Google strikes a big blow: Genie 3 generates interactive worlds in real time from a simple prompt

« What if you could not only watch a generated video, but also explore it? It is in these terms that Google Deepmind unveiled on August 5, 2025 a new model of artificial intelligence (AI), baptized Genie 3. Therefore, Its particularity? Meanwhile, “” Create interactive and playable environments from a single prompt text “Describes the company.

This novelty marks a notable progression compared to the Génie 2 model revealed in December 2024. For example, Now Genie 3 is presented by Deepmind as the first model allowing real -time interaction. Therefore, The company indicates that consistency and realism are improved compared to Genie 2.

google strikes big blow: genie google strikes big blow: genie

Google strikes big blow: genie

24 images per second. For example. Similarly, a consistency of several minutes

« Today, we are announcing Genie 3, a versatile world model capable of generating an unprecedented diversity of interactive environments “Enthuses Deepmind, the Google subsidiary specializing in AI, in its press release. Similarly, And to specifically emphasize that everything starts from an ordinary investigation given by the Internet user.

« GENIE 3 can generate dynamic worlds in which you can navigate in real time at 24 images per second. maintaining consistency for a few minutes at a resolution of 720p ». Either of HD Ready, that is to say an image quality just a notch below the high definition (1080p). Each image is created by the model in response to the actions undertaken by the user.

Deepmind Lists several capacities of his new model:

• model worlds with their physical properties (water, light);
• Simulate natural worlds, such as animal or plant ecosystems;
• Create animations and fiction (scenarios, amino characters);
• Explore historic places and contexts, including over.

To reach this level of interactivity. immersion in real time, Deepmind had to fight against the tendency of AI to accumulate imprecisions over time.

It was necessary to preserve the consistency of the environments created as and when. Thus, Genie 3 retains ” a visual memory going up to a minute before ». If the user revisits a place already explored there is a minute. less. the model draws from his memory to remain coherent – all this being obviously possible at the cost of complex calculations.

Google strikes big blow: genie

Google strikes big blow: genie

Genie 3 of Deepmind is “self -regressive”

GENIE 3 is a model ” self -regressive “. According to Shlomi Fruchter, research director at Deepmind, interviewed by Techcrunch. He generates an image at a time. and ” must refer to what has been generated before to decide what will happen next. It is a key element of architecture “, He said. In summary. the model learns by itself how the environment works-for google google strikes big blow: genie strikes big blow: genie example, how its objects move.

In addition to allowing to navigate these environments. GENIE 3 can take into account the user instructions to make the world evolve in real time. whether it is a change in the weather, or by integrating new protagonists in the scenario. There are numerous possibilities for updating and depend on the inventiveness of the prompt, in short.

Deepmind, however, recognizes limits to Genie 3, in particular his ” limited interaction ». « The model can currently support a few minutes of interaction continues, rather than several hours. Another weakness: Genie 3 is unable to simulate real world. As for her understanding of physics, although already impressive, she still remains to be perfect.

When will we be able to test this new model? We do not know. Currently, only a few rare people have access to Genie 3, specifies Deepmind, mainly creators google strikes big blow: genie and academics.