正文 · AI 翻译
IT Home reported on October 28 that Google's DeepMind team published a blog post yesterday (October 27) announcing Genie 4, a world model it described as a “major research milestone.” It delivers key advances in native audio, real-time changes through natural-language instructions, and more realistic physics and characters, upgrading world models from one-off visual generation to dynamic simulated worlds that can be controlled, heard and reasoned about in real time.
For users, the biggest highlight of Google's Genie 4 world model is the ability to “direct” the world in real time. While a simulation is running, users can change the world with natural-language text instructions, such as changing the weather, moving the camera, or adding new characters or objects.
Unlike mainstream AI systems that regenerate a video clip, Genie 4 makes dynamic changes based on the context of the current scene, making it more like an “interactive virtual sandbox” or an “AI directing system.” Users can set an event and watch how the world continues to evolve, intervening to make adjustments at any time.
Among its other highlights, Genie 4 is DeepMind's first model to generate audio natively in a real-time world simulation, with audio and visuals produced in sync as part of the same process. Google also says Genie 4 performs better in physical scenarios involving aerodynamics and collisions; character movements are more natural, characters can react organically to their surroundings, and real-time control is supported. IT Home includes the relevant demonstration videos below:
Google CEO Sundar Pichai posted on his personal account today, saying that Genie 4 is Google's most realistic and interactive world model to date.
In an interview with technology outlet TechCrunch, Google research director Tejas Kulkarni said that, in the future, Gemini would handle text, code and task planning, while Genie 4 would turn that content into interactive worlds in real time, forming a “hybrid world model.”
DeepMind engineer Diego Rivas acknowledged that Genie 4 is still at an early stage; details may drift during long simulations, and audio accuracy and the readability of text within the visuals have yet to be fully resolved.
Genie 4 is currently available only as a research preview to “trusted testers”; in the coming weeks, access will expand to Google AI Ultra subscribers through Project Genie, as well as early partners using the Gemini API and Google AI Studio.
Related reading:
“Google launches its most powerful world model, Genie 4: Create an interactive virtual world with a single sentence, with synchronized audio and video generation achieved for the first time”