Sprite AI Studio

Inspiration

I already had a working AI sprite generator on my machine. It used Gemini to generate game art, remove backgrounds, and pack sprite sheets automatically. But it was a desktop app with no web presence. I wanted to see how fast MeDo could flip that into something shareable.

What it does

Sprite AI Studio lets you generate game-ready sprites by describing them in plain English. Type "a glowing health potion for a dark fantasy RPG" and the agent handles the rest: picks the art style, generates the image with Gemini, removes the background, and delivers a clean sprite sheet. Generated images appear in the chat in real time as they come back from the model.

How we built it

The backend is a FastAPI server that wraps an existing Python pipeline powered by Gemini image generation on Vertex AI. It exposes REST endpoints for config and chat, a polling endpoint for real-time event streaming, and static file serving for generated sprites. MeDo built the entire React frontend from a requirements document in one session: the chat panel, style references gallery, jobs board, and settings panel. ngrok tunnels the local server so the MeDo app can reach it from anywhere.

Challenges we ran into

The biggest one was the API surface. MeDo built the frontend before the server existed, so there were mismatches between what the frontend called and what the server exposed. Double-slash URLs from a trailing slash in the base URL setting caused 404s for a while. WebSocket connections over ngrok needed to use wss:// not ws://, which meant replacing the stream with HTTP long-polling to keep things reliable across any tunnel or proxy.

The other challenge: the first time I said "hey" to the agent it started generating 300 sprites because it loaded a saved item list from a previous session and auto-approved the plan. Fixed that immediately.

Accomplishments that we're proud of

MeDo generated a complete, functional 4-panel web app from a single requirements conversation. The style references gallery renders real image previews pulled from the local pipeline. The chat panel streams agent events, tool calls, and generated images live as the pipeline runs. The whole stack — Python backend, React frontend, AI image generation, ngrok tunnel — works end to end.

What we learned

MeDo works best when you write a requirements doc first and hand it the full picture at once. Incremental prompts work fine for fixes but the initial generation is much stronger when the context is complete upfront. The two-step approach (requirements pass, then build pass) saved a lot of back-and-forth.

What's next for Sprite AI Studio

Persistent sessions so the chat history survives a server restart. A gallery view of every generated sprite sortable by style and type. One-click export packages for Unity and Godot. And eventually a hosted version so it doesn't require a local server at all.

Built With

• medo