Moonshine: Distilling Game Content Generators into Steerable Generative Models
Aug 2024♦ Synthetic Dataset Creation: Generated a large dataset using a constructive PCG algorithm, with content labeled by a Large Language Model (LLM) to facilitate text-conditioned generation.
♦ Model Training: Conditioned two PCGML models—a diffusion model and the five-dollar model—on the synthetic dataset to enable content-specific generation.
♦ Text-to-Game-Map (T2M) Task: Introduced T2M as an alternative to text-to-image tasks, allowing for the generation of game maps based on textual descriptions.
♦ Evaluation Metrics: Assessed the models' performance by comparing variety, accuracy, and quality against the baseline constructive algorithm, demonstrating the effectiveness of the distillation process.