Posted 3/5/2024, 6:01:13 PM
Google's Genie AI Creates Interactive Games from Images Using 30,000 Hours of Video Game Footage
- Google's Genie AI model converts images into interactive, playable 2D platform environments that can be explored.
- Genie was trained on 30,000 hours of 2D gaming videos to develop a system for predicting how game frames logically connect.
- It builds latent action and dynamics models to guess how frames should change based on button inputs.
- Genie doesn't need textual guides or labels; it infers actions from frame changes in videos.
- Researchers believe Genie shows how robust world modeling trained on large-scale data could empower future AI.