Google DeepMind researchers have introduced Genie, an innovative artificial intelligence model designed to generate playable 2D platformer video games based on various prompts. This breakthrough represents a significant leap in AI-powered content creation during its research phase.
Unlike previous models focusing solely on world generation, Genie takes a step further by allowing users to interact with the environments it creates. The model achieved this capability through training on an extensive dataset of over 200,000 hours of publicly available 2D platformer gameplay videos. Genie’s unsupervised learning process enables it to comprehend character movement, control, and in-game actions.
Accepting inputs in the form of text descriptions, sketches, or existing images, Genie leverages this information to generate game assets and predict future frames based on user actions. This results in a functional and playable gaming experience, showcasing the potential of AI in creating interactive environments.
While Genie is not currently available for public use, the underlying technology and its unsupervised learning approach hold promise for diverse applications. This includes further advancements in AI-powered content generation and the potential development of more interactive and engaging user experiences.
The introduction of Genie follows OpenAI‘s recent revelation of Sora, an AI video generator boasting super-realism. While acknowledging that Genie may not match Sora‘s quality, the complexity of game development compared to video creation is highlighted. Microsoft has confirmed that Sora will be incorporated into Copilot.
Tim Rocktäschel, leading the research team behind Genie, emphasizes the significance of the underlying technology and the novel approaches developed during its creation. These advancements contribute significantly to AI research, laying the foundation for future innovations in the field.