One infrastructure. Every world model.
WorldFoundry treats world modeling as faithful simulation of the physical world, not just reproduction of observed appearances. Its shared capability taxonomy spans perception, manipulation, dynamics, and interaction.
Beyond Perceptual Realism
Recent advances in generative modeling across video, 3D, and emerging 4D modalities have fueled the view that these systems are evolving into world models capable of simulating dynamic environments. Yet existing work often conflates perceptual realism with genuine simulation ability.
Perceptual realism, however impressive, is neither necessary nor sufficient for genuine simulation competence. WorldFoundry clarifies this boundary, separating perceptual distribution modeling from stateful, action-conditioned simulation.
Unified Taxonomy and Operational Definition
A fundamental obstacle to progress in world modeling research is the absence of a precise and shared conceptual foundation. We formalize world modeling as the objective of simulating environment dynamics rather than merely reproducing observations, and introduce a capability gradient that traces the progression from perception and representation to manipulation, dynamics, and interaction.
This taxonomy clarifies the boundary between foundation generative models and generative world models. A generative world model must support three tightly coupled operations: estimating latent state from observations, predicting future states under physical constraints, and coherently adapting futures under agent-driven interventions.
A Unified Inference and Evaluation Infrastructure
Contemporary systems span a broad landscape, yet the infrastructure used to study them remains fragmented. WorldFoundry provides a shared abstraction for model deployment and assessment. It focuses on unified inference across multimodal observations, benchmark evaluation, and interactive visual analytics—training is intentionally out of scope for the current release.
Through Artifact and Representation Contracts, WorldFoundry standardizes how heterogeneous models expose their predictions and world-state estimates, enabling cross-family comparison without collapsing systems into an undifferentiated benchmark wrapper.
The infrastructure is designed for contributors who need repeatable inference demos, reviewable metrics, and simple extension points. Start with the quickstart, validate inference outputs, then run evaluation when the selected model family and benchmark support it.