Jafar: An Open-Source Genie Reimplemention in Jax

Timon Willi, Matthew Thomas Jackson, Jakob Nicolaus Foerster

Keywords: foundational model, video generation, generative model, world model

TL;DR: We reimplement and reproduce Genie: Generative Interactive Environments.

Abstract: We introduce Jafar, an open-source Jax reimplementation of Genie, a foundational world model. Genie was the first world model trained in an unsupervised manner on unlabelled internet data. We follow the reproducibility study in the appendix of Genie and evaluate our reimplementation on the CoinRun environment from the procgen suite. Jafar is implemented in JAX and produces visually consistent video generation and can be trained on a single GPU.

Submission Number: 14