Genie by DeepMind

Genie is a foundation world model developed by DeepMind that can generate playable, interactive environments from a single image prompt. Trained on a large dataset of publicly available Internet videos, Genie learns to create action-controllable worlds without requiring any explicit action labels. It can be prompted with synthetic images, real-world photographs, and even hand-drawn sketches, enabling users to step into and interact with their imagined virtual worlds. The research primarily focuses on 2D platformer games and robotics, but the method is designed to be domain-agnostic and scalable to larger datasets.

Genie stands out for its ability to infer fine-grained latent actions from video data alone, allowing consistent control across different generated environments. By combining Genie with text-to-image models such as Imagen2, users can first generate an initial image from text and then bring it to life as a playable world. This opens up new possibilities for creators, educators, and researchers to prototype games, explore AI-generated content, and train generalist agents in a never-ending curriculum of novel environments.

As a research project, Genie is not a commercial product but rather a demonstration of a new paradigm for generative AI. The model is presented through a paper and a project website, providing details on its architecture and capabilities. While the underlying technology may eventually be integrated into broader platforms, current access and deployment details should be verified directly from DeepMind's official communications.