Hugging Face Open Source AI Game Jam Results

Hugging Face hosted its first Open Source AI Game Jam from July 7th to July 11th, 2023, challenging developers to create games using at least one open-source AI tool. The event resulted in 88 game submissions from over 1,300 signups, demonstrating the versatility of open-source AI in rapid game development.

Game Jam Framework and Evaluation

The competition was centered around the theme "EXPANDING," which developers were encouraged to interpret freely. While proprietary AI tools were permitted, the primary objective was the integration of open-source AI tools into the game's creation or workflow.

Games were evaluated based on three primary criteria:

  • Fun
  • Creativity
  • Adherence to the theme

Peer and contributor evaluations narrowed the field to a top 10, from which a final winner was selected by a jury consisting of Dylan Ebert, Thomas Simonini, and Omar Sanseviero.

Winning Entry: Snip It

The first-place winner of the Open Source AI Game Jam is Snip It, developed by ohmlet (Code: Ruben Gres; AI assets: Philippe Saade; Music/SFX: Matthieu Deloffre). In this game, players visit a museum where paintings come to life and must snip objects within the paintings to uncover secrets. The project utilized Stable Diffusion to generate its visual assets.

Top Finalists and AI Implementations

Eleven games were recognized as top finalists, showcasing a wide array of AI applications ranging from generative art to reinforcement learning:

Generative Art and Assets

  • Yabbit Attack: Used Stable Diffusion for background visuals.
  • Expanding Universe: Utilized the Dream Textures Blender add-on (powered by Stable Diffusion) for planet and star textures, and an LLM for descriptions.
  • Hexagon Tactics: The Expanding Arena: Integrated Stable Diffusion 1.5 for character generation in the executable version.
  • Apocalypse Expansion: Employed Stable Diffusion for images and MusicGen (melody 1.5B) for the soundtrack.
  • Singularity: Used Stable Diffusion for its conceptual demonstration of future game generation.

Gameplay Mechanics and Logic

  • Yabbit Attack: Implemented genetic algorithms based on natural selection and evolution to create an adapting neural network for enemies.
  • Word Conquest: Used the all-MiniLM-L6-v2 model and GloVe embeddings to generate the map based on word unrelatedness.
  • Galactic Domination: Utilized Unity ML-Agents for reinforcement learning-based bot AI.

Coding and Audio

  • Everchanging Quest: Used Starcoder (open source) for coding, alongside GPT-4 (proprietary) for tile and object placement.
  • Fish Dang Bot Rolling Land: Integrated a Text To Speech (TTS) model to generate character voices.
  • Galactic Domination: Used Elevenlabs for voice, Soundful for music, and Scenario for images.

Resources for AI Game Development

Hugging Face provided several technical resources to support the integration of AI into game engines:

  • Unity Hugging Face API: Guidance on installation and integration.
  • AI Speech Recognition: Tutorials for implementing ASR in Unity.
  • Transformers.js: Resources for creating ML-powered web games.
  • Awesome AI Tools for Game Dev: A curated compilation of tools available on GitHub.

Sources