How do you access Genie 4?
Google AI Studio offers an interactive canvas and starter apps. Begin with a short scene description, a starting image or both. A custom app creates a genie-4 session through the Interactions API (v1beta). Google recommends WebRTC for interactive browser apps and documents WebSocket for jobs such as server-side processing and frame capture. A short-lived session token is enough for the browser. Keep the long-lived API key on the server.
What can you prototype?
An exhibition team can explore two fictional booths with different reception locations. A museum studio can compare different openings for a walkthrough. A creative team can discuss the same opening scene under different lighting. Agree on one question before testing, such as whether the reception desk is easy to spot from the entrance. Have two people take the same short route and record when they notice it. These are concept tests. Use appropriate plans, measurements and professional review for final decisions.
How consistent is the scene?
Genie 4 generates the live view in response to input. Google describes a rolling visual horizon, so objects can move or change when you turn away and look back. Test a change of direction and a return to the starting point. If people lose their bearings on a short route, the scene can still prompt a visual discussion. For repeatable routes or distances, use an environment with defined geometry.
What should an API pilot measure?
Measure time to first frame, movement latency and cost per completed session. Google bills for the generated stream on the paid tier. The documentation says closing a browser tab does not immediately end generation on the server. End sessions explicitly in your application and test what happens when a user leaves unexpectedly. Google marks the output with its imperceptible SynthID watermark.
When is it worth moving from Studio to a custom app?
Use the same starting point, route and question for both designs. Use fictional references or material approved for publication. Ask one person to assess the walkthrough as a visitor and another to assess it from the perspective of whoever will implement the design. Record observations about navigation separately from model errors. A reception desk that changes sides when someone turns around indicates a model error. If both visitors walk past the same unchanged desk from the same viewpoint, that is a reason to revisit the design.
For a one-off workshop, the AI Studio trial may be enough. A custom app can make sense when several people regularly need the same starting point or the walkthrough will sit inside an existing design interface. Check input responsiveness at the intended location using its usual connection. Repeat the test in a second short session under the same conditions. The team should be able to tell which observations recurred and which appeared only in the first stream. Finally, check whether the interactive version helped the team make a decision that was harder to assess from a static sketch or an ordinary video.
Keep it verifiable
Primary sources
- Google – Genie 4 is now available on the Gemini APISource checked:
- Google AI Studio – Interactive world generation with Genie 4Source checked:



