Google DeepMind unveiled its latest world model, Genie 4, in a research preview on Tuesday. This model generates the most realistic and interactive virtual worlds to date, allowing users to input prompts during simulations to instantly alter scenes, while also generating real-time audio that matches the virtual environment, including music, ambient noise, and interactive dialogue.
Compared to the previous generation, Genie 3, which was released over a year ago, the advancements in Genie 4 are evident: the visuals are more realistic, audio has been added, and it supports free editing, including real-time weather changes, camera angle adjustments, or adding new characters and objects to the scene. It can maintain a coherent scene for up to 2 minutes, double the previous limit, with more accurate physical representations and character movements that closely resemble real-life actions.
Significant Advances in Virtual World Generation with Genie 4
The integration of Gemini with Genie 4 is more profound. Throughout the entire process, Gemini acts as the "director," responsible for planning steps and frame-by-frame checking of visuals, while Genie is tasked with generating visual content. In a research experiment linked to Gemini Robotics, the Gemini 4 Argon model breaks down tasks into steps, which are then simulated in real-time by Genie 4, demonstrating a feasible path: robots complete validation in a simulated environment before acting in the real world.
Google's research director, Tejas Kulkarni, stated during a briefing, "In the future, we will increasingly see these types of hybrid world models—Gemini is responsible for writing code and text, while Genie receives and simulates reality in real-time at different levels of abstraction." Similar to its peers in world modeling, Genie 4 targets commercial scenarios including gaming, interactive learning, and robotics. Foreign media cited an example of interactive learning: inputting a historical scene painting into the system and then directly conversing with the characters in the painting to understand the events behind the image.
Meanwhile, the "trusted testers" who were among the first to experience the model pointed out another potential use—pre-production prototyping for film and television. Google DeepMind engineer Diego Rivas explained, "The most valuable aspect of real-time capability lies in those segments that require maintaining a coherent creative flow, such as the brainstorming phase and storyboarding phase—any moment that requires rapid concept generation." He also noted that Google's high-definition video generation model, Veo, is becoming increasingly popular among filmmakers, and in the future, users may be able to leverage the multimodal large model Gemini Omni, which can combine images, audio, and text to generate videos, transforming these prototypes into creative materials, although such capabilities are still some distance from realization.