Sumi-e Studio

What you’re looking at

Type a scene, in English or Japanese, and Gemini 2.5 Flash Image paints it as a sumi-e ink wash on washi paper. Each painting takes about ten seconds.

Open “What the model was actually asked” under any painting. Your words are only one line of it. The rest is a fixed style brief: loose brushwork, generous empty space, a single vermilion seal, no text. That brief is why a request for “a train platform at night” comes back as an ink painting and not a photograph. Most of what makes AI images look consistent across a site is this kind of unglamorous prompt scaffolding, not the model.

Things to notice

• The seal is gibberish. Look closely at the red hanko in the corner. It’s convincingly carved, and it isn’t real kanji. Image models learn what writing looks like long before they learn what it says. Ask for text and you’ll see the same thing.
• “Empty space” is hard to ask for. Sumi-e is as much about what’s left unpainted as what’s painted. Models trained mostly on busy photos tend to fill the frame; the prompt has to push for emptiness explicitly.
• Modern subjects work surprisingly well. Train platforms, convenience stores and vending machines translate into ink wash more gracefully than you’d expect. The style is really a set of rules about line and space, not a list of approved subjects.

Behind the scenes

This is the same approach I used for the painted backgrounds, fox sprites and title art in Kitsune Run, just with a live prompt box instead of a batch script. The three paintings in the studio gallery were generated through this exact endpoint.

Guardrails

Images are expensive: each one costs several times as much as a text explanation. So each visitor gets 3 paintings a day, requests pass a Cloudflare human check, and the whole lab has a daily spending cap. The model also has its own safety filters and will decline some scenes. Prompts and images aren’t stored by this site.