Convert a photo of your neighborhood into a playable pixel-art game world with AI characters.
Build an interactive story game set at any real-world address using OpenStreetMap building data.
Play as a game master, injecting events and editing character memories to shape how the story unfolds.
Requires API keys for Google Maps, Google Gemini, and one additional service, plus Node.js 18+ and a Python environment.
OpenPixel-RPG is a web application that turns any real-world location into a playable pixel-art role-playing game. You give it an address, upload a photo, or name a landmark, and it generates a small interactive game world at that location complete with AI-driven characters who wander around, talk to each other, and react to events you trigger. The project is in an early alpha state and runs locally on your own machine. There are three ways to use it. In the first mode you upload any photo and the app converts it into a pixel-art game map. In the second mode you type in a New York City landmark name and it pulls pre-made tile data to build a map of that area. In the third mode you type any address on Earth: the app fetches building geometry from OpenStreetMap, captures a satellite-style view from Google 3D Tiles, and feeds both into an AI image model to produce a pixel-art map that matches the real building layout and colors. Once any map is ready, you click a button to populate it with characters and enter the world. The characters in the game are not scripted. Each one runs its own cycle: it perceives nearby entities and recent events, asks an AI model what to do next, carries out that action, and forms a memory of what happened. Over time those memories decay and are periodically summarized by another AI pass. Characters also track an emotional state that changes based on dialogue outcomes and events. You can play as a kind of game master, injecting events or editing a character's memories and traits to change how the story unfolds. All the AI workloads use Google's Gemini model family. Different Gemini models handle world design, image generation, quality review of generated maps, and real-time character behavior. No third-party AI agent framework is used, the perceive-decide-act loop is built from scratch. The project is built with TypeScript, React, and Python. Running it requires Node.js 18 or newer, Python with a few libraries, and API keys for Google Maps, Gemini, and one other service. The code is released under the MIT license.
← tensor2023 on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.