> This could be super useful for deduplicating large photo datasets!

That is indeed the use case mentioned in the article. A hash like this is useful for deduplicating large _file_ datasets but for deduplicating images in particular, you really want a perceptual hash. Because two images files can _look_ identical but have slightly (or wildly) different byte streams. The trade-off is that perceptual hashes are notorious for false positives in large datasets.