Project Genie: Google DeepMind's Interactive World Model Prototype

Google DeepMind has released Project Genie, an experimental research prototype that enables users to create and explore interactive, real-time generated worlds. This tool is currently available to Google AI Ultra subscribers in the U.S. (ages 18+).

Real-Time World Simulation via Genie 3

Project Genie is powered by the Genie 3 world model, which simulates the dynamics of an environment and predicts how actions affect the state of that environment. Unlike static 3D snapshots, Genie 3 generates the path ahead in real time as a user moves and interacts with the world.

This capability allows for the simulation of diverse scenarios, including:

  • Robotics simulations
  • Animation and fiction modeling
  • Exploration of historical settings and specific locations

Core Capabilities of Project Genie

Project Genie is a web application that integrates Genie 3, Nano Banana Pro, and Gemini to provide three primary user functions:

1. World Sketching

Users can create living, expanding environments by prompting with text, generated images, or uploaded images. This feature allows users to define characters and exploration methods (e.g., walking, riding, flying, or driving).

Integration with Nano Banana Pro enables "World Sketching," which allows users to preview and modify images to fine-tune the world's appearance and select the character perspective (first-person or third-person) before entering the environment.

2. World Exploration

Project Genie generates the navigable environment in real time based on the user's actions. Users can traverse the world and adjust the camera perspective as they move through the generated space.

3. World Remixing

Users can build upon existing prompts to remix worlds into new interpretations. The platform includes a gallery of curated worlds and a randomizer icon for inspiration. Users also have the option to download videos of their explorations and created worlds.

Current Limitations and Responsible Development

As an early research prototype, Project Genie and the underlying Genie 3 model have several known limitations:

  • Visual and Physical Fidelity: Generated worlds may not always appear true-to-life or strictly adhere to prompts, images, or real-world physics.
  • Control and Latency: Characters may occasionally exhibit lower controllability or higher latency in response to user inputs.
  • Duration: Generations are currently limited to 60 seconds.
  • Feature Gaps: Certain capabilities announced in August, such as promptable events that change the world during exploration, are not yet included in the prototype.

Google DeepMind states that the goal of this release is to better understand how world models will be used in AI research and generative media.

Sources