Generative AI can today generate images, videos, voices, and 3D content in seconds. With Genie 3 Google DeepMind goes one step further: The model does not generate a single image, but an An interactive world that continues to evolve in real time during use.
Genie 3 is a so-called World Model. From a text description – and in Project Genie also from images or even Street-View material – a navigable environment is created. As you move around, the model generates the next views of the world. This very approach makes Genie 3 interesting for games, simulation, AI agents, and in the long term also for virtual reality and XR.
In a nutshell: Genie 3 shows how virtual worlds might in the future be created partly only at the moment of interaction. It does not replace a classic 3D engine like Unity or Unreal Engine, but it significantly shifts the boundary between pre-produced 3D worlds and generative simulation.
What is Google Genie 3?
Google DeepMind describes Genie 3 As a general-purpose World Model, it can generate photorealistic, interactive environments from simple text descriptions. The model not only simulates an image, but also attempts to predict how a world changes and how actions affect that world.
Genie 3 was introduced in August 2025. The most significant improvement over Genie 2 is the significantly better real-time capability: Google cites an output with 24 images per second at 720p. The generated worlds can remain consistent over several minutes and can be continuously navigated during that time.
In this way, the system approaches an interactive, real-time application for the first time. The world is not fully modeled in advance; it is created gradually along with the movement of the user or an AI agent.
What is Project Genie?
With Project Genie Google has turned research into an experimental product. Users can create, explore, and vary their own worlds. The world can be described in text, influenced by reference images, and now even relied on by Google Street View to access real-world locations.
Project Genie is based on Genie 3 and makes the principle of a world model directly tangible for the first time: One describes an environment, defines a character or perspective, and then moves through a world that is continuously generated during navigation.
As of May 2026, Project Genie is available for eligible Google users. Google AI Ultra subscribers worldwide Available. The experimental application is intended for users aged 18 and over.
How does an AI-generated world work?
A classic 3D application works fundamentally differently. In Unity or Unreal Engine, there are objects with positions, meshes, materials, animations, collision bodies, and programmed logic. The engine renders the image from these.
A World Model like Genie 3 works more generatively. Simply put, the process can be imagined as follows:
- A prompt or reference image describes the initial state.
- The model generates the current view of the world.
- The user or an AI agent moves or triggers an action.
- Genie 3 takes into account the current situation and generates the next plausible view.
- This process repeats itself continuously.
This means that a environment does not have to fully exist before it can be entered. New areas can only be created when the user actually reaches them.
That is exactly the idea behind a potentially virtually infinitely expandable AI world.
Why are World Models so interesting?
Image generators essentially answer the question: What could it look like? A World Model also tries to answer: What happens next if I do something?
That is a fundamental difference. For interactive media, robotics, and AI agents, you need not only images but a model of how an environment behaves.
Google DeepMind therefore sees World Models not only as a tool for generative media; they can also provide training environments for AI agents in which they learn movement, planning, problem-solving, and the relationship between action and consequence.
Genie 2 vs. Genie 3: What has changed?
Genie 2 was an important step at the end of 2024 because the model was already able to generate interactive, visually three-dimensional environments from a single starting image. However, the sequences were relatively short and more like research demonstrations than permanently navigable worlds.
| Genie 2 | Genie 3 | |
|---|---|---|
| Introduction | 2024 | 2025 |
| Interaction | steerable generated sequences | Real-time navigation |
| Frame rate | not designed as a real-time product | up to 24 fps |
| Resolution | Research demonstrations | 720p |
| Consistency | short periods | several minutes |
| Access | Research model | experimentally usable via Project Genie |
Genie 2 therefore remains technologically relevant, but is no longer the best starting point for current consideration. Genie 3 shows much more clearly where generative world models are heading.
Is Genie 3 a 3D engine?
No. This is particularly important for classification.
Even though the results look like interactive 3D worlds, Genie 3 currently does not generate a classic editable 3D scene with known meshes, materials, transform values, colliders, and game objects.
With an engine like Unity or Unreal Engine, a developer can move an object exactly to position X, define a collision, or program a machine with specific logic. In a generative world, however, a visually and semantically plausible continuation is created.
For many professional XR projects, this distinction is crucial. Technical training, product visualizations, or industrial applications often require precise measurements, reproducible processes, and reliable interactions. This is exactly where classic real-time 3D engines remain indispensable.
Genie 3, Gaussian Splatting, and classic 3D engines
World Models are often discussed alongside other new 3D technologies, fulfilling very different functions.
| Technology | What is being created? | Strength | Border |
|---|---|---|---|
| Unity / Unreal Engine | Explicit 3D scene | Control, logic, interaction, reproducible results | Content must be created or imported |
| Gaussian Splatting | spatial reconstruction of real places | photorealistic capture of real environments | does not automatically generate any new worlds |
| Genie 3 | Generative interactive simulation | Quick creation of new worlds and variants | No classic editable 3D scene and limited determinism |
At Gaussian Splatting For example, the focus is often on the most authentic reconstruction of an existing environment. Genie 3 pursues the exact opposite: The environment is allowed to develop anew and continue to evolve during use.
Hybrid workflows are therefore particularly exciting. A real environment could be captured using Gaussian Splatting or traditional 3D scanning, while generative models generate variations, events, characters, or additional areas.
What does Genie 3 mean for virtual reality and XR?
For XR, Genie 3 is of less interest because of the current product maturity than because of the paradigm shift that follows.
Virtual worlds could change during use
Today, VR applications are typically built in advance. In the future, certain parts of an environment might not be available until the user needs them. This would make much more variable and more personalized experiences possible.
Training could become much more dynamic
At VR Applications For training and simulation, variation is particularly valuable. Instead of repeating the same fixed programmed sequence, generative models could generate new situations, environmental variations, or behaviors.
A safety training could, for example, simulate different disturbances. A communication training could generate other people, reactions, and situations. A emergency response training could generate rare special cases without having to program every variant manually.
Prototyping could be massively accelerated
One of the most interesting applications in the short term lies in prototyping. Instead of building a world entirely in Blender, Unity, or Unreal first, one could test an idea very early as a navigable generative environment.
This can significantly reduce the effort involved for concept development, previsualization, user experience, and pitching.
Can Genie 3 already create real VR worlds?
Not yet in the sense that we would understand a productive VR application.
To achieve a convincing VR experience, it requires, among other things, stereoscopic rendering, very low latency, stable 6DoF tracking, high frame rates, precise camera positions, and reliable hand or controller interactions.
Genie 3 primarily demonstrates the generative simulation of a navigable visual world. A complete XR runtime with the requirements of today’s headsets is not yet available.
The concept is still highly relevant for XR. As soon as World Models is further combined with explicit 3D geometry, spatial consistency, and real-time engines, completely new production methods could emerge.
Where are the limits of Genie 3?
The world remains probabilistic
A generative environment is not necessarily identical with each iteration. This is creatively interesting but problematic when an iteration must be exactly reproducible.
Several minutes are not a lasting world
The significantly improved consistency of Genie 3 is a major improvement over Genie 2. However, the technology is still far from being able to create a persistent virtual world that remains stable over hours or days.
No complete control over geometry and logic
For professional real-time 3D production, developers need to be able to control objects, rules, states, and interactions specifically. This explicit control is a strength of classic engines.
Resolution and frame rate are not yet sufficient for today's VR
720p at 24 fps is impressive for an experimental generative world. However, much higher resolutions and frame rates would be required for current VR headsets.
Will Genie replace Unity or Unreal Engine?
In the foreseeable future, a hybrid approach is much more plausible than a complete replacement.
Game engines are excellent at processing familiar data accurately and reproducibly. World models, on the other hand, are excellent at creating new content and situations.
The interesting combination could therefore look like this:
- Unity or Unreal Engine manage tracking, interactions, multiplayer, UI, and precise 3D data.
- World Models generate environmental variants, behaviors, events, or visual content.
- Generative AI generates assets and alternative scenarios.
- Reconstruction technologies like Gaussian Splatting provide real-world locations.
Also in Generative AI in 3D and XR production It is already evident that the greatest productivity gains often result from the combination of several specialized technologies.
Is Genie 3 the future of infinite AI-generated 3D worlds?
The concept of „infinite“ is currently even greater than the technical reality. Genie 3 cannot create a permanently stable, limitless world. Nevertheless, the system demonstrates a fundamental change.
Until now, a virtual world had to largely exist before a user could enter it. With World Models, this principle can be reversed: The world is created because the user enters it.
That is probably the real meaning of Genie 3. Not that classical 3D production will suddenly disappear, but that virtual worlds could in the future consist of a mix of fixed and generative components.
For XR, simulation, and interactive media, this is one of the most exciting developments in generative AI. Project Genie makes this future practically visible for the first time – even though the path to fully generative VR worlds is still far away.
FAQ about Google Genie 3
What is Google Genie 3?
Genie 3 is a World Model developed by Google DeepMind. It generates interactive, navigable worlds from text descriptions and continues to generate the environment in real time during use.
What is Project Genie?
Project Genie is Google's experimental access to Genie 3. Users can create, modify, and explore their own worlds using text and images in real time. Starting in 2026, Street View images can also be used as a starting point.
Can one use Genie 3 in Germany?
According to Google, Project Genie is now available worldwide for eligible Google AI Ultra subscribers. Users must be at least 18 years old. The availability of individual features may vary depending on the country and account.
What is the difference between Genie 2 and Genie 3?
Genie 3 is significantly more focused on real-time interaction. Google lists 24 frames per second at 720p and world consistency over several minutes. Genie 2 was already capable of creating interactive environments, but was significantly more limited in its short research demonstrations.
Does Genie generate 3 real 3D models?
Not in the traditional sense. Genie 3 generates a navigable generative simulation, but not a known, fully editable 3D scene with Meshes, materials, and GameObjects like Unity or Unreal Engine.
Can Genie 3 be used for virtual reality?
The concept is very relevant for VR, but Genie 3 itself is not yet a complete VR platform. Headsets require, among other things, higher frame rates, stereoscopic rendering, very low latency, 6DoF tracking, and precise interactions.
Will Genie replace Unity or Unreal Engine?
Currently, hybrid workflows are more likely. Classic engines offer precise control and reproducible logic, while World Models can generate dynamic worlds and variations. Therefore, both approaches complement each other more than one replacing the other in the short term.
Why is Genie 3 interesting for XR?
Genie 3 shows that virtual environments could in the future be created partly only during use. This opens up perspectives for dynamic training, personalized experiences, faster prototypes, and significantly more variable simulations.
Sources and further links
- Google DeepMind: Genie 3
- Google DeepMind: Genie 3 – A new frontier for world models
- Google Labs: Project Genie
- Google: Project Genie with Street View and global availability
Status: September 2026.
