Score-Based Generative Model for Minecraft Game Scene Construction

This study explores a new approach to the application of Artificial Intelligence Generated Content (AIGC) in the field of game scene generation, with a particular focus on the Score-Based Generative Model. The model is utilised in the Minecraft game scene construction task, and its performance is assessed through comparative experiments and Bits Per Dimension (BPD) values. We first introduce the background and challenges of game scene generation, followed by an in-depth discussion of the advantages of the Score-Based Generative Model, including its continuity, noise introduction method, training stability, and scene realism. The experimental results demonstrate the gradual improvement of the model's performance in image restoration as the number of training rounds and the maximum noise parameter are adjusted. In particular, optimising the maximum noise parameter enhances the model when adapting to the complexity of the real scene by balancing the noise introduction, and improves the accuracy and stability of image restoration. We found that, when applying the pre-trained model performed on the CIFAR-10 dataset to the Minecraft dataset, the domain mis-match resulted in high BPD values, revealing the importance of domain-specific training. Overall, this study provides game developers with a high-quality and highly realistic solution for scene generation, which provides a solid foundation for game experience enhancement and game development innovation. Future work can further explore strategies such as domain adaptation to improve the applicability of the model in specific domains for game scene generation tasks.