WorldFoundry
English中文

CUT3R

Method overview: Continuous 3D Perception. Given a stream of RGB images as input, our approach enables dense 3D reconstruction in an online, continuous manner, estimating both camera parameters and dense 3D geometry with each.

CUT3R from UC Berkeley: 3d-reconstruction and 3dgs; Runner parity pending.

Edit this pageGitHubMarkdown

CUT3R is a continuous 3D perception model with persistent state: it ingests a stream of images and continuously updates a metric-scale 3D reconstruction, without offline bundling. In this catalog the in-tree point-cloud and two-stage 3DGS routes are registered through Studio discovery. CUT3R also serves as the geometry backbone for the verified VMem entry in the video category.

Method overview: Continuous 3D Perception. Given a stream of RGB images as input, our approach enables dense 3D reconstruction in an online, continuous manner, estimating both camera parameters and dense 3D geometry with each.
Method overview: Continuous 3D Perception. Given a stream of RGB images as input, our approach enables dense 3D reconstruction in an online, continuous manner, estimating both camera parameters and dense 3D geometry with each.

The recorded paper is arXiv 2501.12387.

Project page: cut3r.github.io. Code: CUT3R/CUT3R.

Runner parity is pending. Routes are registered; checkpoint-backed inference waits on staged local or Hugging Face weights. Catalog environment / VRAM / input-contract rows are unset on this card — use worldfoundry-eval zoo model-show --model-id cut3r after staging.

Build your run

Pipeline
CUT3RPipelineworldfoundry.pipelines.cut3r.pipeline_cut3r
Task
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  cut3r \
  --pipeline.task-profile official-export \
  --json

What it is for

Use cut3r for catalog geometry inference. Choose it when the artifact is a reconstruction, not a generated video.

A recurrent transformer maintains a persistent scene state; each new frame updates the state and outputs pointmaps and camera parameters online (per the upstream publication). The recorded routes cover raw point-cloud export and a two-stage 3DGS pipeline; weights are apache-2.0. It is published by UC Berkeley.

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

Runtime contract

Optional input_path, size, vis_threshold, and revisit.

CLI task profiles: official-export and two-stage-3dgs. WorldFoundry binds CUT3RPipeline.

Family status: Runner parity pending.

Recorded official-export defaults: size=512 Recorded two-stage-3dgs defaults: size=512, width=704, height=480.

Runner parity is pending: checkpoint-backed inference depends on staged local or Hugging Face cut3r weights. Stage them before submitting Studio jobs.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run cut3r --pipeline.task-profile official-export. 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 liguang0115/cut3r. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh cut3r. Add --download to fetch missing public checkpoints.

Status: Runner parity pending.