Filenames get changed, pixels get recompressed and metadata gets stripped. An archive needs several identifiers, each built to survive a different insult, and a register that ties them together.
No single identifier survives everything a photograph goes through, so the question is never which one to choose but which three to keep.
A photograph in an archive is at least three things: an object in a sleeve, a master file on a disk, and a picture that will be copied, cropped and recompressed far beyond the archive's sight. Each needs an identifier, and no identifier works for all three. This article's angle is to sort the options by what they survive. An accession number survives everything except being forgotten. A file hash survives nothing but proves everything. A content code such as the new ISO standard, ISCC, sits between them, and it changes the picture for archives because nobody has to issue it.
Why are filenames not enough?
A filename is a label that any operating system, upload form or well-meaning colleague can change. Photographs sent to a publisher come back as final_v2_USE THIS.jpg. A platform renames every upload to a string of its own. Two different scans both called 0001.tif meet in the same folder.
Good naming still matters, and our guide to file naming conventions for photo archives sets out the rules. But the name should be derived from an identifier that lives somewhere safer. The name is where the identifier is displayed, not where it is kept.
What is an accession number, and do photographs need one?
The oldest identifier is the one written on the object. Collections Trust, which maintains the Spectrum standard used by UK museums, puts the requirement in one line: you give a unique number to each accessioned object. Its guidance describes two common patterns, a simple running number, or the year of accession followed by a running number, such as 2024.3, 2024.4, 2024.5. The second is the more usual. Parts of one acquisition take a further suffix, so 2024.24.1 to 2024.24.4 is the twenty-fourth accession of 2024, in four parts.
The same guidance lists habits to avoid: abbreviating the year to two digits, building catalogue information such as classification or storage location into the number, and padding with leading zeros. That last point will surprise anyone who has just read a file-naming guide, including ours, which recommends zero padding so that files sort correctly. Both are right. A museum number is read by people and written on objects; a filename is sorted by machines. The register can hold 2024.24.3 while the file is called 2024_0024_0003.
For a photographic archive that is not a museum, the principle carries over even if the format does not. Every physical original, whether a strip of negatives, a mounted transparency or a Polaroid, needs a number that is unique, never reused, free of meaning that might change, and physically associated with the object by a label on its sleeve or mount. Everything digital hangs from it. A scan without a link to an object number is an image of uncertain origin, however good its metadata.
What does a file hash identify?
A cryptographic hash such as SHA-256 reduces a file of any size to a value of 256 bits, usually written as 64 hexadecimal characters. Identical files give identical values. Any change at all, a single pixel or a single character of embedded caption, gives a completely different one. Nobody assigns it. It is a property of the bytes.
That makes it the right tool for fixity, the term digital preservation uses for demonstrating that a file has not changed. Record the hash of each master scan when it is made, keep the list separately from the files, and you can prove years later that a master is exactly what the scanner produced. The image hash and fingerprint tool computes SHA-256 in the browser without uploading anything, and our piece on backups and checksums covers the routine.
The same strictness makes it useless for recognising a picture. Export the master as a JPEG, or add a copyright notice to its metadata, and the result has a new hash with no visible relation to the old one. A hash identifies a file. Every derivative is, to SHA-256, a stranger.
What survives resizing and recompression?
For that you need an identifier computed from what the picture looks like. A perceptual hash shrinks the image to a small grid, compares neighbouring brightness values and records the pattern as a short string of bits. A resized or recompressed copy produces the same string or one that differs in only a few places, and the number of differing bits measures how alike two images are. We use the technique for housekeeping in finding duplicate photographs in an archive.
Perceptual hashes have two weaknesses as identifiers. There are many algorithms and no agreed one, so two organisations' values are not comparable. And they are fuzzy by design: near-identical frames from the same shoot, which a film archive holds in quantity, can collide. They are a search aid. They do not amount to a name.
What is ISCC, the International Standard Content Code?
ISCC is an attempt to make content-derived identification a standard. It was published by the International Organization for Standardization in May 2024 as ISO 24138:2024, under the committee responsible for identifiers such as ISBN. ISO's abstract describes it as an identification system for digital assets, covering text, images, audio and video, and the ISCC Foundation maintains the open implementation.
Three properties distinguish it.
It is derived, not assigned. An ISCC is computed from the file by published algorithms. There is no agency to apply to and no fee. Anyone holding the same content derives the same code. The standard explicitly leaves registration out of scope.
It is layered. A full code is assembled from several units, each a hash of a different aspect of the asset: the embedded metadata, the normalised content, and the raw data, ending with a unit for the exact instance. So one identifier carries both a "this picture" component and a "this exact file" component.
It preserves similarity. According to the ISCC Foundation, each unit is a compact similarity-preserving hash: the same content yields the same code after re-encoding or compression, and similar content yields similar codes. That is the perceptual hash idea, standardised.
The standard also describes how ISCCs sit alongside assigned identifiers such as DOI and ISBN, which is the right way to think about it. An ISCC does not replace an accession number. It lets a stranger who has only the picture arrive at the same code the archive holds, and so find the archive.
Adoption is early, and a new standard is a bet. But the cost of computing a code at the time of scanning is close to nothing, and it cannot be done retrospectively for a file you no longer have.
Which identifiers should an archive combine?
| Layer | Identifier | Assigned or derived | Survives | Fails when |
|---|---|---|---|---|
| Physical original | Object or accession number | Assigned by the archive | Everything, if the register is kept | The label is separated from the object |
| Master file | SHA-256 hash | Derived from the bytes | Renaming, copying, moving | Any change to the file |
| Picture | Perceptual hash or ISCC | Derived from the content | Resizing, recompression, format change | Heavy cropping or editing |
| Embedded reference | IPTC identifier fields | Written into the file | Renaming | Metadata is stripped |
| Public reference | A stable web address per item | Assigned by the archive | Being cited | The website is reorganised |
The embedded layer has specific homes. The IPTC standard provides Digital Image GUID and Image Supplier Image ID for the file's own reference, and a Source Inventory Number inside its Artwork or Object structure for the number of the physical original. Filled in, they let a stray file identify itself, as we note in IPTC fields for a photo archive.
The register that connects these layers is the real asset. One row per master: object number, filename, SHA-256, content code, date scanned, operator. It is a small table, and it is the difference between an archive and a folder of pictures.
Why does this matter for licensing and for AI?
For licensing, an identifier that survives recompression is how an archive recognises its own picture on a magazine's website five years after the invoice.
For machine learning, it is how a dataset is audited. A buyer wants to show which files were delivered and that none has been altered: file hashes do that. A researcher wants to know whether a test image already exists in a training set: content codes and perceptual hashes do that, and the stakes are set out in evaluation sets for image models. Research showing that generative models can reproduce individual training images has made the question of exactly which images went in more than academic.
Where the Robert Harper archive stands
The Robert Harper Photograph Library's public handle today is the folder. The archive inventory lists 503 of them, each with a name, a description and counts of shots and frames, totalling 95,487 unique photographs and 177,143 frames including duplicates and variants. Those counts come from the inventory and have not yet been reconciled against the physical material.
Item-level identifiers do not yet exist for most of the archive, because more than half of it is still on film and has yet to be scanned. That is a gap, and also an opening. A scheme designed before scanning begins can assign the object number at the sleeve, write it into the file, and record the hash and content code in the same pass, so that no master ever exists without all of them.
Scanning and cataloguing specialists who build that register as part of the job, and not as an afterthought, are invited to get in touch through digitisation partners.
Frequently asked questions
What is the best unique identifier for images?
There is no single best one, because identifiers survive different things. An assigned number identifies the physical original and never changes. A cryptographic hash such as SHA-256 identifies one exact file. A perceptual hash or an ISCC identifies the picture across resizing and recompression. A well-run archive records all three against each other.
What is an ISCC code?
The International Standard Content Code is an identifier for digital content defined in ISO 24138:2024. It is generated from the file by open algorithms, combining several hash units that describe the metadata, the content and the exact data. Anyone can compute it, the same content gives the same code, and similar content gives similar codes.
Is a SHA-256 hash a good identifier for a photograph?
It is an excellent identifier for a file and a poor one for a picture. Two files with the same SHA-256 are byte-for-byte identical, which is ideal for proving that a master scan is unaltered. But saving the image again, resizing it or editing its metadata produces a different hash, so it cannot recognise copies.
What is an accession number?
It is the unique number a collecting institution gives to an object when it formally enters the collection. Under Collections Trust guidance the common form is the year of accession and a running number, with a suffix for parts, and the number is marked on or attached to the object. For photographs it identifies the physical negative, transparency or print.
Should the identifier go in the filename or the metadata?
Both, and also somewhere outside the file. A filename based on the identifier makes files findable; embedded metadata fields carry it when the file is renamed; and a catalogue or register holds it when both have been altered. The register is the authority. Filenames and embedded fields are convenient copies of what it says.
Sources
- Collections Trust: Numbering (Spectrum guidance on accession numbers)
- ISO 24138:2024 Information and documentation: International Standard Content Code (ISCC)
- ISCC Foundation: International Standard Content Code
- Digital Preservation Coalition Handbook: Fixity and checksums
- IPTC Photo Metadata Standard specification (identifier properties)
- Carlini et al., Extracting Training Data from Diffusion Models (arXiv 2301.13188)
Published by Archive.London. Corrections are welcome under our editorial policy.

