Genie 3, a versatile world model, was previewed in August, demonstrating its ability to generate diverse and interactive environments. Early testers created a wide array of engaging worlds and experiences, discovering novel applications for the technology. The next phase involves expanding access through a dedicated, interactive prototype designed for immersive world creation.
Project Genie is now available to Google AI Ultra subscribers in the U.S. (18+). This experimental research prototype enables users to create, explore, and remix their own interactive virtual worlds.
Advancing World Models
A world model simulates environmental dynamics, predicting evolution and the impact of actions. While Google DeepMind has developed agents for specific environments such as Chess or Go, achieving AGI necessitates systems capable of navigating the complexities of the real world.
To address this challenge and further the AGI mission, Genie 3 was developed. Unlike static 3D snapshots, Genie 3 generates the path ahead in real time as a user moves and interacts within the world. It simulates physics and interactions for dynamic environments, and its consistent performance allows for the simulation of various real-world scenarios, from robotics and animation modeling to exploring locations and historical settings.
Based on extensive model research with trusted testers from diverse industries, an experimental research prototype, Project Genie, is now being introduced.
How Project Genie Operates
Project Genie is a prototype web application powered by Genie 3, Nano Banana Pro, and Gemini. It offers users a direct way to experience the immersive capabilities of this world model. The experience focuses on three primary functions:
1. World Sketching
Users can create a dynamic, expanding environment by providing text prompts and generated or uploaded images. This includes designing a character, building the world, and specifying exploration methods like walking, riding, flying, or driving.
For enhanced control, “World Sketching” is integrated with Nano Banana Pro. This feature allows users to preview their world and refine images before entering. Users can also set their character’s perspective, such as first-person or third-person, to control the viewing experience.
2. World Exploration
Each created world is a navigable environment ready for exploration. As a user moves, Project Genie generates the path ahead in real time, adapting to their actions. Camera adjustments are also possible while traversing the world.
3. World Remixing
Existing worlds can be remixed into new interpretations by modifying their original prompts. Users can also browse curated worlds in a gallery, use a randomizer for inspiration, or build upon these existing creations. Completed worlds and explorations can be downloaded as videos.
Responsible Development
Project Genie, an experimental research prototype in Google Labs, is powered by Genie 3. The development of general AI systems prioritizes responsible creation for the benefit of humanity. As an early research model, Genie 3 has several known areas for refinement:
- Generated worlds may not always appear entirely realistic or consistently align with prompts, images, or real-world physics.
- Character control can sometimes be less precise, or users might experience higher latency.
- Generations are currently limited to 60 seconds.
Some Genie 3 model capabilities previously announced in August, such as promptable events that alter the world during exploration, are not yet integrated into this prototype. Further details on model limitations and future improvements are available here.
Based on ongoing work with trusted testers, this prototype is being shared with users of advanced AI to gain insights into how world models will be utilized across AI research and generative media. Project Genie is rolling out to Google AI Ultra subscribers in the U.S. (18+) starting today, with plans for expansion to other regions. The aim is to observe the diverse worlds users create and, eventually, make these experiences and technologies more widely accessible.
