worldmodels.fyi
← Catalog
PaperGoogle DeepMind · 2024

Genie: Generative Interactive Environments

The first foundation world model trained unsupervised from internet video, generating action-controllable worlds.

Summary

Introduces Genie, an 11B-parameter foundation world model trained unsupervised from unlabelled internet videos. It comprises a spatiotemporal tokenizer, an autoregressive dynamics model, and a latent action model, and can generate action-controllable virtual worlds from images, text, or sketches. Basis for the Genie 2/3 line.

Metadata

Authors
Bruce et al., Google DeepMind
Year
2024
arXiv
2402.15391
Venue
ICML 2024 (Best Paper)
Introduces
Genie / Genie 2

Relationships

Introduces

Trained on

Unlabelled internet video
#interactive#foundation-model#latent-action