DeepMind has introduced Genie 2, a model that can create games based on just a single static image. This model is capable of controlling the game using a keyboard and mouse instantly.
Previously, DeepMind had unveiled the GameNGen model for game creation, and the original Genie model was limited in its game creation capabilities.
What sets Genie 2 apart is its ability to recognize the main character that needs to be controlled when given an input image. For example, if there is a robot in a forest, Genie 2 will control the robot instead of the trees. The model can remember the environment it has created for approximately one minute and can simulate physics such as exploding balloons, smoke, or gravitational forces.
Training Genie 2 involves a large number of videos, and during model runtime, developers can see a playback of past video frames along with the current control input.
This approach makes prototyping games much easier, as developers only need a single static image. However, the model comes with limitations. Creating high-resolution images showcased by DeepMind takes a considerable amount of time, and actual gameplay experiments must run on low-resolution image models. Additionally, sometimes the model creates games that are not coherent, like a snowboarding game where the character runs instead of snowboarding.
TLDR: DeepMind introduces Genie 2, a model that can create games from a single image input. The model recognizes the main character to be controlled and can simulate in-game physics. However, creating high-resolution images is time-consuming, and the model may create inconsistent games at times.
Leave a Comment