WorldFoundry Adds Shared Evaluation Tools for Generative World Models

WorldFoundry has introduced an open-source framework for running and comparing generative world models across video, 3D and 4D reconstruction, embodied action and interactive systems, with the aim of making evaluations reproducible across model types.

The framework defines a generative world model as a system that keeps a persistent representation shaped by past observations, updates that representation in response to interventions and produces coherent trajectories over time. Those trajectories can reflect observed events as well as counterfactual outcomes, such as what might follow from a different action. The framework treats visual plausibility alone as insufficient, emphasizing persistence, dynamics and causal responses.

WorldFoundry’s catalog contains 270 model entries, 74 benchmarks and 62 registered metric entries. The entries cover different kinds of systems, from video generation to 3D and 4D reconstruction, embodied action and interactive environments. The framework brings model and benchmark discovery together with execution and evaluation workflows.

Users can run models through command-line tools, a terminal interface or a browser-based Studio. The framework also provides artifact management and visual inspection tools for generated media and geometry. These functions support a workflow in which model runs can be examined alongside the evidence associated with their results.

WorldFoundry’s catalog entries include readiness information and execution evidence. Its scorecards also distinguish execution readiness from evaluative claims, linking those claims to supporting artifacts. The structure is intended to make it possible to trace how an evaluation result was produced rather than treating a score as self-explanatory.

The project’s cross-model evaluation focus differs from systems built for a particular application. Baidu’s Apollo World system for training and validating Apollo Go robotaxis, for example, is designed around autonomous-vehicle development. WorldFoundry instead organizes tools and evidence for comparisons across several categories of generative world models.

Its stated goal is to support reproducible comparisons of persistence, controllable dynamics and causal responses across heterogeneous systems. The framework’s catalog and evidence structure provide a common way to discover models, run evaluations and inspect the artifacts behind claims.