TechPrufer
Get started
DeepMind
DeepMind logoDeepMind · Feature trail

Interactive World Models

1 post · Aug 2025 – May 2026

What this trail covers

Genie 3 advances interactive world models by generating dynamic, navigable environments at 720p resolution with real-time interaction at 24 FPS, maintaining consistency for several minutes. It introduces 'promptable world events' for environmental interventions and demonstrates compatibility with embodied agents like SIMA for complex goal achievement. Limitations include a constrained agent action space, challenges in multi-agent simulation, imperfect real-world location accuracy, inconsistent text rendering, and limited interaction duration.

All DeepMind trails
  1. Simulate real-world places with Project Genie and Street ViewThis post introduces a new Street View grounding capability for Project Genie, allowing generated worlds to be based on real-world imagery. This leverages Maps Imagery Grounding technology to anchor virtual environments in specific locations, enabling exploration and creative reimagining of real places. The capability is powered by real-world imagery and allows for style selection and character description to influence the generated world.
  2. Project Genie: Experimenting with infinite, interactive worldsThis post details the rollout of Project Genie, an experimental research prototype powered by Genie 3, Nano Banana Pro, and Gemini. It outlines three core capabilities: World Sketching (using text/image prompts and Nano Banana Pro for fine-tuning), World Exploration (real-time path generation based on actions), and World Remixing (building on existing prompts or curated worlds). It also discusses responsible AI development and known limitations of the prototype, such as adherence to real-world physics and character control.
  3. Genie 3: A new frontier for world modelsGenie 3 represents a significant leap in world model capabilities by enabling real-time, interactive environment generation. Key technical breakthroughs include managing auto-regressive generation over time to maintain consistency for several minutes (visual memory up to one minute) and responding to user inputs multiple times per second. The introduction of 'promptable world events' allows for dynamic changes to the generated world, enhancing interactivity and enabling 'what if' scenarios for agent training. The post also details the integration and testing of Genie 3 with the SIMA agent, demonstrating its utility for complex goal achievement in simulated environments.