Exploring Google’s Project Genie: Interactive Worlds with Genie 3

I’m Alex, filling in for Matthew Berman, and today I’m trying Google DeepMind’s Project Genie. The experimental web app uses the Genie 3 world model to generate interactive environments as you explore. Genie 3 was previewed in August 2025; this January 2026 video tests the newer Project Genie prototype.

The opening screen offers sample worlds or a way to create your own. I start with a backyard racetrack, controlling a small blue car with keyboard movement and camera keys. The scene looks impressive, although the controls are laggy and difficult to steer. A timer ends the simulation after roughly a minute.

After the run, I can download a video without the interface, though it carries a Genie 3 watermark. The application also lets me reuse the original prompt. Here the description is straightforward: a backyard track and a blue toy car.

To make a new world I use ChatGPT to help write an environment and character prompt: an enormous organic alien structure in space, with moving corridors and a human explorer in a futuristic suit. The creation interface accepts a starting image, randomizes prompts and lets me choose first- or third-person perspective.

Project Genie first produces a sketch that can be edited before entering the world. I change the alien environment from purple to orange and red. The revised preview preserves much of the original layout while changing the color, which makes it useful for adjusting the scene before generation.

Inside the alien world, the result does not behave as I expected from a first-person selection: I appear to control a separate astronaut, and another character follows. The camera and jump controls are unreliable, and the scene shifts from orange to blue as I move. I am not sure whether I selected the wrong perspective or the generation failed to follow it.

I try again with floating cloud platforms and cities above an endless sky, inspired by Skyward Sword. This time the initial sketch shows hands in first person, and the generated scene does too. It looks more like snow than cloud, but I can jump and move around.

I jump off an edge, fall for a while and land in another area below. That is a fun surprise, although latency makes the movement hard to control: I have to press jump ahead of when I want it to happen. These generated scenes are exciting, but the session does not establish that the world has a persistent, physically consistent map.

Google describes Project Genie as a prototype powered by Genie 3, Nano Banana Pro and Gemini. Its three functions are world sketching, exploration and remixing. Nano Banana Pro is used for the editable preview; I speculate about how the other components divide the work, but the demonstration does not verify the internal implementation.

World sketching combines text with generated or uploaded images, a character and a chosen viewpoint. During exploration, Google says the model generates the path ahead in real time based on the user’s actions. Remixing lets people adapt existing prompts or gallery examples and download a video of the resulting exploration.

Google lists limitations: scenes may not follow the prompt, image or real-world physics closely; controls can be imperfect or delayed; and generations in this prototype are limited to 60 seconds. Promptable events shown in the earlier Genie 3 preview were not included in this version. At the January 2026 launch, access began with eligible US Google AI Ultra subscribers aged 18 or older; availability can change. See [Google’s launch description](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie/).

I remix the racing example by asking for a red car and purple grass. The revised sketch matches the request but also changes some leaves toward autumn colors. In the playable scene the car and grass change, while some trees remain green, and the car passes through a wall. The result shows both prompt responsiveness and inconsistent details.

Next I upload an image of Matt converted into a LEGO-like figure and ask for a LEGO city with a LEGO character in third person. The character’s legs look strange, but I can move around the streets and see cars driving. The uploaded image provides a starting point rather than a guarantee that the figure will remain anatomically consistent.

Finally I try the dice icon for randomization. In this session it appears to select from existing example environments and characters rather than inventing an entirely new random pairing. That is an observation from these attempts, not a comprehensive test of every possible result.