Moodboards

Lightdrift is a vibe machine. Type a feeling and it searches over 80 million real photographs by what they feel like, not what they are tagged with, and every result comes back with its usage rights attached. That claim is easy to write and easy to doubt, so this post is just evidence: three moodboards, made by one designer in one evening, with the exact queries and the clock time for each. None took longer than six minutes.

Board one: orange and black

The brief was six words: orange and black, hazy, fast, epic, mysterious. The queries talked to the corpus the way you would talk to a person:

hazy orange sports car at night, cinematic
taillights in a tunnel, motion blur
rusted hull, orange and black
burnt texture, close up
concert crowd, orange stage light
basement show, flash photo, candid
astronaut spacewalk over earth, epic
rocket launch at night
orange wall, black shutters
motel room at night, lonely
graffiti scrawl through glass
old television broadcast, grainy

The tiles nobody would ever keyword-search for are doing the most work. A grilled mushroom as burnt texture. A teddy bear on a motel bed as the one tender image that makes the rest feel dangerous. A 1969 lunar telecast next to a 2006 basement show. Vibe search surfaces them because they are neighbors in feeling; taste keeps them.

Board two: the spectrum

The second board is a color study: six grounds and six skies, walked through the rainbow. The full query log is the concept:

red ground
orange ground
yellow ground
green ground
blue ground
violet ground
red sky
orange sky
yellow sky
green sky
blue sky
violet sky

Look at what the hard colors did. There is no green sky on Earth, so the corpus answered with the only honest one: soldiers under night vision. The green ground is a false-color radar composite from orbit; the violet ground is a stereo anaglyph of river meanders on Mars. When the literal thing does not exist, search by meaning finds the thing that feels like it, and it comes from wherever it lives: Flickr macros, a military archive, two NASA collections.

Board three: heavy industry

The third board took four queries and under three minutes:

steam, trains
industrial, hazy, vintage
industrial, hazy, vintage, machines
heavy machinery in black and white

Here the corpus's depth shows instead of its width. Three of these frames are Office of War Information negatives from 1940s factories, sitting in the Library of Congress; the Artemis crawler-transporter photo is from 2022. Eighty years of machines, one query language, one board.

The part you cannot see

All 36 tiles across the three boards are cleared for commercial use, as declared by their sources: Creative Commons photographers, NASA, the Library of Congress, and a public-domain military archive. The designer never checked a license, they applied one rule while clicking:

const usable = results.filter((r) => r.rights.commercial);

Every credit at the bottom of this page was generated from the attribution strings in the API responses, not written by hand. And nothing in these sessions needed human hands: the whole loop is two endpoints, search and similar, so an agent handed the same six-word brief runs the same loop and gets the same candidates. The taste is yours. The reach is anyone's.

{
  "license": "cc-by-sa",
  "license_url": "http://creativecommons.org/licenses/by-sa/2.0/",
  "as_declared_by": "Flickr (YFCC100M)",
  "commercial": true,
  "attribution_required": true,
  "attribution": "\"Meatlocker-2008-09-28\" by 7278633@N04 — Flickr (CC BY-SA)",
  "basis": "Rights shown as declared by the source, not verified by
   Lightdrift. You are responsible for clearance — copyright, and
   separately any likeness, trademark, or property rights a copyright
   license does not cover."
}