Pick any two references. This walks the straight line between them and, at every step, pulls the nearest real image the archive already holds. Not a generated blend — an actual path of things FIELD chose, from one pole to the other.
loading the archive…
Click any picture in the grid above to load it into the highlighted slot; the other slot arms itself next. Or pick a subject or a mood and it seats the most typical image of that category as the pole. Clicking a picture inside the corridor moves it into the armed slot, so you can walk on from anywhere.
The two vectors are unit length, so they sit on the surface of a sphere. The path between them is the great circle, not a straight line through the middle — a linear blend would bunch at the ends and race through the centre, and the strip would show movement that was an artefact of the arithmetic.
Once an image is used it leaves the pool. Without that the walk stalls: a dense region is the nearest neighbour to eight consecutive points on the line, and the strip returns the same picture eight times.
The break marks the largest single jump — the step where the line passes through territory the archive has no picture of. It is a hole located rather than described: two things FIELD looks at a lot, with a stretch of nothing in between.
This page walks a sample. It holds 5,000 of the 18,182 usable references — a 1-in-3.6 uniform draw on a fixed seed. Its jumps are therefore a little larger than the full archive's, and a hole found here can be a hole in the sample. Re-run anything interesting through corridor.py, which walks all of them.
Two more limits. Every cell is the nearest available image to a point on the line, and nearest can still be far — the strip does not show how far off the route each cell sits. And cosine in CLIP space is not human likeness; the model has its own reasons, and a corridor is worth looking at rather than trusting.
python3 src/viz/corridor.py --from-subject person --to-subject interior --steps 24