Genie 4 opens a new stage for the simulation of physical worlds

Google DeepMind unveils Genie 4, a model capable of generating interactive photorealistic worlds, complete with sound, dialogue, and real-time physics simulation.

A character plays the saxophone on a floating platform. As the camera pulls back, traversing the lake to reveal the surrounding landscape, the music, movements, and reflections continue to evolve. The scene is neither a pre-rendered video nor an environment built in a traditional game engine. It is generated live by Genie 4, Google DeepMind’s new interactive simulation model.

Evolution is important for world models. Until now, their value relied primarily on the ability to produce an explorable visual environment. With Genie 4, Google adds native sound generation, interactive dialogue, a better representation of physical interactions, and the ability to alter the flow of a scene through simple text prompts.

The system is currently being offered to a select group of testers. Google subsequently plans to provide access within Project Genie for Google AI Ultra subscribers, as well as a progressive rollout to developers via the Gemini API and Google AI Studio. Generated environments rather than static videos

The difference between a video generator and an interactive world model comes down to how the images are produced.

A conventional generated video follows a sequence defined in advance. In Genie 4, the environment evolves in response to camera movements, user inputs, and events occurring within the scene.

The demonstrations published by Google DeepMind feature keyboard-controlled navigation combined with text prompts that allow users to influence events as they unfold.

Users can, notably, alter weather conditions, change their point of view, or introduce new characters and objects. The model then adapts the existing scene rather than starting a generation completely over.

In one demonstration, a concrete platform floats in the middle of a foggy lake. A musician plays the saxophone near a glowing sphere. The user can freely explore the environment while the character's movements, the music, and the elements of the scenery continue to be simulated.

Other examples show frozen landscapes traversed by a hot air balloon, a meadow where a huge ribbon of pink fabric meanders, or environments composed entirely of stained glass.

Photorealism is therefore not the only aesthetic direction offered. Genie 4 can also maintain stylized visual universes during exploration. Gemini Omni for visuals, Gemini Flash for orchestrating events

Genie 4 relies in part on the capabilities of the Gemini family, but the various components do not serve the same role.

According to Google, Gemini Omni helps anchor simulations in a believable representation of real-world appearances and behaviors.

Gemini 3.8 Flash acts as a supervisory agent during generation. It regularly observes the scene, breaks goals down into steps, monitors their progress, and adapts the instructions sent to the world model.

This mode of operation adds a layer of reasoning to a system that is no longer limited to predicting the next frame. When a user modifies the environment or interrupts an event, Gemini can reorganize the steps required to continue the simulation.

The distinction is important: Genie 4 does not necessarily perform all planning on its own. Demonstrations of complex behaviors rely on cooperation between the world generator and a Gemini model tasked with guiding its progression.

Google also presents an experiment pairing Gemini 4 Argon with Genie 4. In the published example, the system breaks down the actions of an Apptronik Apollo humanoid robot playing checkers, then progressively transmits the steps to Genie 4 to render them in the simulation.

This is a research experiment, not a demonstration that the physical robot performed these actions autonomously. Objects fall, characters react, and the scene produces its own sounds

Google claims progress in physics simulation, particularly regarding collisions, aerodynamics, and human movements.

One demonstration shows a maple seed descending to the ground. Another features a watermelon placed on a stone pedestal. These examples aim to illustrate the model's ability to replicate dynamic phenomena beyond the mere appearance of an image.

Character generation is also evolving. Genie 4 can depict individuals moving, interacting with their environment, and reacting to changes caused by the user.

The other major new feature is native audio generation. Environments produce their own sounds without necessarily requiring independent audio processing after the images are generated.

The system can create wind, waves, ambient noise, music, or sounds associated with character movements. It also supports interactive dialogues in which characters react to what the user says or does.

In a first-person experience, a knight rides a dragon at high speed. Wing movements, wind gusts, and environmental sounds are generated simultaneously with the image.

For creators of interactive content, the simultaneous generation of imagery and sound reduces the traditional separation between environmental design, animation, and sound design. Explorable film sets prior to shooting

Google is testing Genie 4 with professionals in creative fields, journalism, education, video games, and engineering.

In the audiovisual sector, the model is envisioned as an interactive previsualization tool. A director could explore a scene before its production, change the camera angle, adjust the lighting, evolve dialogues, or experiment with various character movements.

Unlike a traditional storyboard or an isolated generated image, this approach allows movement inside the environment while the scene continues to live.

Video game developers are studying a similar use case: producing an explorable prototype from a description, testing mechanics or an artistic direction, and modifying the universe without immediately building all its components in a game engine.

For journalism and historical documentation, some testers are experimenting with converting archival photographs into simulated spaces. This practice raises an important distinction between what a historical