On a Tuesday in late October, Google DeepMind offered a first look at Genie 4, the newest step in a line of world models the company began in 2024. The system conjures three-dimensional spaces from plain language that a person can walk through, reshape with a sentence, and hear, since it now generates sound as well as sight. Its larger purpose is quieter than the demonstration: it hints at machines that can rehearse the physical world before acting in it, a possibility still held in a limited circle of testers.
Google Previews Genie 4, a World Model That Generates Interactive 3D Environments With Audio
Dozens of detailed prompts, each describing a different step of the workflow
Mark
So what is Genie 4, in plain terms?
Mimi
A model that turns a text description into a 3D world you can move around in, now with sound, like waves crashing.
Mark
What changed from the last version?
Mimi
Smoother motion, finer detail, audio, and people who move more naturally and can be steered in real time.
Luke
Those are Google's own descriptions, though. The authors are DeepMind staff. I haven't seen independent testing.
Mark
Fair. How does it keep the world from falling apart when you edit it?
Mimi
Other Gemini models help. Gemini Omni handles realism, and Gemini 3.8 Flash keeps frames consistent with each other.
Luke
The claim is that edits don't hurt photorealism or consistency. That's attributed, not demonstrated in what we've been given.
Mark
And the robotics angle?
Mimi
Argon turns a short request into dozens of detailed prompts, and Genie 4 renders each step, producing synthetic training footage.
Luke
Note that's under study. Nothing says it works well enough yet, and no release dates or prices have been given.
Le Pouls
- Genie 4 turns a written description into a living 3D space, and the improvement that stands out most is how convincingly its characters move and respond.
- Users can change a world mid-walk by altering the weather or adding objects, and the company claims the edits leave the scene's realism and consistency intact.
- Gemini models act as a quiet supervisor, watching the scene, breaking goals into steps, and revising the plan when a user intervenes or the world shifts.
- DeepMind is pushing the model toward robotics, where a single short request could be expanded into dozens of detailed prompts to generate synthetic training footage.
- Access remains restricted to early adopters, with planned routes through Google AI Ultra, the Gemini API and AI Studio, but no dates or pricing have been announced.
On a Tuesday in late October, Google DeepMind offered a first look at Genie 4, the newest step in a line of world models the company began in 2024. The system conjures three-dimensional spaces from plain language that a person can walk through, reshape with a sentence, and hear, since it now generates sound as well as sight. Its larger purpose is quieter than the demonstration: it hints at machines that can rehearse the physical world before acting in it, a possibility still held in a limited circle of testers.
Google has previewed Genie 4, a model that builds realistic three-dimensional environments from plain-language descriptions, complete with generated sound such as crashing waves. It is the latest in a series of world models the company began in 2024, and this version makes its biggest gains in motion and detail. Audio is new, and the authors of the announcement say the largest leap is in how people are simulated: characters move naturally, respond to their surroundings, and can be steered in real time.
A user begins with a description of the world they want, then walks through it and reshapes it with follow-up prompts or keyboard controls for the camera. Google says these changes do not weaken the photorealism or consistency of the scene. Behind the experience, other Google models do the work. Gemini Omni helps keep visual assets realistic, Gemini 3.8 Flash keeps frames coherent, and Gemini watches the scene, breaks goals into steps, and adjusts its plan when the world changes or a user steps in.
DeepMind is also exploring robotics. Paired with the flagship reasoning model Gemini 4 Argon, Genie 4 could turn a simple request, such as a robotic arm soldering two car parts, into many detailed prompts for each stage of the task. Richer prompts yield better renderings, and realistic synthetic footage could help teach robots the tasks they must perform.
For now, Genie 4 remains in limited testing. Google plans wider access through its AI Ultra subscription, the Gemini API and AI Studio, but it has not said when that will happen, and the robotics work remains a study rather than a product.
Google on Tuesday previewed Genie 4, an AI model that builds realistic three-dimensional environments a user can move through, complete with generated sound such as crashing waves. It is the newest entry in a line of world models the company started in 2024.
The improvements over the previous version are concentrated in motion and detail. Genie 4 renders fluid movement and other complex features better than its predecessor, and it adds audio, which the earlier model did not produce. The three Google DeepMind staffers who wrote the announcement post, Raia Hadsell, Diego Rivas and Hang Qi, said the biggest leap is in how the model simulates people: characters that move naturally, respond organically to their surroundings, and can be steered interactively in real time.
The workflow starts with a plain-language description of the world a user wants. Once the environment exists, the user can walk through it and change it with follow-up prompts, such as altering the weather or adding objects. According to the report, those edits do not degrade the world's photorealism or consistency. Natural-language prompting sits alongside keyboard controls, which handle the camera's position and angle.
Genie 4 leans on other Google models. Gemini Omni, a video generator the company introduced in May, helps keep the visual assets realistic. Gemini 3.8 Flash, a reasoning model tuned for cost-efficiency, keeps individual frames consistent with one another. The authors describe Gemini as continuously observing the scene, breaking goals into steps, and checking incoming frames to confirm each step worked. When the world shifts or the user intervenes, they wrote, Gemini adapts its plan in the moment.
DeepMind is also studying Genie 4 for robotics. That project pairs the model with Gemini 4 Argon, Google's flagship reasoning-optimized large language model. The aim is to let engineers simulate robots from high-level instructions. A user could ask Argon for synthetic footage of a robotic arm soldering two car parts together, and Argon would expand that short request into dozens of detailed, multi-sentence prompts for Genie 4, each describing a different step of the soldering process.
The detail matters because more numerous and more precise prompts improve Genie 4's rendering quality, and better rendering makes the robot simulations more useful for training. Developers teach AI models to control robots by showing them visual examples of the tasks they will be expected to perform, so realistic synthetic footage could serve as that material.
Genie 4 is not yet widely available. Google is testing it with a limited group of early adopters. The company plans to offer it through its Google AI Ultra consumer subscription, the Gemini API and Google AI Studio. No dates or pricing were given in the report, so the next thing to watch is when that access opens and whether the robotics work moves beyond study.
Citations marquantes
The most dramatic leap is in how Genie 4 simulates people, with characters that move naturally, respond organically to their environments, and can be steered in real time.— Raia Hadsell, Diego Rivas and Hang Qi, Google DeepMind (paraphrased)
Envie de l'histoire complète ?
Lire l'original sur SiliconANGLE ↗
Technology