Genie 4 World Model: Build a World, Then Steer It Live
Text in, world out — with sound, physics, and people that respond.
Genie 4 is Google DeepMind's real-time world model. Describe a scene and it builds an explorable world frame by frame while you steer it in plain language: move the camera, change the weather, add a person. Native audio adds ambience, music, and character dialogue as the world runs.
Sexually explicit (NSFW) content is strictly prohibited.
DEMOS
Watch Genie 4 Generate a World in Real Time
Three unedited clips: Google DeepMind’s official reel, an early-access walkthrough, and a scene steered live from a keyboard. Each one is played from its source resolution.
The official reel
Google DeepMindTwo and a half minutes of worlds, people and sound, generated frame by frame.
Early access: one prompt, one world
@venturetwinsA world built from a single starting prompt, then changed while it runs.
Walking through a generated world
@midudevMovement, physics and sound generated live as the camera moves.
Clips by Google DeepMind, @venturetwins and @midudev, taken from their public posts and hosted here for reference. Genie 4 is a trademark of Google LLC; this site is not affiliated with Google.
What Genie 4 Can Do
- Prompt a world, then walk into itType a scene and explore it frame by frame, in any direction.
- Change weather, camera, or characters liveSay what to change; the scene keeps running and evolving.
- Native audio, including interactive dialogueAmbience, music, and characters who answer as the world runs.
- Stronger physics and human motionAir, collisions, and movement that hold together.
- Set the first frame with a reference imageUse a still or an opening frame as the visual baseline.
- Draft shots fast for pre-visBlock a sequence or a storyboard before committing to a shoot.
What Is Genie 4?
Genie 4 is a world model, not a text-to-video model. Ask a video generator for a shot and you get a finished clip; ask Genie 4 for a scene and it builds a world you can move through while it is still being generated. Google DeepMind calls it its most realistic and interactive world model yet, and announced it on 27 October 2026.
The difference is what happens after the prompt. You are inside the world rather than watching a clip: change the weather, move the camera, or drop in a person in plain language, and the world continues from that new state instead of restarting. Gemini runs planning behind the scenes to keep the causal chain intact, so the wet street stays wet and the person you added stays in the scene — and characters can talk back while the world model keeps running.
Sound is generated with the picture: ambience, music, and interactive dialogue, so a character answers you in the same pass that draws the scene. Physics covers airflow and collisions, and human motion reads more naturally than in earlier world models. Google has not opened Genie 4 to the public: the world model goes first to trusted testers and early access partners, and every output carries a SynthID watermark.
HOW A SESSION RUNS
Describe a Place, Step In, Then Change It
Genie 4 is built for interaction rather than one render. A session is a loop: describe, explore, edit, and keep going while the world holds together.
Describe the world
A place, a time of day, a mood. Genie 4 builds the scene frame by frame instead of returning one fixed clip, which is what makes it explorable rather than playable once.
Step inside
Move through the scene with WASD. The world keeps generating what is behind you, to the sides, and ahead, rather than looping a finished shot.
Change it while it runs
Say “make it rain”, “move the camera left”, or “add a person”. The edit lands in the running world and the scene continues from there.
Hear the result
Ambience, music, and dialogue come with the scene. A reference image can set the opening frame, and every output carries a SynthID watermark.
FEATURES
What Makes Genie 4 a World Model, Not a Video Generator
A text-to-video model returns one finished shot. Genie 4 returns a place that keeps running while you change it. The difference shows up in how you work, not only in how the output looks.
The scene keeps going after the prompt ends
You are not waiting on a render. The world is drawn around you as you move, and the next thing you say changes what happens next.
ON THE RECORD
What Google Has Confirmed About Genie 4
No user counts and no benchmark numbers — only what Google DeepMind has published about the model, and what it has not said yet.
The year Google DeepMind announced Genie 4, on 27 October, as a real-time interactive world model.
Native in Genie 4: ambience, music, and interactive dialogue generated while the world runs.
Google describes a session as “a few minutes of continuous interaction”; longer runs drift.
USE CASES
Where a Steerable World Helps
A world model is not a finished-shot tool. These are the jobs where walking through a scene and changing it mid-run does something a rendered clip cannot.
Pre-visualisation
Block a camera move or a chase inside a world you can walk through before anyone books a location.
Moving storyboards
Sketch a beat as a place, step inside it, and see whether the sequence works before drawing the rest.
Level and environment ideas
Walk around a rough location and check whether the space reads before anyone models it.
Exploratory explainers
Let a class move through a concept — a cell, a city, a weather system — instead of watching a clip about it.
Sound and dialogue tests
Hear ambience, music, and a character reply while the scene is still changing around you.
Set and lighting tests
Try weather, time of day, and blocking in the running scene instead of rebuilding a render each time.
World-building drafts
Write a place into existence, walk around it, and keep the parts that read as real.
Look development
Change the mood of a location mid-run and watch the look hold as the camera moves.
Genie 4 vs Genie 3
Genie 3, released in August 2025, generated interactive worlds without audio. Genie 4 adds sound and live direction. Here is only what Google has confirmed.
- Native audio: ambience, music, and interactive dialogue generated with the world
- Live edits in plain language while the scene keeps running
- Stronger physics and more believable human motion
- Gemini planning in the background keeps the causal chain intact
- No public API and no published price
- Interaction runs for a few minutes, and visual detail drifts
- Audio accuracy is still being improved
Describe it, then change it mid-scene
The workbench takes a prompt, a reference image, and a few settings, then returns a clip. Genie 4 has no public API yet, so the workbench runs the video models this site supports today.
FAQs
Frequently Asked Questions
What is Genie 4?
Genie 4 is Google DeepMind’s real-time interactive world model, announced on 27 October 2026 and described by Google as its most realistic and interactive world model yet. Instead of returning one fixed clip from a prompt, it builds an explorable world frame by frame and lets you steer it while it runs.
How is Genie 4 different from Genie 3?
Genie 3, from August 2025, generated interactive worlds with no audio. Genie 4 adds native audio — ambience, music, and interactive dialogue — and live edits in plain language, so you can change the weather, move the camera, or add a character while the world keeps running. Google has not published Genie 4’s resolution or frame rate.
How long can one session last?
Google describes a session as “a few minutes of continuous interaction”. Longer runs are where the known drift shows up: visual detail and object permanence degrade over time, which is one of the limitations Google lists.
Is there a public Genie 4 API or price?
No. There is no public API and no published pricing. Genie 4 is not generally available: Google says it is going first to trusted testers, to Project Genie for Google AI Ultra subscribers, and to early access partners through the Gemini API and AI Studio. There is a waitlist.
Can I try Genie 4 for free, and is it online anywhere?
Not yet. There is no free public version and no online demo anyone can open: Genie 4 is a research preview, so access is limited to trusted testers, Google AI Ultra subscribers inside Project Genie, and early access partners on the Gemini API. The waitlist is the only public route in today.
What are the known limitations?
Google lists five: visual detail and object permanence drift; complex multi-agent and shared-environment interaction is still hard; real places cannot be reproduced exactly; readable text usually appears only when the prompt describes it; and interaction length is limited while audio accuracy is still improving. Every output also carries a SynthID watermark.