Google Unveils Genie 4, a World Model That Generates 3D Environments From Text

Google LLC has unveiled Genie 4, an artificial intelligence system designed to synthesize lifelike three-dimensional environments. The model represents the latest evolution in Google's lineup of world models, which debuted throughout 2024. Genie 4 demonstrates substantial improvements over earlier versions, particularly in simulating smooth motion and intricate visual details. The system also possesses the capability to produce matching audio elements, such as ambient sounds like waves breaking on a shore.

"The most dramatic leap is how Genie 4 simulates people, creating characters that move naturally, respond organically to their environments, and can be interactively steered in real-time," wrote Raia Hadsell, Diego Rivas and Hang Qi from Google DeepMind in a blog post.

Creating and Editing Virtual Worlds

Genie 4 allows creators to construct virtual environments by inputting text-based descriptions of desired scenes. Once generated, users gain the ability to explore the simulation and refine it through additional text commands. A user might instruct Genie 4 to modify atmospheric conditions, introduce new elements, or adjust other aspects of the scene. The system performs these modifications while preserving the photorealistic quality and internal consistency of the generated world.

The model offers both natural language prompting and keyboard-based controls for camera manipulation. Users can adjust camera position and viewing angle through keyboard input.

Technical Architecture

Genie 4 integrates multiple Google AI systems to enhance output fidelity. Gemini Omni, a video generation model released by Google in May, handles the creation of visually convincing assets. Gemini 3.8 Flash, a cost-optimized reasoning model, maintains frame-to-frame coherence throughout simulations.

"Gemini continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded," Hadsell, Rivas and Qi explained. "As the world shifts or the user intervenes, Gemini adapts its plan in the moment."

Robotics Applications

Google DeepMind is exploring robotics applications for Genie 4, pairing it with Gemini 4 Argon, Google's primary reasoning-focused large language model. This combination aims to enable engineers to prototype robotic behaviors through high-level textual instructions. For instance, an engineer could request that Argon generate synthetic video depicting a robotic arm performing precision soldering on automotive components. Argon translates this single request into numerous detailed, multi-step prompts tailored for Genie 4, each describing a distinct phase of the soldering process.

Expanding the quantity and specificity of prompts fed into Genie 4 enhances rendering fidelity. This improvement in visual quality strengthens the utility of robot simulations for training machine learning systems. AI engineers develop robot control models by exposing them to visual training data showing successful task completion.

Availability

Genie 4 is currently undergoing testing with a select group of early adopters. Google plans to roll out broader access through three channels: the Google AI Ultra subscription tier for individual users, the Gemini API for developers, and Google AI Studio.