WorldFoundry

WorldFoundry is an open-source framework for world models. It treats world modeling as faithful simulation of environment dynamics rather than mere reproduction of observed appearances, and provides a shared stack for model runners, local assets, inference, and benchmark evaluation. Recent progress in generative video, 3D, emerging 4D representations, and embodied policies has made it easy to confuse perceptual realism with genuine simulation competence: a system may look convincing while still failing under action, camera change, long-horizon rollouts, or closed-loop control. WorldFoundry starts from that distinction and asks what infrastructure is required once “looking real” is no longer enough.

In practice, world-model research fragments across families. Video generators, reconstruction stacks, interactive worlds, and VLA policies rarely share checkpoint staging, artifact layouts, preview surfaces, or scoring contracts. Teams re-implement the same glue—environment setup, local asset paths, one-off launch scripts, and ad hoc metric notebooks—until small divergences become irreproducible leaderboard gaps. WorldFoundry collapses that glue into one workflow: stage checkpoints and datasets, launch generation through the TUI or CLI, inspect artifacts in Studio, then score outputs with in-tree metrics and reproducible scorecards. The same normalized run contract is meant to serve both interactive debugging and batch evaluation.

1. Follow the main workflow to install the environment, prepare local assets, choose the TUI or CLI, run inference, and score outputs.
2. Set up environments for the unified CUDA baseline and model- or benchmark-specific runtime requirements.
3. Stage Hugging Face checkpoints, datasets, metric weights, and outputs with the local assets guide.
4. Browse models and benchmarks in the TUI.
5. Run discovery, inference, scoring, and reports from the CLI.