Each word is converted to a number, and that number lands on a specific Wikimedia file. The descriptions get pulled apart and rebuilt into a new prompt.
letters become numbers, numbers become files
Assembled Prompt
How the numbers work: each word's own letters (a=1…z=26, weighted by position within the word) produce a number, reduced mod 900000. That number is spread across Wikimedia Commons's page-ID space to compute a starting ID, and the app scans forward from that exact ID — skipping camera-default titles like IMG_0042 or Untitled — until it lands on a real file. Same word, same number, same starting point, same file, every time. Wikimedia's API has no "seed" parameter, so this direct page-ID targeting is the closest real mechanism for consistency; it holds as long as that specific file isn't later deleted or renamed on Commons.
Descriptions get cleared of cataloguing jargon (collection, digitized, copyright, accession, and so on) so what's left reads as visual content, not a museum label. Non-English descriptions are handled by checking for Commons's own English-tagged block first, since a lot of files already have one; if there isn't one, it's run through a translator. The final line skips colons and dashes on purpose — that punctuation reads as a caption, which is part of what pushes image and video models toward rendering the words as on-screen text instead of a scene.