OpenEnvision/WorldFoundry

What it solves

WorldFoundry provides a unified infrastructure for developing, running, and evaluating world models. It eliminates the friction of managing disparate environments, asset staging, and benchmark layouts across different modalities like video generation, 3D/4D representations, and embodied AI action policies.

How it works

The project implements a shared stack that separates model architecture and inference adapters from the actual checkpoints and datasets. It uses a unified GPU environment (via Conda) to handle multiple models, while providing several interfaces for interaction:

• TUI/CLI: An interactive picker for models and benchmarks that generates runnable commands.
• Studio: A browser-based UI for managing inference jobs and reviewing artifacts.
• Evaluation Runner: A system that runs model-benchmark pairs and produces normalized scorecards.
• In-tree Runtimes: Integrated synthesis and pipeline modules that allow for consistent inference paths across various integrated models.

Who it’s for

It is designed for researchers and developers working on world models, video generation, and embodied intelligence who need a standardized way to run inference and benchmark their models against others.

Highlights

• Extensive Model Zoo: Integrates a wide array of models including Wan, HunyuanVideo, LTX2, Cosmos, and various VLA/action policies (e.g., OpenVLA, Octo).
• Unified Backend: Supports advanced attention backends (FlashAttention 2/3, SageAttention, xFormers) and NVFP4 quantization.
• Flexible Asset Management: Supports native Hugging Face Hub loading alongside local checkpoint staging.
• Benchmark Hub: Includes a catalog of manifests and official runner constraints for benchmarks like LaryBench and WorldReasonBench.
• Multi-GPU Support: Features Context/Sequence Parallelism and advanced memory management for high-compute hardware (A100, H100).