OpenEnvision/WorldFoundry

Unified World Model Inference & Evaluation Infrastructure

What it solves

WorldFoundry provides a unified infrastructure for developing, running, and evaluating world models. It eliminates the fragmented nature of model repositories by providing a shared stack for inference, asset management, and benchmarking across various modalities including video generation, 3D/4D representations, and embodied AI actions.

How it works

The project organizes world models into a "model zoo" and benchmarks into a "benchmark zoo," using a standardized catalog system. It provides multiple interfaces for interaction:

  • Inference Interfaces: Users can generate artifacts via a Terminal User Interface (TUI), Command Line Interface (CLI), or a browser-based Studio workspace.
  • Runtimes: Model architectures and inference adapters are kept "in-tree" within the project, while large checkpoints and datasets are staged locally or via Hugging Face.
  • Evaluation Pipeline: An evaluation runner executes model-benchmark pairs, imports results, and generates normalized scorecards to validate performance claims.
  • Optimization: It supports various acceleration backends (like FlashAttention, xFormers, and quantization) with a built-in "truth matrix" to track certified quality and speedups.

Who it’s for

  • AI Researchers: Those working on world models, video generation, and embodied intelligence who need a standardized way to run and compare models.
  • ML Engineers: Developers building interactive worlds or VLA (Vision-Language-Action) policies.
  • Model Evaluators: Users who need rigorous, reproducible benchmark results with formal scorecards.

Highlights

  • Comprehensive Modality Support: Covers video, 3D/4D, and embodied action models.
  • Unified Workspace: A browser-based Studio for managing inference jobs and reviewing artifacts.
  • Integrated Benchmarking: A formal system for running benchmarks and generating eligibility-checked scorecards.
  • Flexible Inference: Supports TUI, CLI, and Studio interfaces for the same underlying model runtimes.
  • Hardware Acceleration: Built-in support for advanced kernels and quantization to optimize GPU performance.

Related

  • Project
    galilai-group/stable-worldmodel

    A platform for reproducible world model research and evaluation that provides a unified interface for data collection, training, and model-predictive control evaluation.

  • Project
    OpenWAM-Official/OpenWAM

    An open research stack for systematically developing World-Action Models (WAMs) that combine world knowledge and action learning for robotics.

  • Project
    benchflow-ai/benchflow

    A universal environment framework that runs AI agents against task environments and scores them through a standardized contract, treating benchmarks as frozen environments.

  • Project
    allenai/vla-evaluation-harness

    vla‑evaluation‑harness is a Python library that lets you evaluate Vision‑Language‑Action models on dozens of robot‑simulation benchmarks (LIBERO, CALVIN, SimplerEnv, etc.) using Docker‑isolated environments and lightweight model‑server scripts. It provides batch‑parallel execution (sharding episodes + GPU batching) for huge speed‑ups, a unified leaderboard of 2 456 models, and a simple CLI/Python API for both research and training‑time evaluation.

  • Project
    TheDesignFounder/DreamLayer-Eval

    An open-source benchmarking platform for image and video diffusion models that automates prompt sweeps, metric calculation, and reproducible run logging.