Genie 3: A new frontier for world models

9 months ago

Genie 3 is a new general-purpose world model capable of generating diverse interactive environments in real-time at 24 FPS and 720p resolution.
Genie 3 improves upon previous models (Genie 1 and Genie 2) by enabling real-time interaction, better consistency, and realism.
Key capabilities include modeling physical properties, simulating ecosystems, creating animations/fiction, and exploring historical settings.
Genie 3 supports promptable world events, allowing dynamic changes like weather shifts or introducing new objects/characters.
The model maintains environmental consistency for several minutes, a significant technical achievement.
Genie 3 was tested with Google DeepMind's SIMA agent, demonstrating its potential for training AI agents in complex environments.
Current limitations include a constrained action space, challenges in multi-agent simulation, and limited interaction duration.
Google DeepMind is releasing Genie 3 as a limited research preview to gather feedback and ensure responsible development.
Potential applications include education, training for robotics/autonomous systems, and generative media.
Acknowledgments highlight contributions from numerous researchers, engineers, and collaborators.

Hasty Briefsbeta