Type “a rainy night market in Tokyo,” and Google’s Genie 4 doesn’t just make a video of it. It builds the market as a place you can walk through with your keyboard, complete with rain sounds, music and vendors who react when you pass.
Tell it to make the rain stop or add someone to the scene, and the world changes while you’re still standing in it.
Google DeepMind announced the model Tuesday and calls it a “world model,” which is the industry term for AI that simulates a space instead of rendering a fixed clip. The difference matters. AI video tools like Veo and Sora give you something to watch, while Genie 4 gives you somewhere to be.
According to Google, objects follow basic physics like collisions and wind, and you can “set an event in motion” and watch the model play out what happens next.
It’s also the first real-time world model from Google that generates its own audio, including music, ambient sound and dialogue. One of the use cases Google lists will get the most attention outside tech circles, and that’s turning archival photos into simulations you can step inside.
Google pitched it at game developers prototyping ideas, filmmakers blocking out scenes before a shoot, and researchers training robots and AI agents in simulated worlds. For creators, it points to a format that doesn’t exist yet, somewhere between a video and a video game, where the audience moves through the scene instead of watching it.
Google also published the catches. Details “can drift” over longer sessions, so a building might not look the same when you walk back to it. Text in the scene often isn’t readable, and complex physics still trips it up. Everything Genie 4 generates carries Google’s invisible SynthID watermark.
ADVERTISEMENT
For now, only a small group of trusted testers can use it. Google says it’s coming next to AI Ultra subscribers through its Project Genie app and to developers through the Gemini API, with a waitlist open now. It didn’t give a date.
