Google DeepMind Launches Project Genie for AI-Generated Interactive Worlds

Google DeepMind has unveiled Project Genie, an AI tool that allows users to create and explore interactive game worlds using text prompts or images. Starting Thursday, Google AI Ultra subscribers in the U.S. can access this experimental research prototype, which combines DeepMind's latest world model, Genie 3, with the image generation model Nano Banana Pro and the Gemini system.

This initiative follows the research preview of Genie 3 five months ago and aims to gather user feedback and training data to advance world model development. World models are AI systems that generate internal representations of environments, enabling predictions of future outcomes and action planning. DeepMind views these models as crucial steps toward achieving artificial general intelligence (AGI) and envisions applications in video games, entertainment, and robotics training.

The release of Project Genie coincides with increased activity in world model development. Fei-Fei Li's World Labs recently launched its first commercial product, Marble, while AI video generation startup Runway has introduced its own world model. Additionally, AMI Labs, founded by former Meta chief scientist Yann LeCun, is focusing on developing world models.

Shlomi Fruchter, a research director at DeepMind, expressed enthusiasm about the broader access to Project Genie, stating, "I think it's exciting to be in a place where we can have more people access it and give us feedback."

DeepMind acknowledges the experimental nature of Project Genie, noting that while it can generate playable worlds, results may vary. Users begin by providing text prompts for the environment and main character, which Nano Banana Pro uses to create an image. This image serves as a foundation for Genie to generate an interactive world. Users can modify the image before world generation and can also use real-life photos as baselines, though outcomes may be inconsistent.

Once the world is generated, users can explore it in first or third-person view, remix existing worlds, or explore curated worlds in the gallery. Currently, DeepMind limits world generation and navigation to 60 seconds due to budget and computational constraints. Fruchter explained, "The reason we limit it to 60 seconds is because we wanted to bring it to more users. Basically when you're using it, there's a chip somewhere that's only yours and it's being dedicated to your session."

He added that extending the duration would diminish the incremental value of testing, noting, "The environments are interesting, but at some point, because of their level of interaction and the dynamism of the environment is somewhat limited. Still, we see that as a limitation we hope to improve on."

In testing, Project Genie performed well with artistic prompts, such as watercolors or anime styles, but struggled with photorealistic or cinematic worlds, often producing results resembling video games rather than real-life settings. Using real photos as baselines also yielded mixed results, sometimes generating sterile, digital-looking environments.

Interactivity is another area for improvement. Users reported instances where characters walked through walls or objects. Additionally, navigation controls were sometimes unresponsive or sent users in unintended directions.

Fruchter acknowledged these shortcomings, emphasizing that Project Genie is an experimental prototype. He stated that the team aims to enhance realism and improve interaction capabilities, including giving users more control over actions and environments.

"We don't think about [Project Genie] as an end-to-end product that people can go back to every day, but we think there is already a glimpse of something that's interesting and unique and can't be done in another way," Fruchter said.

As Project Genie continues to evolve, DeepMind plans to address these limitations and expand access to more users, aiming to refine the tool based on user feedback and further research.