Making of

This site is built with Astro and plain CSS. No UI framework, no cookies, and nothing loaded from anyone else’s servers. The site counts visits without cookies: no addresses are stored, only daily totals. It works with JavaScript turned off, and if your system asks for less motion, it holds still.

The first screen is a photograph I took one night in September, with my camera on a tripod in front of Landskron castle in Leymen. There are two layers from that same spot: a short exposure, where you mostly see clouds, and the starry one, which is around 200 ten-second frames at f/2.8 and ISO 800 on a Fujifilm X-T1, stacked into a single photograph. Moving the pointer stirs a small fluid simulation that wipes the clouds away and lets the stars through.

A camera on a small tripod on a gravel path at night, in front of the dark silhouette of Landskron castle. — The whole studio, the night I shot the first screen.

Every project on the homepage shows the real thing. The before and after on the LocalSR card is a piece of that same photograph, shrunk to a quarter of its size and blown back up: a plain enlargement on the left, LocalSR’s HAT-L model on the right. The painting behind the photograph came out of a model that turns pictures into anime, and it was enlarged with LocalSR too. The piano clip is my own model’s music. The train-tui window is the program’s real screen, redrawn in your browser every second on a made-up training run.

The type is Inter (SIL Open Font License) and the icons in the corner are Tabler Icons (MIT). The photographs are mine. The code was written with the help of Claude Code.

The GPT-2 piano sample is played on the Salamander Grand Piano by Alexander Holm, CC BY 3.0.