DeepMind, Google‘s artificial intelligence company, has just unveiled Genie, a new model capable of generating interactive video games from a simple text or image, without the need for any prior training on game mechanics.
According to the official blog post by Google DeepMind, Genie is a “foundational world model” that is trained with internet videos. The model can “generate an infinite variety of playable worlds (controllable through actions) from synthetic images, photographs, and even sketches“.
The paper “Genie: Generative Interactive Environments” states that Genie is the first generative interactive environment that has been trained in an unsupervised manner using unlabeled Internet videos. In terms of size, Genie has 11B parameters and consists of a spatiotemporal video tokenizer, a self-regressive dynamics model, and a simple and scalable latent action model.
These technical specifications allow Genie to operate in frame-by-frame generated environments even in the absence of training, labels, or any other specific domain requirements.
According to the paper, Genie is a new type of generative AI that allows anyone (even the smallest ones) to dream and immerse themselves in worlds generated similar to simulated environments designed by humans. Genie can be asked to generate a series of interactive and controllable environments, although it is trained only with video data.
Genie can receive images it has never seen before, according to Google DeepMind. This includes photographs from the real world and sketches, allowing people to interact with their imagined virtual worlds. It is what is known as a foundational world model.
Regarding training, the paper highlights that the company has focused more on 2D platform and robotic video games. Genie is trained with a general method, which allows it to function in any type of domain, and is scalable to even larger Internet datasets.
