Google Lets You Rewrite a Generated World While You Are Still Standing in It
Google released Genie 4 in Project Genie on November 18 for AI Ultra subscribers, adding mid-session prompt editing, real-time native audio and a community gallery, while Genie 3 sessions stretch to three minutes.
Google's most capable world model no longer runs a scene from start to finish. From November 18, Genie 4 accepts a new prompt while a session is already playing, so the weather changes, a character walks in, or the action takes a turn without restarting the world.
The launch post from Google DeepMind and Google Labs puts Genie 4 inside Project Genie, the browser experiment that generates explorable spaces from a sentence or a photo. Access goes to eligible Google AI Ultra subscribers first, with AI Pro members on a waitlist.
The change that will decide whether people use it is not the visual fidelity. It is that the scene stopped being a fixed clip. Until now a generated world was something you could walk through and then had to leave. Mid-session prompting turns it into something closer to a game engine with a conversational remote.
Sound arrives with the edit
Genie 4 pairs its visuals with native audio generated in real time, from ambient background to character voices, and the audio follows the scene as it changes. That is a harder problem than adding a soundtrack in post, because the model has to keep a consistent soundscape while the geometry underneath it shifts.
Google also updated Genie 3 rather than retiring it. Sessions now run up to three minutes per generation, up from 60 seconds, and worlds can be chained so a flight continues from forest to underwater ruin to ice. New worlds can be seeded from 360-degree panoramas, and natural language prompts can trigger in-world events while you move.
An agent that builds the room
The new Project Genie agent takes a description and orchestrates Google's multimodal models to produce the initial look and a custom score, so the entry point is a conversation rather than a single prompt. Gemini models can also sit in the session, watching the screen and offering suggestions for where to go next.
Around that sits a community gallery where worlds are published and remixed, including work from outside creators: a gothic manor mystery built with Will aka BLVCKL!GHT, and an ant colony navigated from the insect's point of view with filmmaker Anand Gandhi.
What the model still cannot do
Google is direct that these remain research world models. Its own list of open problems includes camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence, which in plain terms means a world you leave for a while may not be the same one when you return.
Every generated video and audio track carries SynthID watermarking and passes safety screening, which matters more once users can publish and remix each other's worlds.
The business logic is straightforward: world models are compute hungry, and the feature ships behind Google's top consumer subscription tier. The practical question for anyone outside that tier is whether interactive, editable worlds become a creative tool or stay a demo. Google's move from single-shot generation to a live, promptable space, and the earlier developer opening of World Labs Atlas, both point the same way.
Some offers on this page may be paid placements or contain affiliate links.
Subscribe to Techpresso
Free daily newsletter, read in 5 minutes.
Subscribe free