WorldFoundry
English中文

MonST3R

MonST3R from Google DeepMind / UC Berkeley: geometry estimation in the presence of motion; Verified.

Edit this pageGitHubMarkdown

MonST3R extends the DUSt3R pointmap approach to dynamic scenes: it estimates per-timestep geometry directly from video, recovering shape and motion without a static-scene assumption. In this catalog it is fully verified — the CUDA Studio job produced a valid GLB v2 scene with 51 geometries and finite point data on 2026-07-10 through the in-tree demo.py subprocess adapter.

Official paper teaser: Watch the video.
Official paper teaser: Watch the video.

Project page: monst3r-project.github.io. Code: Junyi42/monst3r.

The paper is Paper. Project page: monst3r-project.github.io. Code: Junyi42/monst3r.

Build your run

Pipeline
MonST3RPipelineworldfoundry.pipelines.three_d_four_d.pipeline_runtime
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  monst3r \
  --pipeline.task-profile default \
  --json

What it is for

Pick monst3r for feed-forward geometry: views in, reconstruction / trajectory / point cloud out. Use it when you need the catalog 3D route rather than a depth-only prior.

Per the upstream publication, MonST3R fine-tunes the DUSt3R pointmap regressor on dynamic data so that pointmaps are estimated per timestep; downstream optimization aligns them into a 4D scene. WorldFoundry vendors the infer-time components in-tree and materializes a GLB scene artifact. It is published by Google DeepMind / UC Berkeley.

Catalog tasks include geometry estimation in the presence of motion.

Dynamic scene reconstruction. Sibling of CUT3R / DUSt3R-style cards.

Runtime contract

Optional input_path, frames, fps, weights, and model_name; artifacts generated_4d_scene and monst3r.glb.

CLI task profiles: default. WorldFoundry binds MonST3RPipeline.

Family status: Verified.

The Studio default disables the optional flow loss so missing RAFT2/SAM2 assets do not block preview output. For full official dynamic-mask parity, configure the MonST3R, RAFT, and SAM2 assets in the runtime environment.

WorldFoundry vendors the MonST3R infer-time components in-tree.

Running in WorldFoundry

Launch worldfoundry-eval run monst3r --pipeline.task-profile default to write geometry. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

No public checkpoint id is recorded on this card; follow the usage notes for staging. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh monst3r. Add --download to fetch missing public checkpoints.

Status: Verified.