wtf is worldfoundry?

WorldFoundry is an open source infrastructure project for working with world models, which are AI systems trained to generate and understand video, 3D and 4D scenes, and simulated environments. It brings together the pieces needed to run these models in one place: environment setup, staging for large model checkpoints and datasets, an inference layer you can drive through a terminal interface, a command line tool, scripts, or a graphical Studio app, and a system for scoring model outputs against benchmarks. The day one workflow the project describes has three steps. First you set up your environment and download the model files and datasets you need, kept separate from the code itself. Second you run inference, meaning you generate outputs such as video clips and inspect them using whichever interface suits you. Third, once your generated files match the expected benchmark layout, you can run evaluation to score how well a model performed, with the project cautioning that any performance claims should be backed by a matching run record and scorecard. Under the hood, WorldFoundry integrates a large number of existing video generation and world model systems, including projects named Wan, HunyuanVideo, LTX2, and Cosmos, along with tools for perception and 3D understanding. It supports several GPU acceleration backends and is built to run efficiently across multiple graphics cards. It also folds in action policy models used in robotics research, and ships with a benchmark catalog covering video generation, 3D and 4D representation, embodied action, and interactive worlds. The project is written in Python, requires versions 3.10 through 3.13, and is released under the Apache 2.0 license, which permits commercial use. It is still under active development, so its exact feature set and documentation are expected to keep changing. It maintains community channels on Discord and WeChat for support and updates. The full README is longer than what was shown.