Google Unveils Genie 3: Create Playable 3D Interactive Worlds with Just One Sentence

On August 5, Google DeepMind unveiled Genie 3, the latest iteration of its world model system, capable of generating interactive 3D virtual environments in real time based on simple text or image prompts. Unlike traditional video games that require pre-built assets and environments, Genie 3 creates dynamic, playable worlds on the fly—such as a rainy cyberpunk city or a sunlit fantasy forest—simply from a user’s description. This new version marks a major leap forward in real-time interactivity and environmental consistency. Previous versions, like Genie 2 released in December 2024, could generate interactive worlds but only for short durations—around 10 to 20 seconds—and operated at low resolution (360p) without true real-time rendering. In contrast, Genie 3 runs at 720p resolution and delivers smooth 24 frames per second, enabling continuous interaction for several minutes. A key advancement in Genie 3 is its emerging memory capability. In demonstrations, the model maintains visual consistency for about one minute. For example, if a user draws on a wall in a virtual room and later returns, the graffiti remains visible. This persistent memory helps avoid the inconsistencies and “forgetting” issues that plagued earlier models, making the experience far more immersive and coherent. DeepMind researchers note that this long-term consistency is an emergent property—not explicitly programmed—highlighting the model’s ability to simulate richer, more dynamic worlds. The system also introduces “Promptable World Events,” allowing users to alter the environment in real time using new text commands. For instance, in a peaceful ski scene, typing “add a herd of deer” instantly spawns animated deer. Users can change weather, add objects, or even summon absurd characters like a gorilla in a velvet vest—turning the environment into a flexible, creative sandbox. While entertainment and gaming are immediate applications, DeepMind’s broader goal is to advance the development of artificial general intelligence (AGI). World models like Genie 3 are seen as essential building blocks for training AI agents in diverse, simulated environments. The system is already being used to train SIMA (Scalable, Instructable, Multiworld Agent), an AI agent designed to follow natural language instructions across a wide range of virtual worlds. Through repeated exposure to Genie 3-generated environments, SIMA learns to navigate, interact with objects, and adapt to unexpected situations—such as finding a water pipe in a virtual garden—skills critical for future embodied AI systems like autonomous robots in logistics or manufacturing. Despite its progress, Genie 3 still faces limitations. The range of actions an AI agent can perform remains constrained, with many complex interactions requiring text-based prompts rather than direct physical manipulation. Coordinating multiple AI agents in shared environments remains challenging. The model also struggles with precise geographic accuracy, cannot reliably render readable text unless explicitly prompted, and still falls short of supporting long-term, game-like sessions. Currently, Genie 3 is not publicly available. Google is offering a limited research preview to a select group of academics and creators. This controlled rollout aims to allow DeepMind to study potential risks and ethical concerns while collaborating with the research community to ensure responsible development. While Genie 3 is far from achieving the seamless, fully immersive virtual realities seen in science fiction—like the holodeck from Star Trek—it represents a foundational milestone. As the first general-purpose world model capable of real-time, persistent, and interactive simulation, it opens a promising path toward more intelligent, adaptable, and autonomous AI systems.