Welcome

We present GameCrypto, a generalizable world model that learns from a small-scale dataset of Minecraft game videos. By leveraging the prior knowledge of a pretrained video diffusion model, it can create new games in an open domain.

GameCrypto is a novel framework designed to tackle the challenge of scene generalization in game video generation. Existing methods often struggle with fixed styles and environments, limiting their ability to create diverse and novel games. As shown in the schematic, GameCrypto achieves this by combining the open-domain generative power of pre-trained large video generation models with an action control module learned from a small, high-quality dataset, GF-Minecraft. This decoupling of game style learning and action control, implemented through a multi-phase training strategy, allows the framework to retain its open-domain generalization capabilities while enabling action-controllable video generation.

GameCrypto also has the potential to serve as a generalizable world model, capable of generalizing actions within games and potentially extending to other domains, such as autonomous driving and embodied AI. While these broader applications remain an exciting avenue for future exploration, our framework lays the foundation for such possibilities by demonstrating strong generalization capabilities within the context of game development.