Reproduce “Unbounded: A Generative Infinite Game of Character Life Simulation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jialu Li, Yuanzhen Li, Neal Wadhwa, Yael Pritch, David E. Jacobs, Michael Rubinstein, Mohit Bansal, Nataniel Ruiz

UNBOUNDED introduces a generative infinite game paradigm for character life simulation, employing advanced text-to-image models for real-time visual world creation and a distilled large language model for dynamic narratives and game mechanics. The system achieves real-time interactive speeds, delivering consistent character and environment visuals while enabling open-ended storytelling and emergent gameplay.

Key insights

  • Fine-tuning diffusion models with Latent Consistency Models (LCMs) and DreamBooth LoRAs, then arithmetically merging them, enables personalized characters and rapid inference speeds for visual generation.
  • Dynamically masking cross-attention in a regional IP-Adapter and selectively applying it (block drop) across diffusion blocks is crucial for disentangling character and environment conditioning, preserving visual consistency.
  • A 'teacher-student' distillation approach using powerful LLM agents to generate synthetic interaction data can effectively transfer complex game logic and narrative capabilities to smaller, faster LLMs for real-time interactive applications.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.