ELIT combines text prompting, 3D object generation, and guessing a hidden concept into a fun, fast-paced multiplayer game.
Each round has one Prompter and multiple Guessers. The Prompter describes a hidden concept without prompting it directly, the AI generates a 3D object, and the object is packed and thrown into the arena. The Guessers then race to identify the concept and score points before the round ends.
You receive a hidden concept and cannot prompt it directly. Instead, you create clues by prompting nearby objects for the Guessers. After each prompt, two generated objects appear: one is auto-painted, and one can be painted manually. You choose one, throw it into the arena, and keep building clues until time runs out or everyone guessed correctly.
Guessers explore the city, find the prompted clues, and race to identify the hidden concept. You can slow rivals down with grabbing, throwing, and using power-ups. Each guess costs you points, but you earn points by guessing correctly. So choose wisely when and what to guess.
Prompters earn more points when Guessers identify the concept quickly, so clear prompts are rewarded. Choosing the AI-painted version of a generated object costs the Prompter points. Guessers lose points for each guess attempt, but faster correct guesses earn more. This creates a constant trade-off between guessing early for maximum points and waiting for more clues to avoid penalties.
Faster correct guesses by the team increase the Prompter's reward.
Each guess costs points, but faster correct guesses earn more.
AI-painting and guessing costs points, but helps finding the concept.
With each prompt and generation, object quality decays on purpose: models drift away from the target concept and pick up surface noise. At the same time, score affects player balance: more points make you larger and slower, while fewer points make you smaller and faster. This tames leading players while helping trailing players grow back into the match. That is our interpretation of tamed growth.
We use word embeddings to compare the target concept with each player guess. A glow effect visualizes semantic distance in real time: green means close or correct, while orange to red indicates weaker guesses.
We use Steam for multiplayer, so joining friends is as easy as using your friends list. We also added proximity chat, allowing players to communicate directly in-game. In the background, a university-hosted server runs our 3D object generation pipeline. Prompts are sent to that server and returned as usable in-game objects. The same technology also supports character creation, letting players build highly customized avatars and create even more fun multiplayer moments.