Google Enables Much Faster Game Development With Genie 3 AI

Google DeepMind has introduced Genie 3, the latest version of its AI “world model,” capable of generating 3D interactive environments in real-time.

The model allows both users and AI agents to quickly create and move through simulated game worlds for longer periods than previous versions while maintaining visual memory of objects and scenes.

Advancing AI World Models

World models are AI systems that simulate virtual environments for purposes like education, entertainment, and robotic training. Users can enter a prompt, and the AI generates a 3D space that functions like a video game environment but is entirely AI-generated instead of being built with pre-made assets, which takes longer and also costs more.

DeepMind first demonstrated this approach with Genie 2 in December 2024, which could create interactive worlds from a single image; however, it had major limitations. Interactions typically lasted only 10 to 20 seconds, and objects in the world changed unpredictably when users looked away and back again.

Key Upgrades of Genie 3

Genie 3 significantly extends the capabilities of its predecessor:

• Longer Interactions: Users can now explore worlds for a few minutes, compared to Genie 2’s short, seconds-long experiences.
• Visual Memory: The model remembers object positions for about one minute. For example, writing on a chalkboard or painting on a wall will remain in the same place even if a user looks away.
• Improved Visual Quality: The generated worlds run at 720p and 24 frames per second, providing a smoother and clearer experience than before.
• More Customization: Users can modify conditions with prompts, such as changing weather or adding new characters dynamically.

Limited Research Preview

Despite the improvements, Genie 3 will not be widely available initially. DeepMind is launching it as a limited research preview for a small group of academics and creators to study potential risks and usage patterns.

Google has set strict usage limitations, including:

• Restricted interactions with objects in generated worlds
• Limited generation of legible text, unless provided in the initial prompt
• Controlled testing to mitigate misuse before broader release

DeepMind says it is exploring ways to expand access to more testers in the future.