Genie 4: a world you can walk through, with sound and on command
Google DeepMind drops Genie 4, a world model that keeps running live, makes its own sound and listens to your voice. What that means for previz and scouting.
Miraki MindAI-assisted editorial · 28 October 2026 · 6 min read
On 27 October 2026, Google DeepMind announced Genie 4, a world model for real-time interactive simulation. You move through a scene and the world keeps computing as you go, with its own sound: ambience, music and dialogue. You can change the scene with language, for example the weather, the camera or a new character.
A place isn't made live at our end, it's made as a location sheet. Characters, places and light are approved as sheets before a shot is made. In It Ain't Coming Off, Dee's trailer exists twice: once by day, once at night, same geometry, different light. Those are two separate sheets, both built with Nano Banana, and the shots after that with Seedance 2.5.
Genie 4 promises changes like that on command: different weather, a different camera, a new character, in the middle of the running scene. And DeepMind names film explicitly. So the model touches three things we solve with sheets today: previz, finding the place and developing the look.
What is actually new
A video model delivers a finished clip. A world model like Genie 4 keeps running as long as you're inside it. Navigation and camera go through the keyboard, changes through text. According to DeepMind, Gemini adapts your inputs to the scene and keeps the world going between two inputs.
- Steering with language: change the weather, move the camera, bring in a person or an object while the simulation runs.
- Sound: according to DeepMind, Genie 4 is its first real-time world model with native sound. The ambience fits the place, characters hold dialogue that reacts to your inputs, and there's music.
- Film: under "Learnings from trusted testers", DeepMind describes how people from film are testing Genie 4 as a quick draft mode for previz. They step into a scene, run several "takes" and adjust camera, lighting, dialogue and action as they go.
What's missing are the hard numbers. On resolution, frame rate, price and regions, DeepMind says nothing for Genie 4. For comparison: for Genie 3, from August 2025, DeepMind gave 720p and 24 frames per second. Whether Genie 4 goes beyond that is not documented.
What it means for a production
Previz: this is where the use is most tangible. Walk a scene, try blocking and camera, change the light, before a sheet even exists. It's a tool for finding ideas, not for pinning them down.
Location scouting: a sandbox for invented places. Not for real ones. DeepMind itself writes that Genie 4 can't simulate real-world locations exactly. In the Polestar 3 spec ad, the places were real, among them a bridge in Bern and the Grimsel Pass, and each was rebuilt as a plate and a sheet. That's exactly where an approximate world doesn't help much.
Look development: here's the catch. According to DeepMind, details and objects can drift over time, especially in longer runs, after several prompt edits and while moving around. A sheet does the opposite: it locks a look so it's the same in every shot. So Genie 4 can show where a look might go. Pinning it down is still the job of an approved sheet.
Then there's duration. The model page lists it under limitations: a few minutes of continuous interaction, not long sessions. Whether and at what quality picture and sound can be exported from a session isn't stated anywhere.
Genie 4 remains an early, evolving area of research.
Google DeepMind
What it means for a brand
Labelling: according to DeepMind, all video and audio from Genie 4 carries SynthID, a watermark for AI content in picture and sound. At our end, AI material gets labelled anyway, whatever model it comes from.
Rights: neither the announcement nor the model page says who owns the output or what you're allowed to do with it commercially. Before material from Genie 4 goes into a campaign, you need the terms of use that apply to the access in question. The sources on Genie 4 neither summarise nor link them.
References: an example prompt in the DeepMind post uses a reference image as the first frame. How uploaded images are handled, say a product before launch, isn't described.
Our take and the limits
For us, Genie 4 is still a demo today. Only selected testers have access, in a research preview. Google AI Ultra subscribers are supposed to follow "soon" through Project Genie, plus early access partners through the Gemini API and AI Studio, with no date. Resolution, frame rate, export, price and regions are unknown; the length is only roughly outlined as "a few minutes".
The direction is still interesting. A place you can walk through, relight and hear would be a good tool before the sheet. It doesn't replace the sheet: previz lives on variants, production needs a fixed state. We'll look at Genie 4 as soon as we have access, and then we want to know: can an approved location sheet be the starting point? Does the world hold the look across a scene? And what are you allowed to do with the result?
Sources
- Google: We're announcing Genie 4, our new world model for real-time interactive simulation, 27 October 2026
- Google DeepMind: Genie 4: A new generation of world simulation, 27 October 2026
- Google DeepMind: Genie model page
- Google DeepMind: Genie 3: A new frontier for world models, 5 August 2025
- Miraki Lab: how we work with sheets
- Miraki: It Ain't Coming Off
- Miraki: Polestar 3 (spec ad)