Google expands access to Project Genie, its AI tool that turns photos and text into explorable worlds

The experimental system blends video generation and real-time control inside a browser

Serving tech enthusiasts for over 25 years.
TechSpot means tech analysis and advice you can trust.

What just happened? Google's most exclusive AI tool right now isn't a chatbot or an image generator – it's a world builder. The company has expanded access to Project Genie, an experimental system that transforms simple text instructions into short, explorable video environments. Designed as a research prototype, Genie combines several of Google's latest models to simulate responsive, dynamic worlds within a web app.

Under the hood, Project Genie evolves from Genie 3, a world model first demonstrated by Google DeepMind last year. Unlike fully 3D game engines, world models don't construct continuous 3D geometry. Instead, they generate video sequences that appear interactive, adjusting what's shown in response to input commands. That gives the illusion of real exploration, though, technically, what users see is a constantly regenerated stream of AI-generated frames.

Project Genie expands on that concept with what Google calls world sketching. The system begins by generating a still reference image using its Nano Banana Pro model. Users can either provide a picture or describe what they want.

Once satisfied with the reference image, users can pass it to Genie to build a playable sequence around it. The environment unfolds as a 720p video at roughly 24 frames per second, with controls that mimic standard PC navigation keys. Each world lasts about 60 seconds before resetting, though users can rerun or remix scenes to create new variations.

Behind Genie's creative output are updated integrations with Gemini 3 and Nano Banana Pro, which support the system's spatial consistency and texture generation.

// Related Stories

• Roborock's Rover takes robot vacuums where none have gone before: up the stairs
• Google is putting even more AI into Gmail, and some of it is on by default

A major advantage over earlier versions is the AI's long-term memory, allowing it to retain visual context slightly longer. This results in generated worlds with fewer abrupt shifts in lighting and geometry. However, this "long memory" lasts only a couple of minutes of rendering, highlighting the experimental nature of the technology.

Google has seeded Project Genie with a library of sample environments, but the real appeal lies in custom creation. Users can direct the AI to generate different climates, architectural styles, or character interactions, and then download a video of the traversal.

Google says interactivity is what makes Project Genie noteworthy – it produces results nearly in real time, a capability that other video generators, such as Veo or OpenAI's Sora, have yet to achieve. Those systems typically render video clips passively, offering no user control once generation begins.

Despite these advances, Google stresses that Project Genie remains a research prototype with clear limitations. Each rendering still experiences input lag, playable sessions are of fixed length, and the previously demonstrated ability to edit running simulations has not yet been implemented.

Some environments can behave unpredictably, especially when the AI's physics models falter, resulting in sequences where motion or gravity appear unrealistic. Google cautions testers that such inconsistencies are expected and part of ongoing experimentation.

Access to Project Genie is intentionally restricted. It runs exclusively through a dedicated web app, separate from Google's broader Gemini interface, and is currently available only to subscribers on the AI Ultra plan, which costs $250 per month. Google notes that broader availability may follow as infrastructure and capacity improve.

In a test by The Verge, the system was initially able to replicate familiar gaming styles reminiscent of Super Mario and The Legend of Zelda, though content filters later blocked such prompts due to third-party intellectual property concerns.

For now, Project Genie functions less as a game engine and more as a proof of concept – a demonstration that neural models can approximate a world rather than merely depict one. Even in its brief, one-minute interactive bursts, it offers a glimpse of a future where AI doesn't just describe experiences but dynamically assembles them in real time, frame by frame.