davidpicard/high-fantasy-world-explorer

A retro pixel-art word-guessing game powered by AI image generation.

2

stars

14

commits

Python

primary language

May 21, 2026

updated

README

High Fantasy Word Explorer

A retro pixel-art word-guessing game powered by AI image generation.

Explore AI-generated high-fantasy scenes using novel-view synthesis, and guess the hidden word each scene depicts. Find 10 words to win.

Demo

Demo Video

Screenshots

An enchanted forest shrineA winged creature — first word foundA goblin — 4 words found
Start of a new gameFirst word foundDeep into a run

How to play

  • Scenes are generated by MIRO and rendered with a retro Bayer-dithered pixel aesthetic.
  • Navigate with the camera to explore the scene from different angles, powered by OVIE novel-view synthesis.
  • Press Tab to open the input field, type your guess, press Enter.
  • A correct guess advances to the next scene. A wrong guess costs a life.
  • 3 lives total. Find 10 words to win. Press Enter after game-over to replay with a shuffled word set.

Controls

KeyAction
ZMove forward
SMove backward
← →Yaw (turn left / right)
↑ ↓Pitch (look up / down)
RReset camera to origin
TabOpen / close input field
EnterSubmit guess (or restart after game-over)
BackspaceDelete last character
?Toggle help overlay
ESC / QClose overlay / quit

Requirements

  • Linux, NVIDIA GPU with CUDA
  • Python 3.12 with a PyTorch venv (see install.sh)
  • The OVIE repository cloned locally

Setup

# Clone OVIE alongside this repo (or set OVIE_PATH)
git clone https://github.com/kyutai-labs/ovie ../ovie

# Run the install script (once)
./install.sh

# Launch the game
./run.sh

OVIE_PATH defaults to ../ovie relative to game.py. Override if needed:

OVIE_PATH=/path/to/ovie ./run.sh

Optional arguments

--seed INT      Random seed (default: 3407)
--steps INT     MIRO inference steps (default: 40; lower = faster)
--guidance FLOAT  Guidance scale (default: 7.0)

Example:

./run.sh --steps 50

Model weights

Both models are downloaded automatically from Hugging Face on first run:

  • MIRO: nicolas-dufour/miro
  • OVIE: kyutai/ovie (revision v1.0)

Weights are cached in ~/.cache/huggingface/.

Contributors

davidpicard

14 commits

davidpicard/high-fantasy-world-explorer

A retro pixel-art word-guessing game powered by AI image generation.

2

stars

14

commits

Python

primary language

May 21, 2026

updated

README

High Fantasy Word Explorer

A retro pixel-art word-guessing game powered by AI image generation.

Explore AI-generated high-fantasy scenes using novel-view synthesis, and guess the hidden word each scene depicts. Find 10 words to win.

Demo

Demo Video

Screenshots

An enchanted forest shrineA winged creature — first word foundA goblin — 4 words found
Start of a new gameFirst word foundDeep into a run

How to play

  • Scenes are generated by MIRO and rendered with a retro Bayer-dithered pixel aesthetic.
  • Navigate with the camera to explore the scene from different angles, powered by OVIE novel-view synthesis.
  • Press Tab to open the input field, type your guess, press Enter.
  • A correct guess advances to the next scene. A wrong guess costs a life.
  • 3 lives total. Find 10 words to win. Press Enter after game-over to replay with a shuffled word set.

Controls

KeyAction
ZMove forward
SMove backward
← →Yaw (turn left / right)
↑ ↓Pitch (look up / down)
RReset camera to origin
TabOpen / close input field
EnterSubmit guess (or restart after game-over)
BackspaceDelete last character
?Toggle help overlay
ESC / QClose overlay / quit

Requirements

  • Linux, NVIDIA GPU with CUDA
  • Python 3.12 with a PyTorch venv (see install.sh)
  • The OVIE repository cloned locally

Setup

# Clone OVIE alongside this repo (or set OVIE_PATH)
git clone https://github.com/kyutai-labs/ovie ../ovie

# Run the install script (once)
./install.sh

# Launch the game
./run.sh

OVIE_PATH defaults to ../ovie relative to game.py. Override if needed:

OVIE_PATH=/path/to/ovie ./run.sh

Optional arguments

--seed INT      Random seed (default: 3407)
--steps INT     MIRO inference steps (default: 40; lower = faster)
--guidance FLOAT  Guidance scale (default: 7.0)

Example:

./run.sh --steps 50

Model weights

Both models are downloaded automatically from Hugging Face on first run:

  • MIRO: nicolas-dufour/miro
  • OVIE: kyutai/ovie (revision v1.0)

Weights are cached in ~/.cache/huggingface/.

Contributors

davidpicard

14 commits

Languages

Python

92.7%

Shell

7.3%