Using pHash data from an older version of pHash w/the current version

[sunild]

Apologies, this isn't really a bug, it's more of a question. But it might be useful for someone else down the road...

I was thrilled to discover this project, as we've been using a very old version of pHash (https://github.com/hszcg/pHash-0.9.6, it has some patches applied to it for things I'm assuming you've since addressed here). We have over 2 million existing pHash values for photos in our database calculated by the older version.

We're also using a Ruby wrapper for the pHash lib to compare photos: https://github.com/toy/pHash

It seems that, for a given photo, the pHash values from our old version differ from the value that the latest version produces. It's different enough that we're not able to detect the images as being similar... Is this expected?

And more importantly, is it is possible that this could happen when the pHash library is updated in the future? That a pHash calculated from today's code could be different from one calculated in some future version of the lib?

[DarkShadow44]

Any new on that?

[aetilius]

Hash compatibility across pHash versions has never been guaranteed -- the algorithms have evolved (e.g., #52 just changed which 8×8 block of the DCT is used to match the canonical Krawetz/Zauner convention; #53 fixed the radon projector; #54 changed the audio frame size to match Haitsma-Kalker; #55 switched the text hash from a tumbling to a sliding window).

Your old 0.9.6 hashes are not byte-compatible with current 1.0 hashes, and they won't be cross-version-compatible in general going forward. For an existing 2M-image index, the realistic path is:

1. Keep your existing index built with the old library, and stay on that version for read.
2. When you start writing new entries, write them with whatever pHash version is current at that time, but tag them with the version that produced them.
3. Only compare hashes produced by the same library version.

If you want stable, format-versioned hashes, that's a separate feature request (we'd need a hash-format version byte and a documented backward-compat policy). Worth opening a new issue for that if it matters to you.

Closing this thread -- but #46 is essentially the same question, leaving that one open as the canonical conversation.