Genie by DeepMind
Genie is a research foundation world model developed by Google DeepMind that generates interactive, action-controllable 3D environments from text, images, or sketches. Trained using unsupervised learning on internet videos, Genie simulates a wide range of realistic and imaginative scenarios for both human and agent interaction.
• Generates dynamic environments using text, image, or sketch inputs
• Learns from unlabelled internet video data using unsupervised techniques
• Predicts frame-by-frame visual dynamics via autoregressive modeling
• Employs a latent action interface for structured, scalable interaction
• Designed to evaluate and train AI agents in diverse, open-ended 3D simulations
• Currently available as a research model only, not a deployable development platform