WorldFoundry
English中文
World modelsUnified inference and evaluation across video, 3D/4D, embodied, and interactive worlds.

World models stay heterogeneous. Shared catalogs, artifacts, and scorecards are what make them comparable.

Try it out

Configure a run

Choose a real catalog entry. Runtime and version facts come from the repository manifests, and the command stays copyable.

Selected modelWan2.2Alibaba (Tongyi Lab)
Runtimeworldfoundry-unified-cu128Python 3.11 · CUDA 12.8
ReadinessRunner verifiedSeparate from runner evidence
Command preview
worldfoundry-eval run wan2.2 \
  --pipeline.task-profile text-to-video

Why WorldFoundry

Commensurability without homogenization.

Shared catalogs, durable artifacts, and evidence-aware scorecards — while native inputs, execution semantics, and evaluation protocols stay model-specific.

Cataloged

Declarative model and benchmark manifests, plus explicit runtime and asset management, make heterogeneous systems discoverable before compute is spent.

Artifact-centric

Durable request and result contracts decouple generation from scoring. Videos, geometry, actions, and traces can be inspected or rescored without reloading the model.

Evidence-aware

Scorecards keep declared support, operational readiness, and validated performance separate, so a catalog entry or a demo is never mistaken for a benchmark claim.

Read the introduction

Catalog

275 model entries, one operational vocabulary.

Video, 3D/4D, interactive worlds, and embodied systems share manifests and readiness signals. Catalog inclusion is not uniform runtime maturity.

275 models74 benchmarks

Video

Text-, image-, and video-conditioned generation, editing, and audio-video systems.

62 models

Capability progression

From generation to world intelligence.

Visual plausibility is not world modeling. A system exhibits stronger competence as representations become persistent, transitions become intervention-responsive, and predictions become useful in a closed loop.

Foundational

Perceptual grounding

Interpret, reconstruct, or synthesize observations. This is the perceptual basis for state inference — not yet a persistent model of the environment.

World-model spine

persistent state → dynamics prediction → interactive feedback

  1. Persistent state

    Infer and maintain structured environmental state across observations, identities, geometry, and partial views.

  2. Dynamics

    Predict coherent transitions under endogenous processes, actions, and interventions, so earlier events constrain later outcomes.

  3. Closed-loop interaction

    Act, observe, and update the same state. Feedback revises what the model believes before the next prediction.

Orthogonal

Controllable transformation

Selectively edit represented content while leaving the rest intact. Manipulation can strengthen any stage, but it is not by itself dynamics or interaction.

Read the operational definition

End to end

Operate with shared contracts.

Catalog, runtime, workspace, and evaluation share identities and durable outputs.

Know what exists

Manifests expose stable IDs, sources, capabilities, assets, runtime bindings, readiness, and blockers before compute is allocated.

See what is included

Run through shared boundaries

Pipelines preserve model-native I/O while TUI, CLI, Studio, Python, and MCP reuse the same execution contracts.

Understand the design

Inspect before scoring

Videos, geometry, actions, trajectories, and traces stay visible on disk and in Studio instead of disappearing inside scripts.

Explore Studio

Evidence, not headlines

Benchmark runners preserve per-sample outcomes, coverage, provenance, blockers, reports, and scorecards.

Read about evaluation

Five-stage pipeline

How it works

Every surface follows the same sequence. The artifact is the handoff between model execution and benchmark evaluation.

  1. Discover

    Choose a model and benchmark

    Catalog manifests expose stable IDs, capabilities, readiness, assets, and blockers.

  2. Prepare

    Stage runtime and assets

    Resolve the conda profile, checkpoints, datasets, metric weights, and credentials.

  3. Run

    Generate through one contract

    TUI, CLI, scripts, and Studio dispatch to a pipeline and operator.

  4. Inspect

    Review normalized artifacts

    Videos, geometry, actions, traces, and metadata remain visible and reusable.

  5. Evaluate

    Produce reviewable evidence

    Metrics and official runners write reports, blockers, and a normalized scorecard.

Artifacts are the handoff: model runtimes do not own benchmark logic, and benchmark runners do not load model checkpoints.

Everyone welcome

Got questions? We're here to help.

Whether you're just getting started or debugging a complex run, the community is open to everyone. No question is too basic.

  • Fast & friendly responses
  • Active maintainers
WorldFoundry WeChat community QR code
WeChat CommunityScan to join. The QR code is updated periodically if it expires.

Around the project

Resources & ecosystem

Notes, events, and sibling projects in the OpenEnvision network.

Blog

Updates, technical notes, and release highlights.

Events

Meetups, demos, benchmark sprints, and milestones.

OpenEnvision

The lab organization behind WorldFoundry.

Awesome World Modeling

Curated papers, models, and resources for world modeling.

BlogrXiv

Curated index for technical AI research blogs and writing.

WorldFoundry on GitHub

Source, issues, discussions, and releases.