Google DeepMind Genie 4: a real-time AI world you can steer
Google DeepMind unveiled Genie 4 on October 27, 2026: a model that builds a photorealistic world with sound from a text or image prompt, and lets you change it live. What it does, who can try it, its limits, and what you can already do to plan a shot.
Frank HoubreFounder of Imaginode
6 min read
what Google DeepMind announced on October 27
On October 27, 2026, Google DeepMind introduced Genie 4, a world model that builds a photorealistic world with sound in real time. You walk around it with the keyboard and change it by typing sentences.
DeepMind's post calls it a new generation of world simulation. In practice, you describe a place and a character, or you hand it an image, and Genie 4 creates a scene you can move through with the usual game controls. While you move, you can type "it starts raining", "switch to a top-down camera" or "a passerby walks into the street", and the world changes without stopping.
The big addition over Genie 3 is sound. Every world generates its own ambience, music and even dialogue on the fly: characters can talk and answer you. According to TechCrunch, a scene now stays consistent for up to two minutes, twice Genie 3's limit. Google also points to better physics (collisions, wind) and characters that move more naturally.
Why cover it on a blog about AI images and video? Because Genie 4 often starts from an image. The model page mentions both text and image prompts, and one example in the post uses a reference image as the exact first frame of the scene. In principle, an illustration you generated can become a place you step into.
how it works: Genie draws, Gemini directs
Genie 4 builds on the Gemini family: Gemini Omni grounds the look and behavior of the world in reality, and other Gemini models keep the scene coherent between two instructions.
DeepMind describes a pair. Genie produces the frames and the audio. Gemini acts as the director: it adapts your instructions to the current scene, watches it between requests, and in the agent experiments, Gemini 3.8 Flash breaks a goal into steps and checks they get done. TechCrunch reports that Google calls this "hybrid world models".
What the announcement leaves out: no resolution, no frame rate, no maximum session length. The only number is the consistency window reported by TechCrunch. Be wary of spec sheets quoting 4K or 60 fps for Genie 4: those figures do not come from Google.
On safety, every generated video and sound carries SynthID, Google's invisible watermark, and the model went through internal safety evaluations before the preview.
what the first testers are doing with it
Google's testers mostly use it as a fast draft: previsualizing a film scene, turning an archive photo into a place you can explore, prototyping a game before writing any code.
The most useful part of the post is the tester feedback. On the creative side, filmmakers step into a live simulation of a scene and shoot several "takes", changing camera angles, lighting, dialogue and action as they go. Photojournalists are testing how a single 2D archive photo becomes an explorable environment in seconds. Game developers go from an idea to a playable prototype to try a visual style or materials.
DeepMind engineer Diego Rivas told TechCrunch that real-time interaction is most valuable for rapid ideation. He added that turning those prototypes into usable creative assets with Gemini Omni is "still a ways off". So Genie 4 is for finding an idea, not for delivering a finished shot.
who can try it, and when
As of November 2, 2026, Genie 4 is only open to testers chosen by Google. It is "coming soon" to Google AI Ultra subscribers in Project Genie, and to early access partners in the Gemini API and Google AI Studio.
For everyone else, the only thing you can do today is join the waitlist, through the button on DeepMind's post. Google's blog talks about a limited rollout with no countries or dates. TechCrunch mentions the coming weeks for Project Genie, Google's tool for building interactive worlds from a text or image prompt.
There is no published price for Genie 4 itself and no public API. That means it is not in any third-party model catalog yet, Imaginode included.
the limits Google admits
Google writes that details and object permanence can drift as a session runs longer, after several complex prompt edits, or while moving through the scene.
The limitations section of the post is candid and worth reading in full. Complex physical interactions, real-time audio accuracy, legible text in the scene, spatial accuracy and long-term consistency are all still research problems. In practice, a shop sign may be unreadable, an object can change shape when you come back to it, and after two minutes the world may start to slide.
It is the problem everyone already knows from AI video, only more visible: the more edits you chain, the further the result drifts from where you started. Fine for brainstorming. Not yet for a scene that has to cut cleanly into the next one.
in the meantime: plan your shot in 3D before the video
What testers are looking for in Genie 4, choosing the camera, the light and the movement before producing anything, you can already do today with a 3D previz followed by a video model.
Imaginode's 3D Previz studio starts from the same idea with a different tool. You describe the scene, the AI builds it in 3D in your browser, you place the characters, pick the lens and draw the camera move. Then you export the start frame, the end frame, a reference video and the prompt, and hand them to a video model such as Seedance 2.5. A scene costs from 5 credits, 12 on average, and rendering is free.
It is not a world you walk through live with sound, and I won't pretend it is. But the result is locked: the camera, the path and the timing stay the ones you chose, shot after shot, and you can generate the video right away. The method is in our guides to 3D previz for AI video and to directing the camera. To keep the same face from one shot to the next, the consistent character guide is still the place to start.
what this changes for image makers
Genie 4 can start from an image you made and turn it into a place you walk through, relight and fill with talking characters.
For two years we have treated an AI image as a result. You retouch it, upscale it, animate it into a few seconds of video. Genie 4 offers something else: stepping into the image and looking at what sits outside the frame. For an illustrator or an art director, that is a new way to test a set before drawing it properly.
My advice for now: join the waitlist if this matters to you, keep your best environment images close (they are what you will start from), and keep planning your shots with tools that already produce. That way your first session goes into testing a set you already know by heart, not into learning the tool.
How we analysed this
Written on November 2, 2026 from Google DeepMind's post of October 27 (by Raia Hadsell, Diego Rivas and Hang Qi), Google's own announcement the same day, the Genie model page on deepmind.google, and TechCrunch's October 27 report, which gives the two-minute consistency figure. Screenshots come from DeepMind's post and from the 3D Previz page on imaginode.ai, taken on November 2, 2026.
We have not tested Genie 4: it is only open to a group of testers picked by Google, with no public API. Google gives no resolution, no frame rate and no price. The 3D Previz credit costs are the ones shown on imaginode.ai on November 2.