WorldFoundry
English中文

NeoVerse

NeoVerse from CreateAI (formerly TuSimple): 4D world modeling from in-the-wild monocular videos; Verified.

Edit this pageGitHubMarkdown

NeoVerse is a 4D world model built as a native assembly over canonical Wan components plus model-specific control and reconstruction roles: it enhances world generation with 4D-aware conditioning. In this catalog the runner is verified for the README-style tilt_up video demo with local NeoVerse assets; the official video-input path uses static_scene=false.

Method overview: Framework of NeoVerse. In the reconstruction part, we propose a pose-free feed-forward 4DGS reconstruction model ( Sec. 3.1 ) with bidirectional motion modeling. The degraded renderings in novel viewpoints from 4DGS.
Method overview: Framework of NeoVerse. In the reconstruction part, we propose a pose-free feed-forward 4DGS reconstruction model ( Sec. 3.1 ) with bidirectional motion modeling. The degraded renderings in novel viewpoints from 4DGS.

The recorded paper is NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos (CVPR 2026, 2026).

Project page: neoverse-4d.github.io. Code: IamCreateAI/NeoVerse.

Build your run

Pipeline
NeoVersePipelineworldfoundry.pipelines.neoverse.pipeline_neoverse
TaskOfficial Tilt Up Video official-tilt-up-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  neoverse \
  --pipeline.task-profile official-tilt-up-video \
  --json

What it is for

neoverse is the recorded interactive world line. Start here when the job is stepping a world with actions.

Per the upstream release (arXiv 2601.00393), NeoVerse layers 4D reconstruction and control roles over a Wan video backbone. WorldFoundry expresses this as a native assembly over shared Wan infrastructure — real inference stays checkpoint/GPU gated with no metadata-only runner shim. It is published by CreateAI (formerly TuSimple).

Catalog tasks include 4D world modeling from in-the-wild monocular videos. Catalog aliases include NeoVerse: Enhancing 4D World Model.

Recorded evaluation entry points include WorldScore (Scores navigable generated worlds), 4DWorldBench (Matches the entry's 4D world-model focus).

Runtime contract

Optional video, prompt, predefined_trajectory, interactions, and trajectory_file; artifacts generated_world, generated_4d_scene, and neoverse.mp4.

CLI task profiles: official-tilt-up-video. WorldFoundry binds NeoVersePipeline.

Family status: Verified.

Recorded official-tilt-up-video defaults: seed=42.

Runtime execution requires NeoVerse, Wan, reconstructor, LoRA, and CUDA assets.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run neoverse --pipeline.task-profile official-tilt-up-video. 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).

Recorded weight repositories include Yuppie1204/NeoVerse. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh neoverse. Add --download to fetch missing public checkpoints.

Status: Verified.