Google previews Genie 4 real-time worlds with sound and responsive characters

• Genie 4 generates explorable worlds with native sound from an environment-and-character prompt, as venturetwins demonstrated in early access.
• Scenes can be redirected while running: bilawalsidhu’s demo changes camera angles, lighting, sound and action between takes.
• Setup typically takes under 30 seconds for two minutes of interaction, according to venturetwins’ reply.
• API access is coming soon for the first time, according to the preview announcement. Google is starting with trusted testers.

Google’s Antarctic balloon prompt puts the camera’s position in the “character” field, while the watermelon setup puts a fruit there. An early tester also stepped inside a photograph of Berlin days after the Wall opened.

With early access from Google, bilawalsidhu built a black-triangle UFO scene while adjusting camera angles, lighting, sound and action between takes.

Live direction is the star of this preview. Google’s technical introduction describes the workflow:

1. A starting text prompt or reference image establishes the scene.
2. Keyboard controls handle navigation and camera movement.
3. New text prompts change the running simulation, including its weather, objects and people.

Voice direction remains on the tester’s wish list. He also suggested using Genie as a real-time draft model for a future Veo, a proposed workflow rather than an announced integration.

Google’s announcement describes sound generated alongside the world:

• Ambient noise matched to the setting.
• Music generated within the simulation.
• Character dialogue that responds to inputs and events in real time.

The published saxophonist example specifies sound as tightly as the image: lake ambience, wind, water and the instrument itself, with no background music beyond instruments physically present in the scene.

Google publishes the starting prompts in its demo gallery. The examples collected by minchoi range from a short documentary brief to densely specified production designs:

• Motorcycle close-up: The Hong Kong scene specifies a 1995 romantic-drama aesthetic, 35mm film, overlapping faces, shallow depth of field, black bars and one continuous shot.
• Porcelain salon: In the wizard scene, the brief specifies wand materials, ceramic glaze, skin texture, anamorphic optics and small sounds such as floorboard creaks and fingers adjusting their grip.
• Stained-glass landscape: Every surface must retain the material treatment, including water, clouds and sky, according to the stained-glass prompt. It explicitly excludes natural textures that would break the style.
• Japanese station: Crowd density and composition get explicit constraints in the train-station prompt: the red train occupies the nearest track, passengers fill the platform, and the framing stays asymmetrical.
• Crane in snow: Google uses a much shorter brief in the wildlife example: “crowned crane walks across the snow beneath the clear dusk gradient sky. Natural colors in a documentary style”.

The motorcycle scene, prompted as one continuous 35mm shot

An explorable landscape with a consistent stained-glass treatment

Two short demos probe physical interactions: a statue drops onto a watermelon, and wind moves silk through a meadow.

Genie 4 draws on Gemini models with distinct roles. Gemini Omni anchors appearance and behavior, while Gemini 3.8 Flash acts as a guiding agent:

1. Observes the generated scene.
2. Breaks goals into a sequence of steps.
3. Reviews incoming frames to check whether each step succeeded.
4. Adapts its plan when the world changes or the user intervenes.

In a separate robotics experiment, Gemini 4 Argon supplies step-by-step text captions for Genie to simulate Apptronik’s Apollo humanoid playing checkers.

A simulated UR5 arm also dispenses liquid into a lab plate, triggering tiny electrical sparks. These examples are generated task simulations; Google describes exploring them as environments for developing robots and agents.

The October 27 release is a research preview for trusted testers. Google’s rollout plan names Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio as the next audiences.

A Project Genie waitlist is open. The announcement gives no public release date or API pricing.

Early-access user venturetwins reports under 30 seconds of setup followed by two minutes of real-time interaction. They also clarified that the Waymo-style driving video collects favorite clips from sessions.

DeepMind’s limitations statement identifies several unfinished areas:

• Visual details and object permanence can drift during longer runs, repeated complex edits or navigation through a scene.
• Complex physical interactions and spatial accuracy remain research problems.
• Real-time audio accuracy still needs improvement.
• Rendering legible text remains difficult.

Google also says generated video and audio will carry SynthID watermarks identifying them as AI-generated content.