Thanks for the helpful critique. Your proposed algorithm is reasonable and much, much faster than simulated annealing, and I appreciate the live demo. It looks like your algorithm is deterministic (except for tie-breaking of a pair of columns that have the same difference?), so it doesn't matter which way the image is shuffled.

You are also correct that I overstated how hard the problem is; in retrospect it isn't that hard at all to unshred clean digital images (as opposed to paper scans).

On your web page:

> It takes roughly the same amount of time to run but correctly reconstructs all of the images.

This is not true, but is very close to correct. The image "Blue Hour in Paris" has some disturbances in the lower part of the sky, "Alaska Railroad" has a single wrong column on one side of the image, and "Abstract Light Painting" sometimes has a few incorrect columns on the edge. At least that's what I could spot by eye at a glance; I didn't rigorously compare them against the original images.