Google DeepMind's Genie 3 Lets Anyone Build a Playable 3D World From Text

Google Labs just gave the world a live look at Project Genie at this year's Google I/O sandbox, and the demo is hard to ignore: pick your characters, describe a scene, and within minutes you're walking around inside a fully interactive world you just conjured from text. No game engine. No assets. No code. The pitch is that the gap between playing a game and designing one has just collapsed to a prompt.

The product is powered by Genie 3, the third generation of Google DeepMind's world model research. Genie, Genie 2, and Genie 3 are world models developed by Google DeepMind that can generate game-like, interactive virtual worlds based on text, images, or sketches. A world model is different from a video generator: instead of producing a fixed clip you watch passively, it simulates an environment that responds to your actions in real time, predicting what happens next frame by frame as you move through it.

From research lab to your browser

The first version of Genie was introduced in 2024, capable of generating two-dimensional interactive environments. Its successor, Genie 2, expanded these capabilities to generate three-dimensional environments. Genie 3 followed with higher-resolution world generations and multiple minutes of visual consistency. DeepMind then released Project Genie to AI Ultra subscribers.

World models are AI systems that can use their understanding of the world to simulate aspects of it, enabling agents to predict both how an environment will evolve and how their actions will affect it. They are also a key stepping stone on the path to AGI, since they make it possible to train AI agents in an unlimited curriculum of rich simulation environments. The gaming use case is almost a side effect of a much bigger research agenda.

How Genie 3 actually works

At its core, Project Genie works by teaching an AI model to understand how worlds behave, rather than by programming specific game rules. The foundation model, Genie 3, was trained on tens of thousands of hours of gameplay videos. By studying these diverse examples, it learned the visual logic and physics that make interactive worlds believable: how characters move, how objects respond to forces, and how environments update with each new frame.

Built as autoregressive latent diffusion models, they enable frame-by-frame interactivity without requiring labeled action data for training. That last part is significant: the model learned what actions do to a world just by watching, not by being told. The platform utilizes Genie 3 as its foundation, which creates internal representations of environments and predicts future outcomes. This model works alongside Nano Banana Pro for image generation and Gemini for natural language processing.

The user-facing workflow breaks down into three modes:

• World Sketching , describe or upload an image to seed the environment
• World Exploration , navigate the generated world in first or third person in real time
• World Remixing , modify existing generated worlds or take someone else's creation as a starting point

Genie 3 is a general purpose world model that can generate an unprecedented diversity of interactive environments. The demos shown at I/O span fantasy forests, ancient Greek ruins, volcanic terrain, underwater canyons, and animated cartoon worlds , all navigable, all generated on the fly.

One technically notable feature is promptable world events: beyond moving through a world, you can issue a text command mid-session to change weather, introduce new objects, or trigger environmental changes. This makes the system useful for generating counterfactual scenarios , essentially asking "what if it started raining?" and having the world respond.

What it's good at , and where it falls short

It can generate three-dimensional environments in 720p, at 24 frames per second. For a model that's synthesizing a navigable world frame by frame in real time, that's a meaningful technical milestone. Genie 3 powers the prototype, generating real-time paths as you move and interact.

The limitations are real and worth knowing upfront:

• Session length: each exploration session caps at around 60 seconds before resetting
• Input latency: characters can sometimes be less controllable, or experience higher latency in control
• Physics accuracy: the prototype has limitations around world realism and character control
• No export: there is currently no way to export generated worlds or integrate them into a game engine
• Access: access is rolling out for Google AI Ultra subscribers in the U.S. (18+)
• Copyright guardrails: TechCrunch noted it was overly censored and not generating content related to mermaids due to a cease and desist from Disney that Google received in December 2025

Fast Company criticized Project Genie for being limited to 24 frames per second and lacking enjoyable gameplay when creating video games, and noted that a future version of Genie could be used for procedural generation or game generation based on player input.

The industry context nobody's ignoring

When Google launched the product, videogame-related stocks like Unity and Take-Two took a dip. Whether that reaction is warranted is debatable , the tool cannot produce commercial-grade games today , but it signals how seriously the market is reading the trajectory of this technology.

Fei-Fei Li's World Labs released its commercial product Marble, while Runway has also launched its own world model. Former Meta chief scientist Yann LeCun's startup AMI Labs has announced its focus on world model development. Waymo has also adopted Genie 3 and created a specialized world model for autonomous driving simulation. The race to own the world model layer is on, and Google is currently the only player with a publicly accessible consumer prototype.

Who should actually pay attention

The honest answer right now is: researchers and early adopters, not production teams. Project Genie is best for rapid prototyping of interactive 2D environments from text or image prompts, great for AI agent training, research experiments, and early-stage creative exploration , but not ideal for commercial game production or simulations requiring precise, physics-accurate control.

The more durable use cases, if the technology matures