Project Genie Turns Prompts Into Interactive Worlds Users Can Change
Google DeepMind presents Project Genie as a browser-based tool for turning a text prompt or reference image into an explorable world that users can change while inside it. The company says its Genie 4 model supplies simulated physics, lighting and scene-matched audio, while a creation agent helps develop starting visuals and music and Gemini can suggest or trigger events during a session. Users can also publish worlds for others to explore or remix, with access and reuse controlled by separate settings.

A prompt or reference becomes a place to enter
Project Genie begins with a description or an image and turns it into a world the user can explore. The opening example is a prompt for an aerial view of Antarctica dotted with hot-air balloons. Another starts with a photo of yellow sunglasses and asks Genie to create a pop-up store based on the product’s brand world, shapes and colors. The distinction matters: the image is not merely the object to display; it is a reference for imagining the environment around it.
The product’s stated aim is to let someone create a world from a sentence and change it while inside. Google DeepMind describes a world as a place of sound, motion and moments that respond to the user. The demonstrations make that range concrete: a snowy balloon landscape, a door onto Mars or the bottom of the sea, a robot at work, and an ant colony viewed up close. They are examples of different settings and uses, rather than evidence that each scene behaves in the same way.
Project Genie is built with Genie 4, which Google DeepMind calls its newest world model. The announcement describes simulated physics, natural lighting, and audio and character dialogue matched to a scene. In the ski example, the user takes a downhill run while an announcer calls the race; a later prompt asks to see the course at night. The intended interaction is not only to specify an initial scene, but to issue a new instruction from within an ongoing experience.
That session limit belongs to upgraded Genie 3, not Genie 4. The announcement also lists 360-degree panorama grounding among the Genie 3 options in the model picker. Those are separate choices from the Genie 4 experience described in the demonstrations.
The agent helps develop a world; Gemini can act within the session
The Project Genie agent is presented as a collaborator for users who are unsure what to make next. Google DeepMind says users can bounce ideas off it and use its help to find what works in a world. The product announcement adds a more specific role: describe a story or simulation, and the agent generates starting visuals and custom background music. Gemini, meanwhile, observes in-session and can suggest or trigger live events.
These are distinct contributions. The agent helps establish a starting point and supports the process of developing an idea; Gemini is described as watching the session and responding with possible events. The announcement’s examples of natural-language changes include shifting the weather, introducing new characters and changing the action mid-session. Together, those functions frame the experience as a loop: define a world, explore it, and make changes without treating the first generated scene as final.
Publication settings determine what others can see and reuse
Sharing is not a single public-or-private switch. The Project Genie dialog presents three access levels: private, visible only to the creator; unlisted, viewable by anyone with the link; and public, listed in the gallery. The gallery itself is organized into categories including adventure, challenges, stories, simulation, creativity, education and exploration. A downhill ski adventure appears there with a description of a high-speed course beneath the Matterhorn.
Creators can separately allow remixing. The dialog says that enabling it also shares the underlying ingredients, including uploaded reference images and prompts. That means a creator deciding to permit a remix is not only letting others alter the visible world; the source materials and instructions may be shared as well. The interface also notes that an experience listed in the gallery may be reviewed for safety and policy compliance.
Other users can visit a listed world, copy its link or remix it into something new. The ski-course example shows how those actions connect to in-session creation: a world someone has published can become another person’s starting point, and that person can ask for a changed version, such as the same course at night. Sharing therefore extends the creation process, while the access and remix settings define who can enter and what they can build on.
Access and model options are separate from the creation workflow
Google DeepMind says Project Genie is rolling out to eligible Google AI Ultra subscribers aged 18 and older; Google AI Pro members can join a waitlist. The closing screen directs viewers to labs.google/ProjectGenie. These access details sit alongside, but do not change, the distinctions among the product’s creation tools: Genie 4 is presented as the world model behind the experience, the Project Genie agent helps generate starting visuals and music, Gemini can suggest or trigger live events, and upgraded Genie 3 capabilities are available through the model picker.