Rebuilding it yourself

The corpus is one tier of one run. Everything that made it is open, so you can rebuild this one or make a different one.

What you need

Stegobench, the harness the corpus was built with. It fetches the covers, builds the arms, packs the shards and derives the metadata.

Stegobench is not public yet

The harness has not been released, so the repository link is not open to readers today and everything below describes how a rebuild works rather than something you can run right now. The corpus itself is fully published and needs none of it. If you want to rebuild rather than download, say so on the issue tracker and it will help make the case for bringing the release forward.

What determinism means here

Pixels	Seeded per cover and per payload, so two runs produce the same images
Shards	Fixed member metadata, so the same corpus packs to the same bytes on any machine on any day

A shard whose digest moves with the clock cannot be checked against a published one, which is why the packing is fixed rather than merely repeatable.

What you cannot reproduce exactly

The cover set itself depends on what Wikimedia Commons held when it was fetched. Photographs are added, and occasionally removed. The manifest records every cover's source URL and digest, so you can re-obtain the same set as long as those files are still there, and see plainly which ones are not.

That is the same approach ImageNet takes, and it is the reason the corpus can be redistributed at all.