Unbounded: A Generative Infinite Game of Character Life Simulation

Unbounded is a generative infinite game that allows users to interact with a custom wizard character named Archibus in a dynamic virtual world. The game utilizes advanced generative models to create an open-ended life simulation where the character's hunger, energy, and fun levels are influenced by user interactions through natural language. The gameplay is designed to be spontaneous, with the character exploring various environments and engaging in numerous actions and interactions. The development of Unbounded includes innovations in both large language models (LLMs) and visual generation techniques. A specialized LLM generates game mechanics and narratives in real-time, while a dynamic regional image prompt Adapter (IP-Adapter) ensures visual consistency of the character across different settings. The system has been evaluated through qualitative and quantitative analyses, demonstrating improvements in character life simulation, narrative coherence, and visual consistency compared to traditional methods. The game engine is capable of real-time image generation and maintains character consistency while adapting to environmental changes. The research highlights the effectiveness of distilling a specialized LLM for enhanced game simulation performance, achieving results comparable to leading models like GPT-4o.

- Unbounded allows for interactive character life simulation using natural language.

- The game features real-time updates to character metrics based on user input.

- Innovations include a specialized LLM and a dynamic IP-Adapter for visual consistency.

- The system shows significant improvements over traditional game simulation methods.

- The research validates the effectiveness of distilling LLMs for enhanced performance.