.

Method overview: Framework of StreamVGGT. Our model consists of three main components: an image encoder, a spatio-temporal decoder, and multi-task prediction heads. During training, we utilize full-sequence inputs to provide the.
Method overview: Framework of StreamVGGT. Our model consists of three main components: an image encoder, a spatio-temporal decoder, and multi-task prediction heads. During training, we utilize full-sequence inputs to provide the.

The recorded paper is Streaming 4D Visual Geometry Transformer (2025).

Project page: wzzheng.net/StreamVGGT. Code: wzzheng/StreamVGGT.

Build your run

Pipeline
InfiniteVGGTPipelineworldfoundry.pipelines.infinite_vggt.pipeline_infinite_vggt
TaskInteractive Video interactive-video
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  infinite-vggt \
  --pipeline.task-profile interactive-video \
  --json

What it is for

infinite-vggt is the recorded reconstruction line. Start here to emit geometry from the documented views.

A causal transformer with cached-token knowledge distillation from VGGT (per arXiv 2507.11539). License is CC-BY-NC-4.0 — non-commercial. It is published by Tsinghua University.

Catalog tasks include 3d-reconstruction and streaming-3d. Catalog aliases include stream-vggt and streamvggt.

Runtime contract

Optional input_path, interactions, and output_format.

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

Family status: Runner parity pending.

Use the vggt entry for offline reconstruction and this one when frames arrive as a stream.

In-tree StreamVGGT (Infinite VGGT) pipeline is registered for streaming 3D reconstruction.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run infinite-vggt --pipeline.task-profile interactive-video. 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 lch01/StreamVGGT. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh infinite-vggt. Add --download to fetch missing public checkpoints.

Status: Runner parity pending.